Brain Organization and Memory

This page intentionally left blank

Brain
Organization
and Memory:
Cells, Systems, and Circuits

Edited by

JAMES L. McGAUGH
NORMAN M. WEINBERGER
GARY LYNCH

New York Oxford

OXFORD U N I V E R S I T Y PRESS
Oxford University Press

Oxford New York Toronto

Delhi Bombay Calcutta Madras Karachi
Kuala Lumpur Singapore Hong Kong Tokyo
Nairobi Dar es Salaam Cape Town
Melbourne Auckland
and associated companies in
Berlin Ibadan

Copyright © 1990 by Oxford University Press

First published in 1990 by Oxford University Press, Inc.,

200 Madison Avenue, New York, New York 10016
First issued as an Oxford University Press paperback, 1992
Oxford is a registered trademark of Oxford University Press
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise,
without prior permission of Oxford University Press.

Library of Congress Cataloging-in-Publication Data

Brain organization and memory : cells, systems, and circuits /
edited by James L. McGaugh, Norman M. Weinberger, Gary Lynch.
p. cm.
Based on the Third Conference on the Neurobiology of Learning and Memory
organized by the Center for Neurobiology of Learning and Memory of
the University of California, Irvine, held at Irvine on Oct. 14-17, 1987.
Includes bibliographies and index.
ISBN 0-19-505496-2 ISBN 0-19-507712-1 (Pbk.)
1. Memory —Congresses. 2. Cerebral cortex Physiology—Congresses.
3. Neural circuitry—Congresses. L McGaugh, James L.
II. Weinberger, Norman M. III. Lynch, Gary.
IV. University of California, Irvine. Center for the Neurobiology of Learning and Memory.
[DNLM: L Cerebral Cortex physiology -congresses. 2. Memory—physiology-congresses.
3. Models, Neurological—congresses. 4. Nervous System- physiology—congresses.
WL 102 B8132 1987] QP406.B735 1989 153.T2 dc19
DNLM/DLC for Library of Congress 88-38920

9 8 7 6 5 4 3 2 1
Printed in the United States of America
Preface

This volume is the third in a series. Like its predecessors, Neurobiology of Learn-
ing and Memory and Memory Systems of the Brain, this book is the outgrowth
of a conference that was organized by the Center for the Neurobiology of Learn-
ing and Memory of the University of California, Irvine. The Third Conference
was held at Irvine on October 14-17, 1987. One might ask, as we did, whether
a book summarizing work on how the brain encodes and uses memory is needed
every two to three years. An examination of the tables of contents of the three
books clearly answers this question. Some of the central theoretical and meth-
odological issues addressed in this book are not so much as noted in the indexes
of the previous volumes. This attests to the active and expanding interest in
research on brain, learning, and memory.
This book focuses on three related topics: forms of memory, regulation of cor-
tical function in memory, and representations. Some of the themes developed
in the first two volumes are given further emphasis in this book. Prominent
among these are the efforts to relate learning and memory as they are studied in
animals to those processes observed in humans. Despite steady experimental
and conceptual progress, development of an adequate taxonomy of memory
continues to hamper efforts to achieve unification of memory theory. This
annoying fact should temper our hopes for a rapid and complete integration of
the subtle and surprising phenomena of memory with those of brain cells, sys-
tems, and circuits. Nonetheless, it is clear that the rate of progress in understand-
ing the cellular processes underlying memory continues to accelerate. We are
beginning to see the emergence of a consensus regarding the physiological and
chemical events that translate activity patterns into long-term changes in the
strength of connections between brain cells. We are also now seeing behavioral
studies of memory using predictions derived from understanding of synaptic
plasticity. One can only be impressed at how quickly the world of ions and
receptors is being translated into the realm of memory systems and mazes.
Special emphasis is given in this volume to computational and modeling
approaches to the study of brain and memory. The development of network
vi Preface

models and computer simulations of memory permeates current research. This

effort is explicitly recognized in the final section of the book, but the use of net-
work models is seen throughout the book. Experimentalists working on cortex
and hippocampus are attempting to link cells to systems with interesting and
powerful models. However, the gap between formal mathematically tractable
models and the stuff of brains remains clear and provides a continuing challenge
for experimentalists and theorists alike.
In all, the leitmotif of this book, like that of its predecessors, is excitement.
Excitement about what we have learned as well as what we can now ask. We
hope that readers of Brain Organization and Memory: Cells, Systems, and Cir-
cuits will share our excitement as well as our eagerness to follow this intellectual
drama in volumes yet to come.

Irvine, Calif. J. L. McG.

October 1988 N. M. W.
G. L.
Acknowledgments

The conference on which this book is based was planned and organized by the
members of the Center for the Neurobiology of Learning and Memory at the
University of California, Irvine. The selection of topics and authors grew out of
extensive discussions with our colleagues John Ashe, Michel Baudry, Richard
Granger, Mary-Louise Kean, Mark Rosenzweig, Gordon Shaw, Larry Squire,
and Larry Stein. Funding for the conference was provided by a number of orga-
nizations: the UCI Office of Research and Graduate Studies, the Office of Naval
Research, the Air Force Office of Scientific Research, the National Science Foun-
dation, the Sloan Foundation, the Irvine Company, Searle Research and Devel-
opment, Beckman Instruments, and Hoechst-Roussel Pharmaceuticals. We
thank Lynn Brown, Nan Collett, and Lori LaSalle for their skilled coordination
of all aspects of the conference. Special thanks are also due to Lynn Brown for
editorial assistance.
This page intentionally left blank
Contents

1. Neurobiology of Memory: The Significance of Anomalous

Findings 3
JAN BURES

I Forms of Memory
Introduction 23
MICHELA GALLAGHER
2. The Development of Learning and Memory in Aplysia 27
THOMAS J. CAREW
EMILIE A. MARCUS
THOMAS G. NOLEN
CATHARINE H. RANKIN
MARK STOPFER

3. Synaptic Plasticity, Neural Architecture, and Forms of Memory 52

RICHARD G. M. MORRIS

4. Forms of Memory in Pavlovian Conditioning 78

PETER C. HOLLAND

5. Functional Forms of Human Memory 106

MARCIA K. JOHNSON

COMMENTARIES AND ALTERNATIVE PERSPECTIVES

6. Neuromnemonics: Forms and Contents 137

NORMAN M. WEINBERGER
7. Time and Memory 145
ROBERT W. DOTY

8. Forms of Memory: Issues and Directions 159

ARTHUR P. SHIMAMURA
x Contents

II Regulation of Cortical Function in Memory

Introduction 177
MARK R. ROSENZWEIG

9. Functions ofNeuronal Networks in the Hippocampus

and of Backprojections in the Cerebral Cortex in Memory 184
EDMUND T. ROLLS
10. Ontogenetic Self-Organization and Learning 211
WOLF SINGER

11. The Dissection by Alzheimer's Disease of Cortical

and Limbic Neural Systems Relevant to Memory 234
GARY W. VAN HOESEN

COMMENTARIES AND ALTERNATIVE PERSPECTIVES

12. The Neocortex and Memory Storage 265

HERBERT P. KILLACKEY

13. A Network Model for Learned Spatial Representation

in the Posterior Parietal Cortex 271
RICHARD A. ANDERSEN DAVID ZIPSER
DAVID ZIPSER

14. Cortical Localization of Working Memory 285

PATRICIA S. GOLDMAN-RAKIC

III Representations: Beyond the Single Cell

Introduction 301
GORDON L. SHAW
15. Neural Networks: Test Tubes to Theorems 306
LEON N. COOPER
MARK F. BEAR
FORD F. EBNER
CHRISTOPHER SCOFIELD

16. Notes on Neural Computing and Associative Memory 323

TEUVO KOHONEN

17. Building Network Learning Algorithms from Hebbian

Synapses 338
TERRENCE J. SEJNOWSKI
GERALD TESAURO
18. A Neural Architecture for the Representation of Scenes 356
CHRISTOPH VON DER MALSBURG
 Contents xi

COMMENTARIES AND ALTERNATIVE PERSPECTIVES

19. Representations: Who Needs Them? 375

WALTER J. FREEMAN
CHRISTINE A. SKARDA

20. Interactions Within Neuronal Assemblies:

Theory and Experiment 381
GEORGE L. GERSTEIN
21. Neural Networks and Networks of Neurons 390
GARY LYNCH
JOHN LARSON
DOMINIQUE MULLER
RICHARD GRANGER

Index 401
This page intentionally left blank
Contributors

RICHARD A. ANDERSEN
Department of Brain and Cognitive Sciences
Massachusetts Institute of Technology
Cambridge, Massachusetts 02139

MARK F. BEAR
Center for Neural Science
Brown University
Providence, Rhode Island 02912

JAN BURES
Institute of Physiology
Czechoslovak Academy of Sciences
Prague 4, Czechoslovakia

THOMAS J. CAREW
Departments of Psychology and Biology
Yale University
New Haven, Connecticut 06520

LEON N. COOPER
Department of Physics and Center for Neural Science
Brown University
Providence, Rhode Island 02912

ROBERT W. DOTY
Department of Physiology
University of Rochester
Rochester, New York 14642
xiv Contributors

FORD F. EBNER
Center for Neural Science
Brown University
Providence, Rhode Island 02912

WALTER;. FREEMAN
Department of Physiology-Anatomy
University of California
Berkeley, California 94720

MICHELA GALLAGHER
Department of Psychology
University of North Carolina
Chapel Hill, North Carolina 27514

GEORGE L. GERSTEIN
Department of Physiology
University of Pennsylvania
Philadelphia, Pennsylvania 19104

PATRICIA S. GOLDMAN-RAKJC
Section of Neuroanatomy
Yale University School of Medicine
New Haven, Connecticut 06510

RICHARD GRANGER
Center for the Neurobiology of Learning and Memory
and Department of Information and Computer Science
University of California
Irvine, California 92717

PETER C. HOLLAND
Psychology Department
Duke University
Durham, North Carolina 27706

MARCIA K. JOHNSON
Department of Psychology
Princeton University
Princeton, New Jersey 08544

HERBERT P. KILLACKEY
Department of Psychobiology
University of California
Irvine, California 92717
 Contributors xv

TEUVO KOHONEN
Laboratory of Computer and Information Science
Helsinki University of Technology
Espoo, Finland
JOHN LARSON
Center for the Neurobiology of Learning and Memory
University of California
Irvine, California 92717
GARY LYNCH
Center for the Neurobiology of Learning and Memory
and Department of Psychobiology
University of California
Irvine, California 92717
EMILIEA. MARCUS
Department of Psychology
Yale University
New Haven, Connecticut 06520
RICHARD G. M. MORRIS
Department of Pharmacology
University of Edinburgh Medical School
Edinburgh EH8 9JZ
Scotland
DOMINIQUE MULLER
Department of Pharmacology
Centre Medical Universitaire
Geneva 4, Switzerland
THOMAS G. NOLEN
Department of Biology
University of Miami
Coral Gables, Florida 33124

CATHARINE H. RANKIN
Department of Psychology
University of British Columbia
Vancouver, British Columbia, Canada V6T1W5
EDMUND T. ROLLS
Department of Experimental Psychology
University of Oxford
Oxford OX 1 3UD
England
xvi Contributors

MARK R. ROSENZWEIG
Department of Psychology
University of California
Berkeley, California 94720

CHRISTOPHER SCOFIELD
Nestor Inc.
Providence, Rhode Island 02906

TERRENCEJ. SEJNOWSKI
The Salk Institute
La Jolla, California 92037

GORDON L. SHAW
Center for the Neurobiology of Learning and Memory
and Department of Physics
University of California
Irvine, California 92717

ARTHUR P. SHIMAMURA
Veterans Administration Medical Center
San Diego, California 92161

WOLF SINGER
Max-Planck-Institute for Brain Research
Frankfurt 71
Federal Republic of Germany

CHRISTINE A. SKARDA
Department of Physiology-Anatomy
University of California
Berkeley, California 94720

MARK STOPPER
Department of Psychology
Yale University
New Haven, Connecticut 06520

GERALD TESAURO
IBM Thomas J. Watson Research Center
Yorktown Heights, New yo rk 10598

GARY W. VAN HOESEN

Department of Anatomy
University of Iowa
Iowa City, Iowa 52242
 Contributors xvii

CHRISTOPH VON DER MALSBURG

Computer Science Department
University of Southern California
Los Angeles, California 90089

NORMAN M. WEINBERGER
Center for the Neurobiology of Learning and Memory
and Department of Psychobiology
University of California
Irvine, California 92717

DAVID ZIPSER
Institute for Cognitive Science
University of California, San Diego
La Jolla, California 92093
This page intentionally left blank
Brain Organization and Memory
This page intentionally left blank
1
Neurobiology of Memory:
The Significance of Anomalous Findings
JAN B U R E S

The neurobiology of learning and memory is a relatively young discipline with

a vague status that lacks a unifying theory. Facts are accumulating at a rate that
vastly exceeds our capability to explain them. Research strategies highly valued
in physics find little application in neurobiology. The almost obsessive interest
of physicists in anomalies, an attitude that has led to the discovery of planets,
subatomic particles, big bang, and superconductivity, is not seen in neurobiol-
ogy, where it is often difficult to distinguish an anomaly from the normal state
of affairs, to recognize an anomalous phenomenon against the background vari-
ability of the process studied. Even when the oddity of a finding is obvious and
proved beyond any doubt, it is easier to consider it an enigmatic exception than
to seek an explanation requiring a fundamental revision of the accepted beliefs.
I believe that our discipline is mature enough to face embarrassing facts and that
the study of anomalies will not undermine the coherence of the field but open
the way to more comprehensive cross-level synthesis of cellular, systemic, and
behavioral investigations.

ROBUST LEARNING CAN PROCEED WITHOUT AWARENESS

One of the basic tenets of contemporary neurobiology is that associative learning

requires not only the specific input of signals to be processed (i.e., CS and US),
but also various auxiliary mechanisms—for example, attention, motivation,
emotion—that together account for a particular central state characterized as
consciousness, awareness, or alertness and expressed by the corresponding level
of arousal. In spite of countless examples supporting this belief, it would be a
mistake to assert that awareness is a necessary condition of learning. Condi-
tioned taste aversion (CTA) is a dramatic counterexample. The rat's capability
to avoid a flavor whose ingestion has been followed by illness (Barker, Best, &
Domjan, 1977; Braveman & Bronstien, 1985; Garcia, Kimeldorf, & Koelling,

3
 Neurobiology of Memory 5

1955) is not diminished when immediately after presentation of the taste stim-
ulus the animal is anesthetized with pentobarbital and maintained under deep
anesthesia throughout the duration of the poison-elicited illness (Berger, 1970).
The gustatory stimulus has left in the rat's brain a trace that is later associated
with the visceral symptoms of intoxication. Anesthesia interferes neither with
the persistence of this trace nor with its association with the poison-elicited sig-
nals. Similar results were obtained with other comatose states elicited by func-
tional decortication (Buresova & Bures, 1973), deep (20°C) hypothermia
(Ionescu & Buresova, 1977), cerebral ischemia, or prolonged seizures (Bures &
Buresova, 1987). When these states are elicited after presentation of the taste
stimulus (Fig. 1.1) they do not induce CTA when applied alone, and they do not
prevent CTA acquisition when inserted between the CS and US, that is, when
applied before administration of the poison. In fact, we have not yet found any
reversible intervention that would disrupt CTA acquisition under the above
conditions. This is perhaps not surprising if we take into account the biological
significance of CTA, which protects animals against the fatal error of repeated
ingestion of lethal poisons.
Accurate memory record of the sensory properties of food whose ingestion has
preceded poisoning gives the animal a clear advantage in the struggle for sur-
vival. The value of this behavior is greatly increased when the learning is not
disrupted by a comatose state that may be a part of the poisoning. Evolution has
recognized the vital significance of the taste-poisoning association and has
equipped animals with failure-proof brain mechanisms that ensure reliable
recording of the relevant information as long as the animal is alive; thus an ani-
mal that recovers can profit from the experience and prevent fatal poisoning by
avoiding the particular flavor.
The various instances of learning form a continuum ranging from the most
universal associative phenomena, allowing combination of any pair of stimuli
while requiring prolonged training and activation of auxiliary mechanisms, to
dedicated associative processes automatically entering variables of a definite
class into highly prepared behavioral schemes. Learning mechanisms supporting
food selection are typical examples of the latter questionnaire-type learning. The
sensory properties (particularly the taste) of ingested food are automatically

FIGURE 1.1. The effect of pentobarbital anesthesia (50 mg/kg, A), picrotoxin convulsions (5
mg/kg, C), cerebral ischemia (5-min occlusion of neck vessels under ether anesthesia, I), or sham
treatment (ST) on the acquisition of CTA. The columns indicate mean (± SEM, n = 10) saccharin
preference during the retention test on day 5. W = water; S = saccharin. The conditions during
CTA acquisition on day 3 are schematically shown on the left. (A) Saccharin drinking followed
by any of the foregoing experimental treatments alone does not elicit CTA. (B) Saccharin drinking
followed by intraperitoneal injection of lithium chloride (0.15 mol/1, 2% body weight) elicits strong
CTA, which is not prevented by the foregoing interventions applied after the gustatory CS and
before the visceral US. (C) The same interventions (with the exception of ST) prevent CTA
acquisition when applied prior to systemic administration of the taste stimulus (2% saccharin, 1%
body weight) followed by lithium chloride poisoning.
6 NEUROBIOLOGY OF MEMORY

recorded in a short-term memory file where they wait during several hours for
visceral signals generated by the passage of the food through the gastrointestinal
tract. Only two outcomes are monitored: (1) the food has no adverse conse-
quences and satisfies hunger or thirst and (2) the food elicits gastrointestinal dis-
tress and other signs of sickness. After a time sufficient for food digestion and
resorption has elapsed, the neutral input signal is transferred from short-term
memory into permanent memory with the label acceptable or aversive. The only
intentional part of the process is the initial food intake, which starts an auto-
matic chain of events generating a reliable record of gustatory-visceral contin-
gencies, essential for the future feeding behavior of the animal.
The gustatory trace-poisoning association in anesthetized animals was dis-
covered by Berger (1970) and has since been replicated in a score of studies in
different laboratories (Buresova & Bures, 1977, 1979a, 1979b, 1984; Millner &
Palfai, 1975; Rabin & Rabin, 1984; Roll & Smith, 1972). Although these reliable
and easily reproducible experiments clearly demonstrate that robust learning
can proceed without awareness, the evidence was dismissed as an interesting
peculiarity that does not require any fundamental revision of the conventional
view. Of course this is not true. A single demonstration of this type shows that
the theory is not generally valid and that other, not immediately obvious alter-
natives must be considered.

CONDITIONED-UNCONDITIONED STIMULUS PAIRING PROCEEDS

IN STAGES DIFFERENTIALLY SENSITIVE TO DISRUPTION

Some of the problems involved can again be illustrated through CTA research.
Whereas the gustatory trace-poisoning association is not disrupted by the deep-
est coma, formation of the gustatory trace is a fragile process that can be blocked
by mild hypothermia (Buresova & Bures, 1981), previous paradoxical sleep
deprivation (Danguir & Nicolaidis, 1976), or anterograde ECS effects (Buresova
& Bures, 1979a). Interventions that block spontaneous drinking and feeding, for
example, functional decortication by cortical spreading depression (CSD) (Bur-
esova & Bures, 1973) or general anesthesia, prevent formation of the gustatory
trace by forcefeeding or by systemic application of the concentrated flavor (e.g.,
intraperitoneal injection of 2% saccharin, 1% body weight; Bellingham & Lloyd,
1987; Buresova & Bures, 1977), which is effective in intact animals. Thus the
same intervention, for example, anesthesia, disrupts or does not disrupt CTA
acquisition according to its timing with respect to the various phases of the
learning process (Fig. 1.2). An important advantage of the CTA model is that
the long CS-US delay makes it possible to differentially influence CS processing
and formation of the gustatory trace, persistence of this trace, and its association
with the symptoms of poisoning. This is not likely to occur in other types of
learning characterized by CS-US delays of less than a few seconds. Even such
instances of conditioning may proceed in several successive stages that are dif-
ferentially sensitive to disruption but cannot be easily separated from one
another.
 Neurobiology of Memory 1

FIGURE 1.2. The effect of pentobarbital anesthesia (50 mg/kg) on CTA acquisition on day 3.
Ordinate: Mean saccharin preference (±SEM, n = 10) during the retention test on day 5. Ab-
scissa: Time of pentobarbital injection (50 mg/kg) before (—) or after systemic administration o
the taste stimulus (2% saccharin, 1% body weight) followed two hours later by lithium chloride
poisoning (0.15 mol/1, 2% body weight). Note that CTA is disrupted when the taste stimulus is
applied to anesthetized rats or to rats recovering from anesthesia, but that strong CTA develops
when the gustatory stimulus precedes the onset of anesthesia.

The successive stages of CTA acquisition are implemented by different brain

structures. The finding that formation of the gustatory trace is disrupted by
reversible functional decortication suggests that this stage of CTA acquisition
requires cortical participation. But the same intervention does not impair per-
sistence of the gustatory trace and/or its association with the visceral symptoms
of poisoning because the underlying plastic change takes place at a subcortical
level, probably in the lower brain stem. It is important to realize that this dif-
ferential sensitivity to interference cannot be established by brain lesions, which
always affect the most failure-prone link of the process but do not allow recog-
nition of the better protected links. The vast literature on the effects of brain
lesions on CTA (Riley & Tuck, 1985) almost exclusively describes the signifi-
cance of the damaged structures for the formation of the short-term gustatory
trace but does not assess their role in the gustatory trace-poisoning association.
Papers describing CTA acquisition deficits after ablation of the gustatory cortex
(see Kiefer, 1985, for review) do not consider the possibility that the cortex is
8 NEUROBIOLOGY OF MEMORY

not necessary for the persistence of the gustatory trace and for its association
with poisoning. A more realistic picture of the neural mechanisms involved can
be obtained with the use of functional ablation procedures. Thus the finding that
CTA acquisition is impaired in amygdalectomized rats was analyzed by Ber-
mudez-Rattoni, Rusiniak, and Garcia (1983), who demonstrated that reversible
inactivation of amygdala by novocaine prevents taste-potentiated odor aversion
when employed before presentation of the compound odor-taste stimulus. The
same intervention applied after the CS but before the visceral US was ineffec-
tive. Like the gustatory cortex, amygdala participates in the initial formation of
the short-term memory trace but is not indispensable for the poisoning-induced
aversive labeling of this trace.
Another deeply ingrained concept represents the association process by direct
interaction of the CS- and US-activated circuits. In fact, classical conditioning
is best expressed when CS and US overlap in time and deteriorates with increas-
ing CS-US delay in case of trace conditioning. It is possible to surmise that the
CTA blockade observed when the gustatory CS has been presented during CSD
or anesthesia is due to the failure to form the gustatory trace and to bridge the
temporal gap between the CS and US. The gustatory trace might be considered
unnecessary when the taste signal overlaps with poisoning and when there is an
opportunity for direct convergence of the CS and US volleys on the circuits
mediating aversive labeling of the gustatory experience (Fig. 1.3), but experi-
ments show that this is not the case. Whereas forcefeeding saccharin to intact
rats poisoned with lithium chloride elicits moderate CTA, no CTA is acquired
under the same conditions by rats under bilateral CSD (Bures & Buresova,
1981). This indicates that CSD or other CTA-disrupting interventions prevent
transcription of the gustatory signal into a memory buffer accessible to visceral
signals. The purpose of this buffer is not only to cope with the CS-US delay, but
also to provide a device for interaction between the temporarily stored neutral
signal with other relevant signals and for its eventual transformation into an
interpreted, labeled engram stored in permanent memory. It is conceivable that
a similar scheme applies to other forms of conditioning that are implemented
not by direct interaction between CS- and US-elicited volleys but between rep-
resentations of the CS and/or US signals. The indispensability of the short-term
gustatory trace for CTA formation is thus an important argument in support of
the cognitive theories indicating that the multistage organization of the learning
process cannot be replaced by a simpler system, corresponding to a verbatim
interpretation of the S-R or S-S theories.

MOTOR LEARNING DOES NOT ALWAYS OBEY THE LAW OF EFFECT

Thus far we have discussed the sensory aspects of learning, the processes that
store a neutral sensory stimulus and change it into a signal of biologically sig-
nificant events. In most cases this is sufficient for effective coping behavior, con-
sisting of inhibition of some and activation of other preprogrammed items of
 Neurobiology of Memory 9

FIGURE 1.3. Schematic representation of the putative stages of CTA acquisition. Above: Serial
model. Below: Serioparallel model. CS = conditioned gustatory stimulus; US = unconditioned
visceral stimulus; GSTM = gustatory short-term memory; CTA = aversive labeling of the gus-
tatory stimulus; (CS + US)/CSD = simultaneous presentation of CS and US during bilateral
cortical spreading depression. Heavy lines indicate the part of the circuit that is active during the
particular stage of the experiment.

the animal's repertoire such as eating, running, and freezing. Sometimes, how-
ever, adaptive behavior cannot be achieved by the already available motor pro-
grams. In this case, learning consists not only of the detection and interpretation
of stimulus contingencies, but also of the modification of the motor response of
the animal in the acquisition of new skills.
An obvious advantage of motor learning for brain and behavior research is
that the putative plastic change takes place close to the motor output and can
therefore be readily observed, localized, and analyzed by electrophysiological
methods. The remarkable progress in research into the motor functions (Brooks
& Thach, 1981) has significantly advanced our understanding of motor learning,
particularly in primates. But the validity of some general principles must be
reassessed in light of the physiological analysis of motor learning. The most
important of these principles is the "law of effect," formulated almost a century
ago by Thorndike (1911), which states that the strength of an operant not fol-
lowed by reinforcement gradually decreases. This law, borne out by countless
experiments, is the theoretical basis of operant conditioning and represents an
10 NEUROBIOLOGY OF MEMORY

important tenet of the universalistic concepts asserting that learning can asso-
ciate any pair of stimuli or establish bonds between any stimulus and any reac-
tion of the organism. Whereas important species-specific constraints described
by Breland and Breland (1961) have demonstrated the limited validity of the
law of effect in general, analysis of the physiological mechanisms of motor learn-
ing shows that the law requires a wider interpretation, taking into account not
only the delivery of reinforcement but also the effort required for the necessary
reorganization of the existing motor programs. This assertion can be illustrated
by experiments attempting to modify rhythmic licking in rats by motor learning.
Licking, like mastication and swallowing, belongs to feeding-related oral
behaviors implemented by a central rhythmic generator residing in the lower
brain stem (Halpern, 1977; Malmo, Malmo, & Weijnen, 1987; Weijnen & Men-
delson, 1977). The generator is switched on and off from higher forebrain centers
and is influenced by trigeminal, glossopharyngeal, and vagal inputs. In a thirsty
rat, the generator is switched on in the presence of a drinking spout and switched
off only some time after the spout has been suddenly removed (Mamedov &
Bures, 1984). The nonrewarded licks emitted in the absence of the spout
decrease the overall efficiency of behavior and the law of effect requires that their
incidence is maximally reduced. This assumption was tested by training rats to
alternate between two retractable drinking spouts accessible through adjacent
openings in the wall of the cage (Mamedov, Hernandez-Mesa, & Bures, 1987).
After the first lick at the left spout, this spout was removed by a solenoid out of
the reach of the animal's tongue, while the other spout was advanced to the right
opening. After the first lick at the right spout, this spout was retracted and the
left spout was advanced (Fig. 1.4). The licks at the two spouts were monitored
with photoelectric lickometers and their incidence was analyzed with a labora-
tory computer.
After up to 10 days of training and thousands of spout alternations, the ani-
mals continued to emit more than one lick per spout presentation (2.5 licks on
average) and this value even grew when the spout separation was increased by
introduction of a vertical partition. This apparent violation of the law of effect
is not a result of the inability of the rat to emit single licks: if the animal had to
alternate between an empty spout and a spout containing water, the additional
licks were produced only at the rewarded spout, whereas they were almost
absent at the empty spout (Fig. 1.5). It seems that the additional licks reflect the
tendency of the animal not to interrupt the activity of the central timing network
between spout presentations. Although no overt licks are visible during transi-
tion between spouts, the timing network is probably not stopped but only dis-
connected from the intracycle pattern generator (Lennard & Hermanson, 1985)
and from the motor output. The noninterrupted activity of the pacemaker of
licking is indicated by good synchronization of the posttransition licks with the
pretransition licking.
The postulate that animals tend to maximize the value of their behavioral
output must be based on a complete cost-benefit analysis (Rachlin & Burkhard,
1978), taking into account not only the response-reinforcement ratio but also
FIGURE 1.4. Forced spout alternation. (1) First lick at the left spout. (2,3) Left spout is retracted and right spout advanced; the rat continues to lick
at the left wall opening, although it cannot contact the distant spout. (4) After two nonrewarded licks at the left opening the rat moves to the right
opening. (5) First lick at the now available right spout. (6) The right spout is retracted, but the animal continues to lick at the right wall opening.
FIGURE 1.5. Cumulative histogram of successive licks emitted during forced spout alternation at the left (L) and right (R) spouts which either
contain water (W) or are empty (E). Ordinate: Number of licks (vertical bar denotes 100 licks). Abscissa: Number of licks per spout presentation
(LPSP). Note that the two distributions were symmetric when both spouts contained water, but that LPSP approached 1 at the empty spout while
additional licks continued to be emitted at the other water-containing spout.
 Neurobiology of Memory 13

the difficulties connected with generation of the new response pattern. It seems
that rats find it acceptable to get only one reinforced lick out of five if they can
avoid in this way the need to switch the central timing network on and off. It is
also conceivable that the consummatory licks, resulting in transport of a small
amount (about 5 ^1) of water into the mouth cavity, require continued licking
in vacuo in order to induce swallowing, which usually follows after 3 to 5 licks
(Weijnen, Wouters, & Van Hest, 1984). This may explain why the instrumental
lick at the dry spout is not followed by additional licks, whereas the consum-
matory licks at the water-containing spout are.
This simple experiment demonstrates that the law of effect formulated on the
basis of behavioral observations exerts only limited control over specific physi-
ological mechanisms and must therefore be applied with great caution for the
interpretation of cellular phenomena.

MEMORY READOUT IS NOT IMPAIRED

BY COMPLETE RECONSTITUTION OF SENSORY INPUT

Finally, let us consider yet another anomaly in the neurobiology of memory

access, concerned with the problem of central stimulus representation. Although
it is not yet clear whether information encoding employs a hierarchical system
of neurons converging through progressively more complex feature detectors to
the so-called gnostic units (Konorski, 1967), distributed stimulus representation
by spatiotemporal patterns of activity in large neuronal populations (Freeman,
1975), or a combination of both principles (Perrett, Mistlin, & Chitty, 1987),
orderly mapping of the receptor surface on the appropriate brain centers is a
basic prerequisite for each of these alternatives. The fact that neurons are not
replaced in the adult brain is usually considered to be an organizational principle
supporting the invariance of such mappings. This is the reason why three gen-
erations of neurobiologists have been fascinated by Sperry's (1944) demonstra-
tion that the regenerating optic nerve axons in frogs and fish can find their
respective targets in the tectum and reestablish through appropriate synaptic
connections (Jacobson, 1978) apparently normal retinotectal mapping and
vision. Recent progress in transplantation of embryonal brain tissue makes it
possible to perform analogous experiments in mammals. McLoon and Lund
(1980) demonstrated that embryonal retina transplanted close to the superior
colliculus of adult rats retains its morphology and light sensitivity and that the
axons of the ganglion cells form synaptic contacts with the host tectum. Freed
(1983) used light stimulation of the grafted retina through an implanted light-
conducting rod to test the possible transmission of CS signals through this
ectopic input. Experiments of this type resemble Doty's (1969) analysis of con-
ditioning with the use of electrical stimulation of brain as the CS. There is little
doubt that the retinal graft can transmit information about presence or absence
of light to the host and allow at least some crude form of brightness discrimi-
nation learning.
14 NEUROBIOLOGY OF MEMORY

Whereas the exciting attempts to provide the rat with the additional Hatteria-
type parietal eye are in their infancy, another mammalian sensory system resem-
bles the optic nerve regeneration in amphibia more closely. Graziadei and
Monti-Graziadei (1978) convincingly demonstrated that the sensory neurons of
the olfactory epithelium in the nasal cavity are in a permanent state of renewal.
After several months they degenerate and are replaced by new neurons devel-
oping from the germinative layer of basal cells. Since the olfactory nerve is
formed by axons of sensory neurons, all fila olfactoria are completely exchanged
within a few months, as are their synaptic connections in the glomerular layer
of the olfactory bulb.
The regenerative power of the olfactory epithelium is better expressed after
transection of the olfactory nerve: all sensory neurons whose axons were cut
degenerate, but the newly formed sensory neurons send their axons through the
cribriform bone to the olfactory bulb again. After several months the sense of
smell is restored and old olfactory memories are reliably retrieved through new
sensory elements, nerve fibers, and synapses. Since the efficiency of this reorga-
nized input is comparable to that observed in case of the optic nerve-tectum
reconnection in frogs, olfactory nerve regeneration can serve as a convenient
model for the investigation of the molecular mechanisms governing reconnec-
tion of millions of axons to their respective targets.
There is an important difference between the two models, however. Whereas
in the case of the optic nerve the regenerating axons connect the surviving gan-
glion cells of retina with tectum, transection of the olfactory nerve leads to initial
degeneration and subsequent regeneration of the neuroepithelium so that appro-
priate connections must be formed between a substantially reduced and recon-
stituted population of input elements and the original target structure.
Olfactory nerve regeneration makes it possible to go still further: Graziadei
and Monti-Graziadei (1986), using rats and mice with olfactory bulbs removed,
showed that the olfactory axons can form the characteristic glomeruli and the
corresponding synapses in other forebrain structures extending into the bulbec-
tomy-created cavity (frontal lobes, oflactory tract, anterior olfactory nucleus).
Whether the synapses between primary olfactory fibers and nonolfactory brain
structures or higher olfactory centers are functional is a matter of controversy.
Wright and Harding (1982) reported good recovery of smell in bulbectomized
mice, but their results were questioned by Butler, Graziadei, Monti-Graziadei,
and Slotnick (1984), who found irreversible loss of smell in animals with mor-
phologically verified complete ablation of the olfactory bulb. Amemori, Soukup,
Valouskova, and Bures (1986) confirmed recovery of smell in bulbectomized
rats (Fig. 1.6). In most animals the positive results could be explained by rein-
nervation of small remnants of the olfactory bulb adjacent to the olfactory tract,
but several rats succeeded in the olfactory tests in spite of complete bulbectomy.
In such cases, intensive glomerulization was seen in the anterior olfactory
nucleus and frontal cortex or in embryonal grafts transplanted into the cavity
produced by bulbectomy, and units responding to olfactory stimuli were found
FIGURE 1.6. Recovery of smell after bulbectomy (BEC) and replacement of the olfactory bulb
with grafts of embryonal olfactory bulb (OBT) or neostriatum (StT). Ordinate: Average latency
(±SEM, n = 10) to find a food pellet buried under a 5-cm-deep layer of polystyrene balls.
Abscissa: Time after bulbectomy in weeks. C = unoperated control animals.
16 NEUROBIOLOGY OF MEMORY

in such nonbulbar brain structures. On the other hand, glomeruli and olfactory
responses were absent in the brains of rats failing in the olfactory test.
These results point to the possibility that regenerating olfactory axons can
bypass the first relay of the olfactory pathway and transmit olfactory signals to
higher olfactory centers. This possibility raises a number of fundamental ques-
tions at the molecular level (matching of presynaptic and postsynaptic compo-
nents of chemical transmission) and system level (relaying the information to
appropriate neural circuits), which should be examined by morphological, neu-
rochemical, electrophysiological, and behavioral methods.

CONCLUSION

The preceding examples of experimental evidence incompatible with the con-

ventional view of brain organization and function are certainly not unique. I am
sure that others could name even more striking anomalies which show that the
accepted explanations are incorrect in a number of specific situations. It is prob-
ably more than a coincidence that anomalous findings are encountered more
frequently at the system level of investigation than in cellular and behavioral
studies. This is understandable because the attempts to find physiological mech-
anisms for well-established behavioral phenomena led from the days of Pavlov
(1927) and Lashley (1929) to construction of conceptual brains whose putative
properties approached the properties of real brains at a much later date. The
most influential concept of this type, Hebb's (1949) model of neuronal plasticity,
decisively influenced present research into cellular and subcellular events medi-
ating simple forms of learning in Aplysia (Kandel, 1976) and Hermissenda (Far-
ley & Alkon, 1985) as well as long-term potentiation in the hippocampal slice
(Dingledine, 1984). We are clearly approaching the moment when at least some
instances of plasticity will be described as accurately as have been the generation
and conduction of nerve impulse (Woody, 1986).
Achievements at the molecular and cellular levels must not lead to the erro-
neous conviction that by disclosing the mechanisms of changes occurring at the
individual nodes of the neural network we may understand learning in general.
It must be stressed that learning is meaningful only at the level of the organism,
as an extension of the adaptive phenomena governing the evolution of species
to the needs of an individual. Stating that a cell has learned a CS-US contin-
gency, that it recognizes a complex visual stimulus, or that it controls delivery
of reward is an unwarranted ascription to a neuron of the properties of a system
of which it is a part. Progress at the molecular and cellular levels will not resolve
the enigma of learning without commensurate progress at the system level.
Unfortunately, the pace of system investigations has been dissappointingly slow
because the problems encountered here are necessarily much more complex
than at the cellular or the behavioral level. Nature has employed similar ele-
ments (neurons, synapses) and signals (nerve impulses, transmitters) to assem-
ble and operate the full scale of brain types, from the most primitive inverte-
 Neurobiology of Memory 17

brates to humans. And all these brains are used to perform a limited list of
essential behavioral programs (e.g., feeding, defense, mating, locomotion), obey-
ing the same fundamental principles.
To use a technical analogy, you can run the same BASIC program on all types
of computers and be sure that it is carried out by transmission of electrical
impulses through the computer's circuits, but there will be great differences in
the architecture of different computers and in the chips employed. Like com-
puters, all brains use similar gates and similar impulses but differ in the density
of gates, in their connectivity, and in the overall organization of the system.
Hardware-transparent study of behavior, espoused by Skinner (1938), has
greatly advanced our knowledge of higher level programming of brain function
and is now being supplemented by detailed knowledge of the processes taking
place at the individual nodes of the neural networks, but we are still far from
understanding how these elementary processes are organized to implement
behavior.
In many respects neuroscience faces the same difficulties as the technical dis-
ciplines simulating brain functions—artificial intelligence, robotics, computer
science—which have only recently discovered that efficient solution of their
problems requires not only faster gates, higher density of components, and
advanced programming languages, but mainly new principles of computer archi-
tecture, characterized by massive parallel processing, reconfigurable setups, and
fault-tolerant computation. This is best illustrated by the recent interest in neu-
rocomputers, which should emulate brain function by following biological prin-
ciples not in the construction of individual gates but in their organization, that
is, at the system level. Most needed now is a synthesis of different approaches
from molecular and cellular investigations to behavioral experiments and tech-
nical simulations, aimed at accelerating the progress of system research.

REFERENCES

Amemori, T., Soukup, T., Valouskova, V., & Bures, J. (1986). Embryonal brain grafts and
recovery of smell in bulbectomized rats. Physiologia Bohemoslovaca, 35, 519.
Barker, L. M., Best, M. R., & Domjan, M. (1977). Learning mechanisms in food selection.
Waco, Tex.: Baylor University Press.
Bellingham, W. P., & Lloyd, D. (1987). Injected flavor as a CS in the conditioned aversion
preparation. Animal Learning and Behavior, 15, 62-68.
Berger, B. D. (1970). Learning in the anesthetized rat. Federation Proceedings, 29, 749.
Bermudez-Rattoni, F., Rusiniak, K. W., & Garcia, J. (1983). Flavor-illness aversions: Potenti-
ation of odor by taste is disrupted by application of novocaine into amygdala. Behav-
ioral and Neural Biology, 37, 61-75.
Braveman, N. S., & Brostein, P. (Eds.). (1985). Experimental assessments and clinical appli-
cations of conditioned food aversions. Annals of the New York Academy of Sciences,
443.
Breland, K., & Breland, M. (1961). The misbehavior of organisms. American Psychologist, 16,
681-684.
18 NEUROBIOLOGY OF MEMORY

Brooks, V. B., & Thach, W. T. (1981). Cerebellar control of posture and movement. In V. B.
Brooks (Ed.), Handbook of Physiology: Vol. 2. The Nervous System (pp. 877-946).
Washington, D.C.: American Physiological Society.
Bures, J., & Buresova, O. (1981). Elementary learning phenomena in food selection. In G.
Adam, I. Meszaros, & E. I. Banyai (Eds.), Advances in Physiological Science: Vol. 17.
Brain and behavior (pp. 81-94). Budapest: Akademiai Kiado.
Bures, J., & Buresova, O. (1987). Conditioned taste aversion is neither elicited nor prevented
by deep anesthesia, anoxia or convulsions. Neuroscience, 22 (Suppl.), S517.
Buresova, O., & Bures, J. (1973). Cortical and subcortical components of the conditioned sac-
charin aversion. Physiology and Behavior, 11, 435-439.
Buresova, O., & Bures, J. (1977). The effect of anaesthesia on acquisition and extinction of
conditioned taste aversion. Behavioral Biology, 20, 41-50.
Buresova, O., & Bures, J. (1979a). The anterograde effect of ECS on the acquisition, retrieval
and extinction of conditioned taste aversion. Physiology and Behavior, 22, 641-645.
Buresova, O., & Bures, J. (1979b). Extinction of a newly acquired conditioned taste aversion:
Effect of gustatory CS administered under anesthesia. Behavioral Processes, 4, 329-
339.
Buresova, O., & Bures, J. (1981). Threshold hypothermia disrupting acquisition of conditioned
taste aversion and attenuation of neophobia in rats. Behavioral and Neural Biology,
31, 274-282.
Buresova, O., & Bures, J. (1984). Central mediation of the conditioned taste aversion induced
in rats by harmaline. Psychopharmacology, 34, 384-389.
Butler, A. B., Graziadei, P. P. C, Monti-Graziadei, G. A., & Slotnick, B. M. (1984). Neonatally
bulbectomized rats with new olfactory neocortical connection are anosmic. Neuro-
science Letter, 48, 247-254.
Danguir, J., & Nicolaidis, S. (1976). Impairments of learned aversion acquisition following par-
adoxical sleep deprivation in the rat. Physiology and Behavior, 17, 489-492.
Dingledine, R. (Ed.). (1984). Brain slices. New York: Plenum.
Doty, R. W. (1969). Electrical stimulation of the brain in behavioral context. Annual Review of
Psychology, 20, 299-320.
Farley, J., & Alkon, D. L. (1985). Cellular mechanisms of learning, memory and information
storage. Annual Review of Psychology, 36, 419-494.
Freed, W. J. (1983). Functional brain tissue transplantation: Reversal of lesion-induced rota-
tion by intraventricular substantia nigra and adrenal medulla grafts. With a note on
intracranial retinal grafts. Biological Psychiatry, 18, 1205-1267.
Freeman, W. J. (1975). Mass action in the nervous system. New York: Academic Press.
Garcia, J., Kimeldorf, D. J., & Koelling, R. A. (1955). A conditioned aversion towards saccha-
rin resulting from exposure to gamma radiation. Science, 122, 157-159.
Graziadei, P. P. C., & Monti-Graziadei, G. A. (1978). Continuous nerve cell renewal in the
olfactory system. In M. Jacobson (Ed.), Handbook of sensory physiology (Vol. 9, pp.
55-83). New York: Springer-Verlag.
Graziadei, P. P. C., & Monti-Graziadei, G. A. (1986). Principles of organization of the verte-
brate olfactory glomerulus: An hypothesis. Neuroscience, 19, 1025-1035.
Halpern, B. P. (1977). Functional anatomy of the tongue and mouth of mammals. In J. A. W.
M. Weijnen & J. Mendelson (Eds.), Drinking behavior (pp. 1-92). New York:
Plenum.
Hebb, D. O. (1949). The organization of behavior. New York: Wiley.
lonescu, E., & Buresova, O. (1977). Effects of hypothermia on the acquisition of conditioned
taste aversion in rats. Journal of Comparative and Physiological Psychology, 91,
1297-1307.
Jacobson, M. (1978). Developmental biology (2nd ed.). New York: Plenum.
Kandel, E. R. (1976). Cellular basis of behavior. San Francisco: Freeman.
Kiefer, S. W. (1985). Neural mediation of conditioned food aversions. In N. S. Braveman & P.
Bronstein (Eds.), Experimental assessments and clinical applications of conditioned
 Neurobiology of Memory 19

food aversions (pp. 100-109). (Annals of the New York Academy of Sciences, Vol.
443). New York: New York Academy of Sciences.
Konorski, J. (1967). Integrative activity of the brain: An interdisciplinary approach. Chicago:
University of Chicago Press.
Lashley, K. S. (1929). Brain mechanisms and intelligence: A quantitative study of injuries to the
brain. Chicago: University of Chicago Press.
Lennard, P. R., & Hermanson, J. W. (1985). Central reflex modulation during locomotion.
Trends in Neuroscience, 8, 483-486.
Malmo, H. P., Malmo, R. B., & Weijnen, J. A. W. M. (1987). Individual consistency and mod-
inability of lapping rates in rats: A new look at the variance-invariance question.
International Journal of Psychophysiology, 4, 111-119.
Mamedov, Z,., & Bures, J. (1984). Sensory feedback modulates the central pacemaker of licking
in rats. Neuroscience Letter, 45, 1-6.
Mamedov, Z., Hernandez-Mesa, N., & Bures, J. (1987). Inefficient licking during forced spout
alternation in rats: Violation of the law of effect? Physiology and Behavior, 39, 153-
160.
McLoon, S. C, & Lund, R. D. (1980). Specific projections of retina transplanted to rat brain.
Experimental Brain Research, 40, 273-282.
Millner, J. R., & Palfai, T. (1975). Metrazol impairs conditioned aversion produced by LiCl: A
time dependent effect. Pharmacology Biochemistry and Behavior 3, 201-204.
Pavlov, I. P. (1927). Conditioned reflexes. London: Oxford University Press.
Perrett, D. I., Mistlin, A. J., & Chitty, A. J. (1987). Visual neurons responsive to faces. Trends
in Neuroscience, 10, 358-364.
Rabin, B. M., & Rabin, J. S. (1984). Acquisition of radiation- and lithium chloride-induced
conditioned taste aversions in anesthetized rats. Animal Learning and Behavior, 12,
439-441.
Rachlin, H. C., & Burkhard, B. (1978). The temporal triangle: Response substitution in instru-
mental conditioning. Psychological Review, 85, 22-47.
Riley, A. L., & Tuck, D. L. (1985). Conditioned taste aversions: A behavioral index of toxicity.
Annals of the New York Academy of Sciences, 443, 272-292.
Roll, D. L., & Smith, J. C. (1972). Conditioned taste aversion in anesthetized rats. In M. E. P.
Seligman & J. L. Hager (Eds.), Biological boundaries of learning (pp. 93-102). New
York: Appleton-Century-Crofts.
Skinner, B. F. (1938). The Behavior of organisms. New York: Appleton-Century-Crofts.
Sperry, R. W. (1944). Optic nerve regeneration with return of vision in Anurans. Journal of
Neurophysiology, 7, 57-69.
Thorndike, E. L. (1911). Animal intelligence. Experimental studies. New York: Macmillan.
Weijnen, J. A. W. M., & Mendelson, J. (1977). Drinking behavior. New York: Plenum.
Weijnen, J. A. W. M., Wouters, J., & Van Hest, J. M. H. H. (1984). Interaction between drink-
ing and swallowing in the drinking rat. Brain, Behavior and Evolution, 25, 117-127.
Woody, C. D. (1986). Understanding the cellular basis of memory and learning. Annual Review
of Psychology, 37, 433-493.
Wright, J. W., & Harding, J. (1982). Recovery of olfactory function after bilateral bulbectomy.
Science, 216, 322-324.
This page intentionally left blank
I
Forms of Memory
This page intentionally left blank
Introduction
M I C H E L A GALLAGHER

It is not a real contradiction that the most creative scientists are fre-
quently just those, not only willing to accept the existence of disorder,
but also positively attracted to it. The evident reason is that the recog-
nition of disorder is an opportunity, and in fact a necessary preliminary,
for the creative act of ordering.
G. GAYLORD SIMPSON, Principles of Animal Taxonomy (1961)

Memory, defined in general terms, refers to the effects of experience that are
manifest at a later time. This capacity permits animals to acquire, retain, and
use a great variety of information. Thus memory can be captured by an encom-
passing definition, but can it be profitably studied as a unitary function? Alter-
natively, does the phenomenon of memory represent a diversity of systems or
mechanisms that have evolved to execute different essential types of memory?
In this introduciton it is only fair to acknowledge that there is no general con-
sensus that different forms of memory should be distinguished; dissenting views
can be found (e.g., Craik, 1983).
In discussing this question, Tulving (1983)—a proponent of the concept that
memory is not unitary—uses an analogy in which he compares the problem of
understanding memory with that of understanding locomotion. An encompass-
ing definition of locomotion can be formulated along the lines of a change in the
position of an animal from one location to another location. From this point,
however, any further attempt to describe, or model, or specify the mechanisms
used for locomotion requires identification of some type of this general phenom-
enon. Indeed, a conspicuous and amazing variety of systems/mechanisms have
evolved for the purpose of locomotion. Animals swim, jump, burrow, or fly
from one place to another. A study of locomotion that provides an appreciation
of the behavioral properties, the neural systems, or the cellular mechanisms
involved requires an acknowledgment that locomotion is not a unitary phenom-
enon with a single set of governing principles and underlying mechanisms. If
isolated empirical observations are to increase our understanding of the phe-
nomenon, it is also important to know whether different specific instances of

23
24 FORMS OF MEMORY

locomotion belong to some larger classes that share common features. However,
a general theory of locomotion would not be nourished by many empirical facts
that apply overall.
Properties that distinguish different forms of memory are conspicuous to
many who study memory. There is a growing vocabulary to describe types of
memory. Distinctions are made based on the durability of information that is
acquired. Short-term, long-term, and intermediate forms of memory often figure
in contemporary discussions of memory. The content of information that is pre-
served has also been used as a basis for distinguishing different forms of mem-
ory; there is, for example, memory for facts, or skills, or spatial information.
Other distinctions are based on operational characteristics such as working and
reference memory.
Different information about memory—its varieties and organization—comes
from studies with a biological focus. In one such line of investigation, different
forms of memory are characterized by brain damage, which alters performance
on some types of memory tasks while sparing performance on others. It is well
known, for example, that relatively circumscribed brain damage in diencephalic
and temporal lobe structures in humans is associated with an amnesic syn-
drome. Perhaps the most striking realization of the last decade or so is that these
amnesic patients possess some apparently intact memory capacities: they per-
form as well as normals in retaining certain motor, perceptual, or cognitive skills
and show residual effects of exposure to items such as individual words, a phe-
nomenon known as the repetition priming effect (see Squire, 1987, for recent
discussion).
Other work with a neurobiological focus has indicated that mechanisms for
memory may be distinguished at a cellular level. The short- and long-term rep-
resentations of nonassociative memory for sensitization in Aplysia are likely to
require the engagement of qualitatively different intracellular processes (Mon-
tarolo et al., 1986). In addition, it is apparent that changes in synaptic function
can be controlled by distinct mechanisms in different neural circuits. From the
properties of the circuitry, to the identity of neurotransmitters/neuromodulators
that regulate neural plasticity, to the species of conductances or nature of mor-
phological changes that occur, there is at least a hint that the cast of characters
used to preserve biological records of experience is highly diverse.
The nature of much thought and many empirical studies on memory at dif-
ferent levels of analysis, both behavioral and neurobiological, indicates an
expectation that some taxonomy of memory is likely to emerge. At this point, it
is perhaps worthwhile to ask whether the criteria applied in this endeavor are
useful and/or valid.
In a recent review of this general topic, Sherry and Schacter (1987) provide
some guidance about the defining features of a memory system. They propose
that a memory system refers to the acquisition, retention, and retrieval mecha-
nisms for information that are governed by distinctive rules of operation. Mul-
tiple forms of memory thus emerge when two or more systems are "character-
ized by fundamentally different rules of operation." By this definition, the type
of information—the content of memory—is not alone a sufficient basis for iden-
 Introduction 25

tification of a distinct form of memory. The handling of information by diiferent

bits of neural circuitry or different regions of the brain is likewise an insufficient
basis for making critical distinctions between types of memory. Indeed, different
operations might well be performed by systems that share at least some neural
components.
An emphasis on rules of operation—as distinct from content-based or neu-
roanatomically sequestered systems of memory—emerges in the chapters that
follow. It is also evident from some of these presentations that behavioral and
neurobiological data may converge to provide a mutually supportive basis for
making distinctions. The opening chapter is a developmental analysis for the
emergence of nonassociative learning in an invertebrate in which Carew and
colleagues provide behavioral and neurobiological evidence for distinctive pro-
cesses underlying dishabituation and sensitization, forms of nonassociative
learning that had previously been subsumed by a common process. Morris then
offers a strong argument for examination of the rules of operation at behavioral
and neurobiological levels; in the case of the latter a particular emphasis is
placed on the information that is likely to be embedded in the neural architec-
ture used to implement the storage of information. By this view, knowledge of
neural circuitry is not only a vehicle for discovering sites of plasticity that can
be subjected to cellular analysis, but also places interesting and important con-
straints on the nature of information that is processed and preserved within it.
Later chapters in this volume further discuss processing in cortical circuitry and
characteristic algorithms used in computational models of such neural
architecture.
The chapters by Holland and Johnson share a common thread in viewing
memory as a less segregated psychological/neurobiological function. Johnson's
discussion of a multiple-entry modular memory system provides a cognitive
framework for examining the operating characteristics for information process-
ing in normal humans and amnesic patients. Holland shows that Pavlovian con-
ditioning provides a powerful framework for examining even representation in
memory in laboratory animals. These chapters remind us that neurobiological
accounts of memory depend critically on study at a behavioral level. Indeed,
with current rapid technical advances in the field of neurobiology, it is possible
that progress in understanding the neurobiology of memory will be mainly lim-
ited by the rate of developing an appropriate behavioral analysis.
These chapters cover an enormous range of research, from behavioral studies
of both laboratory animals and human subjects to studies of circuits and cellular
mechanisms in mammals and invertebrates. The problem of developing a tax-
onomy of memory will provide a challenging agenda for many years to come.

REFERENCES

Craik, F. I. M. (1983). On the transfer of information from temporary to permanaent memory.

Philosophical Transactions of the Royal Society (London) 302, 341, 359.
26 FORMS OF MEMORY

Montarolo, P. G., Goelet, P., Castellucci, V. R, Morgan, J., Kandel, E. R., & Schacher, S.
(1986). A critical period for macromolecular synthesis in long-term heterosynaptic
facilitation in Aplysia. Science, 234, 1249-1254.
Sherry, D. F., & Schacter, D. L. (1987). The evolution of multiple memory systems. Psycholog-
ical Review, 94, 439-454
Squire, L. R. (1987). Memory and brain. New York: Oxford University Press.
Tulving, E. (1983). Elements of episodic memory. New York: Oxford University Press.
2

The Development of Learning

and Memory in Aplysia
T H O M A S J. CAREW
E M I L I E A. M A R C U S
T H O M A S G. N O L E N
C A T H A R I N E H. R A N K I N
MARK STOPPER

Few issues have captured the imagination of psychologists more than the fun-
damental question of how, over a lifetime, animals are able to acquire a vast
amount of information, store it in their brains, and retrieve it in one form or
another for use at a later time. The depth and complexity of this question are
reflected in the many rich and diverse approaches psychologists and neurobiol-
ogists have taken in analyzing mechanisms of learning and memory. As exem-
plified by the contributors to this volume, these approaches range from the anal-
ysis of membrane currents and molecular events in individual neurons to the
investigation of computational strategies in highly complex neural networks. A
unifying theme underlying all these approaches is a fascination with the remark-
able capacity of animals to learn and remember.
One powerful approach to the analysis of neuronal mechanisms of learning
and memory is the use of simple invertebrate animals whose nervous systems
are well suited for detailed cellular and molecular studies (for reviews, see Byrne,
1987; Carew & Sahley, 1986; Hawkins, Clark, & Kandel, 1987; Mpitsos & Luko-
wiak, 1985). Among these invertebrate preparations, the marine mollusc Aplysia
californica has proved extremely useful, since the animal exhibits a variety of
types of learning that range in complexity from nonassociative forms such as
habituation, dishabituation, and sensitization to associative forms such as clas-
sical and operant conditioning. Moreover, these various forms of learning can
be retained on a time scale from short-term memory (lasting hours) to long-term
memory (lasting days and weeks). Finally, it has been possible to specify in some
detail the synaptic, biophysical, and molecular events that contribute to the
expression of these various forms of learning and memory (Byrne, 1987; Carew,
1987; Hawkins et al., 1987).

27
28 FORMS OF MEMORY

In our laboratory in recent years we have extended the analysis of learning

and memory in Aplysia by combining it with a developmental analysis, aimed
at examining the emergence and assembly of different forms of learning and
memory during ontogeny. The rationale for this approach is that by examining
the assembly of different forms of learning as they sequentially appear during
development, it may be possible to obtain unique insights into underlying cel-
lular and molecular mechanisms. In this chapter we will illustrate our approach
by showing how a developmental analysis in Aplysia has revealed multiple com-
ponents underlying simple forms of nonassociative learning. These different
components can be dissociated by examining their developmental timetables,
which turn out to be quite different. This in turn has afforded us the experimen-
tal advantage of examining early emerging forms of learning in developmental
isolation from later emerging forms. In this way we have not only been able to
dissociate different components of learning, but we have also been able to reveal
previously unappreciated behavioral and cellular processes. Finally, the devel-
opmental dissociation of different components of learning in juvenile Aplysia
prompted us to examine adult animals, where we found the same clearly dis-
sociable components. Thus an analysis of the ontogenetic assembly of learning
can also provide important insights into the final phenotypic expression of learn-
ing in the adult.

DEVELOPMENT OF HABITUATION AND DISHABITUATION

IN THE SIPHON WITHDRAWAL REFLEX

In adult Aplysia the gill and siphon withdrawal reflex exhibits both nonassocia-
tive and associative forms of learning (Carew, 1987; Hawkins et al., 1987). The
three different forms of nonassociative learning are illustrated in Figure 2.1.
Habituation refers to a decrease in response amplitude occurring as a function
of repeated stimulation; dishabituation describes the facilitation of a habituated
response by the presentation of a strong or novel stimulus; and sensitization
refers to the facilitation of a nondecremented response by a similar strong or
noxious stimulus.
In our developmental analysis we focused on the juvenile stages of develop-
ment because it is during the juvenile period that the gill and siphon first emerge
(Kriegstein, 1977a; Rankin, Stopfer, Marcus, & Carew, 1987). The juvenile
stages we analyzed are shown in Figure 2.2. In the first set of experiments (Ran-
kin & Carew, 1987a) we investigated whether habituation and dishabituation
emerge at the same developmental stage or whether there is a difference in their
developmental timetables. We found that habituation of siphon withdrawal is
present in very early stages of juvenile development. However, it exists in a rel-
atively immature form. In the youngest animals, significant within-session
habituation was produced only by very short interstimulus intervals (ISIs). As
the animals continued to develop, significant habituation was produced in
 The Development of Learning and Memory in Aplysia 29

FIGURE 2.1. Three forms of nonassociative learning: habituation, dishabituation,

and sensitization.

response to progressively longer ISIs. Thus in stage 9, significant habituation

occurred to only a 1-sec ISI, but not to 5- or 10-sec ISIs; in stage 10, to both 1-
and 5-sec ISIs but not to a 10-sec ISI; and in stage 11, to all three ISIs. Moreover,
even in later juvenile stages, further maturation of habituation was evident. For
example, in response to a 30-sec ISI, stage 12 animals show significantly less
habituation than do adult animals to this same ISI. Figure 2.3 summarizes the
development of habituation: the magnitude of habituation (as an average per-
centage habituation score) in response to each ISI tested is expressed across juve-
nile developmental stages. We examined habituation to comparable ISIs across
developmental stages and found a clear and significant developmental trend in
the magnitude of habituation exhibited to a variety of ISIs (ranging from 1 to
30 sec) throughout most of the juvenile life of Aplysia.
Although habituation was present at the earliest stage of development studied,
dishabituation was not. Specifically, dishabituation was absent in stage 9, first
appearing several days later in stage 10. This can be seen in Figure 2.4, which
provides a comparison of habituation (to a 1-sec ISI) and dishabituation (to tail
shock) in stages 9, 10, and 11. Whereas dishabituation in response to tail shock
was significant in stages 10 and 11, it was completely absent in stage 9. Thus
habituation is present in early juvenile development and dishabituation emerges
in a distinct later stage (stage 10) approximately one week later.
FIGURE 2.2. The Aplysia life cycle has been divided into five major phases: embryonic (lasting about 10 days), planktonic (lasting
about 30 days), metamorphic (lasting about 3 days), juvenile (lasting about 120 days), and adult. We have focused our attention on the
juvenile phase, which can be further subdivided into four stages based on discrete morphological criteria. The approximate duration of
each stage (in days) is shown.
 The Development of Learning and Memory in Aplysia 31

FIGURE 2.3. The mean percent habituation (100% = complete response decrement) to four
different ISIs (1, 5, 10, and 30 sec) as a function of developmental stage. There was a significant
increase in habituation (i.e., a positive slope) as a function of age for all four ISIs. Thus progres-
sively older animals were capable of expressing habituation to progressively longer ISIs. (From
Rankin and Carew, 1987a)

DEVELOPMENT OF SENSITIZATION IN THE SIPHON

WITHDRAWAL REFLEX

Until recently, dishabituation and sensitization had commonly been considered

to reflect a single underlying facilitatory process. This view was supported by the
general observation that both nondecremented and decremented responses were
simultaneously facilitated by the presentation of a single strong or noxious stim-
ulus. The most parsimonious explanation for this kind of result was that a nox-
ious stimulus initiated a general arousal process that was widespread in the ner-
vous system, facilitating habituated and nonhabituated responses alike (Carew,
Castellucci, & Kandel, 1971; Groves & Thompson, 1970). Although reasonable,
this explanation did not rule out the possibility that dishabituation and sensiti-
zation could reflect separate facilitatory processes that are activated in parallel
by a strong stimulus. In fact, direct cellular evidence for two separate facilitatory
processes has recently been provided by Hochner, Klein, Schacher, and Kandel
(1986a, 1986b), who suggested that dishabituation and sensitization in adult
Aplysia may be produced, at least in part, by different cellular mechanisms.
One way to behaviorally address the question of whether dishabituation and
sensitization reflect a unitary process is to determine when they emerge during
development. If a single process is involved, both forms of learning should be
expressed at the same time ontogenetically. Alternatively, if more than one pro-
32 FORMS OF MEMORY

FIGURE 2.4. Summary of habituation and dishabituation for stages 9,10, and 11. Summarized
data are for a 1-sec ISI. The hatched bar shows the initial response (100%). The clear bar shows
habituation (the mean of the last trial for the habituation series). The black bar shows dishabituation
(the mean of the first trial following tail shock). Older animals showed greater habituation to a 1-
sec ISI than did younger animals. Moreover, there was no dishabituation in stage 9, whereas there
was significant dishabituation in stages 10 and 11. (From Rankin and Carew, 1987a)

cess is involved, it might be possible to separate them developmentally. To

address this question Rankin and Carew (1988) examined the effects of tail
shock on both decremented and nondecremented responses in developmental
stages 11, early 12, and late 12. In each stage two groups of animals were exam-
ined. One group received dishabituation training (habituation followed by tail
shock); the other received sensitization training (several baseline stimuli were
delivered at an ISI too long to produce response decrement, followed by a tail
shock identical to that of the dishabituation group). Both groups were then
tested for response facilitation resulting from tail shock. Our results showed (1)
that sensitization emerges between early and late stage 12 of juvenile develop-
ment and (2) that tail shock actually produces inhibition of the reflex response
in young animals prior to the development of sensitization.
 The Development of Learning and Memory in Aplysia 33

Sensitization Emerges in Late Stage 12 of Juvenile Development

The results of dishabituation and sensitization training for each of the stages
examined are summarized in Figure 2.5, which shows the mean response ampli-
tudes immediately prior to and immediately following tail shock for each of the
three developmental stages studied. During habituation trials, animals at each
stage showed significant response decrement as a result of repeated stimulation
(not shown). In all stages habituated animals showed significant facilitation of
response amplitude following tail shock. Thus, confirming previous results
(Rankin & Carew, 1987a), dishabituation was present in each of the develop-
mental stages examined. In contrast, the results for animals that received sen-
sitization training showed a strikingly different pattern: sensitization was com-
pletely absent in stage 11 and early stage 12. In these stages the same animals
that exhibited dishabituation lacked sensitization. Moreover, sensitization was
absent regardless of the intensity of siphon stimulation used to elicit the reflex

FIGURE 2.5. Differential emergence of dishabituation and sensitization during juvenile devel-
opment. Dishabituation: Significant habituation is exhibited in all developmental stages: compare
initial response (hatched bars) to final habituated responses (clear bars). Moreover, significant
dishabituation is also evident in all stages: compare postshock responses (black bars) with habit-
uated responses. Data are expressed as a percentage of initial response. Sensitization: In contrast
to dishabituation, sensitization is not evident until late stage 12; compare preshock responses (clear
bars) to postshock responses (black bars). Significant response facilitation is seen only in late stage
12 animals. Data are expressed as a percentage of mean preshock responses. (From Rankin and
Carew, 1988)
34 FORMS OF MEMORY

or the intensity of tail shock used to attempt to produce sensitization (ranging

from weak to very strong). In fact, as we will discuss later, in these early stages
in which sensitization was absent, tail shock produced a modest but significant
depression in response amplitude. However, in late stage 12, tail shock did
indeed produce sensitization of the reflex (Fig. 2.5). Thus these experiments
showed that sensitization emerges between early and late stage 12 of juvenile
development.

Tail Shock Produces Inhibition of Reflex Responding Before the Emergence

of Sensitization
In examining the developmental emergence of sensitization, we assessed the
effects of tail shock on nondecremented responses in three juvenile stages: stage
11, early stage 12, and late stage 12. This analysis revealed an unexpected result.
In early developmental stages (stage 11 and early stage 12) a modest but reliable
inhibitory effect of tail shock was apparent. Then, between early and late stage

FIGURE 2.6. Developmental transition from inhibition to facilitation of nondecremented re-

sponses following tail shock. A composite facilitation ratio was computed for each stage. In stage
11 and early stage 12, tail shock produces significant reflex inhibition. In contrast, in late stage
12, tail shock produces significant facilitation. (From Rankin and Carew, 1988)
 The Development of Learning and Memory in Aplysia 35

12, there was a dramatic transition after which tail shock could produce its facil-
itatory effect. This transition is illustrated in Figure 2.6, in which the data are
shown as a facilitation ratio for each developmental stage by expressing the first
response following tail shock as a percentage of the average preshock responses.
In any single group of animals the inhibitory effect of tail shock was modest.
However, it was quite consistent: in six out of six different groups of stage 11
and early stage 12 animals the first postshock response was lower than the mean
of the preshock responses. Thus significant response inhibition was seen in early
stages, whereas significant facilitation was seen in the later stage (Rankin &
Carew, 1988). The functional significance of this novel inhibitory process is an
interesting question for further exploration.
Figure 2.7 summarizes the behavioral results for all of the developmental
stages we examined. To permit a comparison of the magnitude of both disha-

FIGURE 2.7. Summary of the emergence and maturation of dishabituation and sensitization.
Data are expressed in terms of a mean difference score. (From Rankin and Carew, 1988)
36 FORMS OF MEMORY

bituation and sensitization across different stages, the data are expressed in
terms of a mean difference score, which was obtained by computing the differ-
ences in amplitude between the first response following tail shock and the mean
preshock responses. There was a significant developmental increase in both dis-
habituation and sensitization. In dishabituation there were no signficant differ-
ences between stage 11 and early stage 12; however, there was a significant
increase in late stage 12. The same was true for sensitization; there was no sig-
nificant difference between stage 11 and early stage 12 (sensitization was absent
in both), but there was significant facilitation following tail shock in late stage
12, reflecting the expression of sensitization for the first time.
The results shown in Figure 2.7 permit several interesting conclusions. First,
although dishabituation is present at all developmental stages examined, sensi-
tization does not emerge until sometime between early and late stage 12 in juve-
nile development. Second, inspection of the data on the effects of sensitization
training shows that the inhibitory effects of tail shock appear to diminish
between stage 11 and early stage 12, and in parallel with this reduction of the
inhibitory effect there is a concomitant modest increase in the magnitude of dis-
habituation. It is not clear from these results whether the inhibitory process is
truly diminishing during development and/or whether an independent facilita-
tory process is beginning to emerge in early stage 12, with the net effect of reduc-
ing the inhibition expressed at this stage.
Since both the facilitatory process involved in dishabituation and the inhibi-
tory process are produced by tail shock, it is possible that the inhibitory process
competes with dishabituation. In fact, evidence has recently been obtained in
stage 11 juvenile Aplysia (Rankin & Carew, 1987b, 1989) that weak tail shock
(which produced no inhibition of nondecremented responses) produces signifi-
cant dishabituation, whereas strong tail shock (which produced strong inhibition
of nondecremented responses) produces no dishabituation, suggesting that the
inhibition produced by strong tail shock may directly compete with dishabitu-
ation. Moreover, it is striking that at the time in development when the effects
of tail shock on nondecremented responses show a transition from inhibition to
facilitation (i.e., sensitization emerges), the magnitude of dishabituation also sig-
nificantly increases. It is likely that this increase is due to the sensitization pro-
cess adding to dishabituation and/or to the reduction of the inhibitory process.
Our observations during development suggest the hypothesis that the process
of dishabituation may be more complex than previously thought (see, e.g.,
Carew et al, 1971; Carew & Kandel, 1974; Groves & Thompson, 1970). Specif-
ically, the magnitude of dishabituation may ultimately be determined by the
interaction of three separate processes: (1) the dishabituation process itself; (2)
the inhibitory process that may subtract from dishabituation (see also Marcus,
Nolen, Rankin, & Carew, 1987; Rankin & Carew, 1987b); and (3) the sensiti-
zation process that may add to dishabituation (see also Hochner et al., 1986a,
1986b). These behavioral results reveal striking differences in the development
of dishabituation and sensitization that are not compatible with the notion that
 The Development of Learning and Memory in Aplysia 37

dishabituation and sensitization are a single process; rather they suggest that
they are, at least in part, separate processes.

THE DEVELOPMENT OF SENSITIZATION IN THE ESCAPE

LOCOMOTION SYSTEM

In the work presented thus far, the development of learning has been analyzed
in the siphon withdrawal reflex. Having established the timetable for the emer-
gence of sensitization in this reflex system, we then asked if the emergence of
sensitization was due to the development of a process restricted to the siphon
withdrawal circuit or to a more general change in the whole animal. One way to
examine this question is to analyze the development of sensitization in other
response systems; thus we turned our attention to escape locomotion, a dramatic
stereotypic response in which an animal rapidly locomotes away from a source
of noxious stimulation. This response offers a variety of advantages for a devel-
opmental analysis: (1) it is expressed very early in development (as early as stage
8; unpublished observations); (2) it is simple to quantify in an unrestrained ani-
mal (Hening, Walters, Carew, & Kandel, 1979; Walters, Carew, & Kandel,
1979); (3) the response in the adult is known to be modulated by several different
forms of both nonassociative and associative learning (Carew, Walters, & Kan-
del, 1981; Walters, Carew, & Kandel, 1981); and (4) the neural circuitry under-
lying escape locomotion in the adult has been examined in some detail (Hening
etal, 1979;Jahan-Parwar&Fredman, 1978a, 1978b).
We examined the development of escape locomotion in juvenile stages 10, 11,
and early and late 12, as well as in adults (Stopfer & Carew, 1988). Juvenile
animals were observed through a stereomicroscope in appropriately scaled
open-field test chambers (Fig. 2.8). There were three phases to the experiment:
(1) pretesting, (2) training, and (3) testing. In the pretest phase, animals first
received two weak pretest stimuli (mild tail shock), each followed by an obser-
vation period during which the distance traveled by the animal was recorded.
For each developmental stage two groups of animals were run: a sensitization
group, which received sensitization training (strong electric shocks to the tail);
and a control group, which received another weak test stimulus. Following train-
ing, in the test phase, both groups received another two test stimuli, identical to
pretest.
In stage 10, stage 11, and early stage 12 there were no significant differences
between animals that received sensitization training and animals that received
the control procedure; that is, there was no sensitization in these early juvenile
stages. However, in late stage 12 animals there was significant sensitization as a
result of strong tail shock (Fig. 2.9). Thus sensitization of escape locomotion
appears to emerge between early and late stage 12 of juvenile development.
The emergence of sensitization of escape locomotion in late stage 12 is strik-
ingly similar to the timetable for sensitization in the siphon withdrawal reflex.
FIGURE2.8. Experimental procedure for studying escape locomotion. (A)Testarena. Different-
sized open-field test arenas were appropriately scaled for each developmental stage. Juvenile stages
10 to early 12 were observed through an adjustable stereomicroscope which could be kept centered
over the animal. Escape locomotion was quantified by the number of boxes entered during an
observation period. (B) Experimental procedure. Both experimental and control groups received
two mild pretest stimuli and the distance traveled after each stimulus was recorded. A second
experimenter then delivered a training stimulus to each animal. All animals were then tested with
two mild stimuli (test) identical to the pretest. Each animal contributed a single score, the mean
test response expressed as a percentage of the mean pretest response. (From Stopfer and Carew,
1988)
 The Development of Learning and Memory in Aplysia 39

FIGURE 2.9. Sensitization of escape locomotion (A) emerges in the same juvenile stage as
sensitization of siphon withdrawal (B). Data are expressed as the percentage of experimental
animals within each stage exhibiting test scores greater than their prescores. The dashed line through
the 50% level represents the percentage of animals that would be expected to show an increase in
responding on the basis of chance alone. A clear developmental trend is exhibited, indicating that
sensitization of both escape locomotion and siphon withdrawal emerges in late stage 12 of juvenile
development.

Figure 2.9 compares the development of sensitization in escape locomotion with

that of the siphon withdrawal reflex. The data are expressed as the percentage of
animals in the experimental groups that exhibited sensitization (i.e., mean
postshock score higher than their mean preshock scores). The dashed line
through 50% indicates the proportion of preparations that would be expected to
show an increase in response amplitude by chance alone. In both siphon with-
drawal and escape locomotion there is no significant increase above the chance
level until late stage 12 of juvenile development, when significant sensitization
emerges in both response systems.
The fact that sensitization emerges within the same developmental stage in
the escape locomotion system and the siphon withdrawal reflex is additionally
interesting because the siphon withdrawal reflex and escape locomotion repre-
sent very different classes of response systems: siphon withdrawal is a graded
response that involves a restricted effector organ system (the mantle organs: gill,
siphon, and mantle shelf), whereas escape locomotion involves a widespread
effector organ system (involving virtually the entire body, including the head
40 FORMS OF MEMORY

and neck, the whole foot, the bilateral body walls, and the two parapodia). In
addition, the two response systems have very different underlying neural cir-
cuits: the gill and siphon withdrawal reflex is mediated by a relatively simple
circuit located in the abdominal ganglion (for review, see Hawkins et al., 1987),
whereas escape locomotion is mediated by more complex circuitry involving the
coordination of triggering, oscillatory, and effector circuits in the cerebral, pleu-
ral, and pedal ganglia (Hening et al., 1979; Jahan-Parwar & Fredman, 1978a,
1978b). Thus despite several differences in ontogeny, response topography, and
underlying circuit complexity, both response systems begin to express sensiti-
zation within the same developmental stage.

DEVELOPMENT OF THE CELLULAR ANALOGUES OF HABITUATION,

DISHABITUATION, AND SENSITIZATION

The developmental separation of different learning processes that we have

observed in the siphon withdrawal reflex ofAplysia affords the opportunity to
examine the unique contributions of different cellular and molecular mecha-
nisms to each form of learning. A first step in this analysis is to establish that
cellular analogues of these forms of learning can be identified and analyzed in
the central nervous system of juvenile Aplysia at the same stages of development
that these forms of learning are first behaviorally expressed. Although the neural
circuit in the abdominal ganglion that controls siphon and gill withdrawal is
relatively well understood in adult Aplysia, it has not yet been mapped in juve-
nile animals. However, the giant mucus motor neuron R2, also found in the
abdominal ganglion, can provide an excellent system in which to explore cellular
analogues of learning. This neuron is readily identifiable as a unique individual
throughout juvenile development (Kriegstein, 1977b). In addition, R2 receives
afferent input from the siphon (Kandel & Tauc, 1965; Rayport, 1981). Thus R2
can serve as a cellular vantage point to monitor plastic changes that emerge in
the reflex pathways for gill and siphon withdrawal.
The development of cellular analogues of habituation and dishabituation in
R2 was first examined by Rayport (1981) and by Rayport and Camardo (1984).
They found that the analogue of habituation (synaptic decrement) was present
at a very early stage in juvenile development, stage 9, and that the analogue of
dishabituation (facilitation of decremented synaptic potentials) emerged in stage
10, about one week later. These findings are consistent with the behavioral
observations of Rankin and Carew (1987a) that habituation in the siphon with-
drawal reflex was present in stage 9 and dishabituation emerged in stage 10.
Since behavioral studies indicated that sensitization emerged during stage 12
of juvenile development, Nolen and Carew (1988) focused on this stage for their
studies on the developmental emergence of the cellular analogue of sensitiza-
tion, the facilitation of nondecremented excitatory postsynaptic potentials
(EPSPs). The physiological preparation is shown in Figure 2.10. While recording
intracellularly from R2, stimulation of the siphon skin was mimicked by brief
 The Development of Learning and Memory in Aplysia 41

FIGURE 2.10. Physiological preparation for the examination of the development of the cellular
analogues of dishabituation and sensitization. Intracellular electrodes were used to record the
membrane potential (Vm) of the identified mucus motor neuron R2 in the abdominal ganglion of
stage 12 animals. In late stage 12 animals, a second electrode (/,,„) was used to hyperpolarize the
cell. Afferent input from the siphon was activated by a brief shock to the siphon nerve, which
evoked a complex EPSP in R2. Input from the tail pathway was activated by a train of brief shocks
to the right connective. (From Nolen and Carew, 1988)

electrical pulses delivered to the siphon nerve, which evoked a complex EPSP
in R2. To mimic tail shock, a train of electrical pulses was delivered to the pleu-
ral-abdominal connective (which carries the input from the tail to the abdom-
inal ganglion). The analogue of sensitization was examined by analyzing the
ability of connective stimulation to facilitate nondecremented complex synaptic
potentials in R2. In behavioral studies stage 12 had been divided into two sub-
stages, early stage 12 and late stage 12; sensitization was absent in early stage 12
but present in late stage 12 (Rankin & Carew, 1988). To permit greater resolu-
tion of temporal emergence of the analogue of sensitization for the cellular anal-
yses, stage 12 was divided into three substages: early, mid-, and late stage 12.
As would be predicted on the basis of the behavioral findings on the devel-
opment of sensitization in both the siphon withdrawal reflex and escape loco-
motion, the analogue of sensitization was clearly present in late stage 12 animals.
42 FORMS OF MEMORY

Specifically, connective stimulation produced significant facilitation of nonde-

cremented EPSPs. Moreover, in mid-stage 12 animals, connective stimulation
also produced clear facilitation of nondecremented EPSPs. Thus the analogue of
sensitization was present in both late and mid-stage 12 of juvenile development
(Fig. 2.11). However, in contrast to the older substages, connective stimulation
produced no significant facilitation of nondecremented EPSPs in early stage 12.
Thus the analogue of sensitization was absent in this early stage. In fact, at this
stage the connective stimulation actually produced significant depression of PSP
amplitude (Fig. 2.11). These cellular results therefore show a clear developmen-
tal trend in the emergence of the cellular analogue of sensitization, which sup-
ports the behavioral observation that sensitization is absent in early stage 12 of
juvenile development (Fig. 2.6; and Rankin & Carew, 1988).
Since the emergence of the analogue of sensitization occurs between early and

FIGURE 2.11. Emergence of the cellular analogue of sensitization. The median difference score
in EPSP amplitude (percent post minus pre) in R2 is shown for each developmental substage.
While significant facilitation was present in middle and late stage 12, significant inhibition was
present in early stage 12. (From Nolen and Carew, 1988)
 The Development of Learning and Memory in Aplysia 43

FIGURE 2.12. Comparison of the effects of connective stimulation in early and mid-stage 12.
The amount of facilitation of the first test response after connective stimulation is plotted for each
preparation. All mid-stage 12 preparations (clear bars) exhibited facilitation, whereas all but one
early stage 12 preparation showed depression (shaded bars). There was a highly significant differ-
ence between the two substages. (From Nolen and Carew, 1988)

middle stage 12, a detailed comparison of the effects of connective stimulation

in these two developmental substages is instructive. Such a comparison is seen
in Figure 2.12, in which the amplitude of the test EPSPs following connective
stimulation is expressed as a percentage of its amplitude before connective stim-
ulation. There was a clear and significant difference between the two substages:
whereas all mid-stage 12 preparations exhibited facilitation, all but one early
stage 12 preparation exhibited depression; and there was virtually no overlap
between the groups.
In summary, there are two striking parallels between the development of
behavioral sensitization and the development of its cellular analogue. First, the
cellular analogue of sensitization (produced by activation of the tail pathway)
emerges in the same late juvenile stage of development as the appearance of
behavioral sensitization (Rankin & Carew, 1988). Second, the cellular analysis
of sensitization revealed an inhibitory process produced by connective stimu-
lation similar to the behavioral depression produced by tail shock in animals
prior to the emergence of behavioral sensitization (Figs. 2.6 and 2.7); specifically,
prior to the emergence of the cellular analogue of sensitization, activation of the
44 FORMS OF MEMORY

tail pathway produced significant depression of the afferent reflex input. Thus
there is a close temporal correspondence between the development and matu-
ration of behavioral sensitization and a cellular representation of that process in
the CNS of juvenile Aplysia.

DISHABITUATION, SENSITIZATION, AND INHIBITION CAN BE

BEHAVIORALLY DISSOCIATED IN ADULT APLYSIA

The developmental studies we have described thus far show that dishabituation
and sensitization, as well as a novel inhibitory process, can be dissociated behav-
iorally in juvenile animals. Moreover, cellular analogues of each of these behav-
ioral phenomena can also be dissociated during juvenile development. It is pos-
sible, however, that these processes, although separable during ontogeny, are not
dissociable in the adult. Thus an important question was whether the same
forms of behavioral plasticity could be identified and separated in adult animals.
Marcus et al. (1987; 1988) addressed this possibility by examining the effects of
a wide range of tail shock intensities, at several times after tail shock, on both
habituated and nonhabituated siphon withdrawal reflex responses. They found
that dishabituation and sensitization were clearly dissociable in adult animals
in two independent ways: (1) on the basis of their differential time of onset, and
(2) on the basis of their differential sensitivity to stimulus intensity.

Dishabituation and Sensitization Have Different Times of Onset

The dissociation of dishabituation and sensitization on the basis of their differ-
ential time of onset is shown in Figure 2.13A, which illustrates results obtained
in a test session 90 sec after tail shock. The data are expressed as difference
scores comparing postshock to preshock responses. Dishabituation was signifi-
cantly expressed at a variety of stimulus intensities, whereas in this early test,
sensitization was not exhibited at any stimulus intensity (in fact, examining
nondecremented responses revealed that tail stimulation produces inhibition of
reflex amplitude). Although sensitization was not expressed in the 90-sec tests,
when nondecremented responses were examined in later tests, for example, 20
min and 30 min after tail shock, significant reflex facilitation was observed (data
not shown). Therefore, dishabituation has an early onset, within 90 sec after tail
shock, whereas sensitization has a very delayed onset, 20 to 30 min after tail
shock.
An interesting aspect of the data seen in Figure 2.13A is that the strongest
stimulus (4X, see below) produces no dishabituation, and in parallel, the same
stimulus intensity produces significant inhibition. This suggests that the recruit-
ment of an inhibitory process (by strong stimuli) competes with the expression
of dishabituation. As mentioned earlier, a very similar result was recently
obtained in juvenile Aplysia by Rankin and Carew (1987b; 1989), who found
that increasing tail shock produced progressively greater reflex inhibition and
also in parallel, progressively less dishabituation.
FIGURE 2.13. Behavioral dissociation of dishabituation and sensitization in adult Aplysia. (A) Significant dishabituation (asterisks) at a number of stimulus
intensities is produced soon after the tail stimulus in the 90-sec test. In contrast, in this early test, no sensitization is produced at any intensity; in all cases,
tail stimulation produced suppression of nondecremented responses, with the strongest stimulus (4X) producing significant inhibition. Data are expressed as
median difference scores (post-pre). (B) Reflex magnitude as a function of tail shock intensity for dishabituation and sensitization is shown. The dashed
line through 100% indicates baseline. Maximal dishabituation is produced by weak stimuli (which produces no sensitization), and maximal sensitization is
produced by strong stimuli (which produces no dishabituation).
46 FORMS OF MEMORY

Dishabituation and Sensitization Are Differentially Sensitive

to Stimulus Intensity
The dissociation of dishabituation and sensitization on the basis of stimulus
intensity is seen in Figure 2.13B. A range of stimulus intensities to the tail was
employed, from a mild tactile stimulus (touch) to increasing amplitudes of a
single electric shock (weak, int, and strong), and, finally, to multiple tail shocks
(4X). Maximal dishabituation was produced by weak tail stimuli, whereas max-
imal sensitization was produced by strong stimuli. Moreover, the stimulus
intensity that was most effective in producing dishabituation (weak) produced
no sensitization, and the stimulus intensity that was most effective in producing
sensitization (strong) produced no significant dishabituation. These behavioral
observations show that the processes of dishabituation, sensitization, and inhi-
biton, first behaviorally dissociated developmentally, could also be dissociated
in adult animals.

CONCLUSIONS

The experiments that we have described in this chapter allow us to draw five
principal conclusions. We will briefly discuss each of them in turn.

/. Habituation, Dishabituation, and Sensitization and Their Cellular Analogues

Have Different Developmental Timetables in Aplysia
Figure 2.14 compares the developmental timetables for behavioral habituation,
dishabituation, and sensitization and their cellular analogues. Habituation is
present as early as we can study the siphon, in stage 9, and continues to develop
in terms of the maximal ISI that produces habituation throughout the juvenile
life of the animal. Dishabituation emerges approximately one week later in stage
10. Finally, sensitization emerges several weeks after dishabituation, during
stage 12. A clear parallel between each of these forms of learning and its respec-
tive cellular analogue is apparent: habituation and its analogue, synaptic decre-
ment, are present in stage 9 (Rayport & Camardo, 1984), the earliest stage tested;
dishabituation and its analogue, facilitation of decremented EPSPs, emerge in
stage 10 (Rayport & Camardo, 1984); and sensitization and its analogue, facili-
tation of nondecremented EPSPs, emerge considerably later, during stage 12
(Nolen & Carew, 1988; Rankin & Carew, 1987b; 1988). The developmental par-
allel between the emergence of each form of learning and its respective cellular
analogue is striking; it suggests that gaining an understanding of the cellular
mechanisms underlying the expression of the synaptic plasticity we have
observed, prior to the emergence of additional processes that develop later in
time, may contribute importantly to an understanding of the mechanisms
underlying the expression of the learning processes themselves.
 The Development of Learning and Memory in Aplysia 47

FIGURE 2.14. Comparison of the developmental timetables of behavioral habituation, disha-

bituation, and sensitization and their cellular analogues. Habituation and its cellular analogue (ho-
mosynaptic decrement of EPSPs) emerge by stage 9; dishabituation and its analogue (facilitation
of decremented EPSPs) emerge in stage 10. Sensitization and its cellular analogue (facilitation of
nondecremented EPSPs) emerge much later (approximately 60 days) in mid- to late stage 12. (From
Nolen and Carew, 1988)

2. The Simultaneous Emergence of Sensitization in Different Response Systems

Suggests the Development of a General Process
The observation that sensitization or its cellular analogue emerges at the same
time in development in at least two different systems (siphon withdrawal and
escape locomotion) suggests the intriguing possibility that a set of integrating
mechanisms responsible for an animal-wide activation of this form of learning
may be expressed during stage 12 of juvenile development. Consistent with this
general idea is the recent observation that in this same stage of juvenile devel-
opment there is a dramatic and highly nonlinear proliferation of neurons
throughout the entire CNS (Cash & Carew, 1989), raising the possibility that
some aspect of the developmental trigger for neuronal proliferation might also
contribute to the general expression of sensitization. Such integrating mecha-
nisms for the coordinated expression of sensitization might include the induc-
tion of particular hormones, neurotransmitters, or facilitatory interneuronal sys-
tems, or the expression of one or more second messenger systems, which are
important for neuromodulation in the different response systems. Finally, if
such integrating facilitatory mechanisms exist, they are likely to be important
48 FORMS OF MEMORY

not only for the coordinated expression of sensitization in different response sys-
tems during ontogeny, but also for the integration of sensitization in multiple
response systems in the adult animal as well.

3. Analysis of Nondecremented Responses Prior to the Emergence

of Sensitization Reveals a Novel Inhibitory Process
Our behavioral and cellular studies examining the development of sensitization
have shown that in young animals, tail shock actually produces an inhibition of
response amplitude. In the behavioral analysis of the siphon withdrawal reflex,
Rankin and Carew (1987a, 1987b, 1988, 1989) found that before the emergence
of sensitization in late stage 12, tail shock actually produced a modest but sig-
nificant depression of nondecremented reflex responses. Similarly, Nolen and
Carew (1988) identified a cellular analogue of this depressive effect of tail shock
in nondecremented EPSPs in R2 prior to the development of sensitization. A
similar observation of inhibition in R2 was made by Rayport and Camardo
(1984) in stage 9 juvenile animals in their analysis of the emergence of the cel-
lular analogue of dishabituation. They found that when facilitation of decre-
mented EPSPs in R2 was absent in this early developmental stage, connective
stimulation produced a short-lived depression of peak EPSP amplitude. Thus in
both stage 9 and early stage 12, when the facilitatory processes underlying dis-
habituation and sensitization respectively have not yet emerged, an inhibitory
effect of connective stimulation is revealed.
To further examine this inhibitory process, Rankin and Carew (1987b, 1989)
recently studied the effects of different intensities of tail shock (very weak and
very strong) on the amplitude of siphon withdrawal in stage 11 and early stage
12 animals. Not surprisingly, they found that greater inhibition was produced
by the strong tail shock than by the weak shock. However, counterintuitively,
they found that significantly less dishabituation was produced by strong tail
shock than by weak tail shock, a finding also confirmed in adult animals (Marcus
et al., 1987, 1988). This observation is consistent with the hypothesis that the
inhibition produced by the strong tail shock actually competes with the expres-
sion of a facilitatory process (dishabituation) that is present in early develop-
mental stages. Taken collectively, these data suggest that two antagonistic pro-
cesses, one facilitatory (dishabituation) and the other inhibitory, emerge early in
development, and that at a later stage another facilitatory process (sensitization)
emerges.

4. Dishabituation, Sensitization, and Inhibition Can Be Separated

in Adult Aplysia
The behavioral separation of dishabituation and sensitization, as well as the
identification of an inhibitory process in developing juvenile Aplysia, raised the
interesting question of whether these processes can be identified behaviorally
and dissociated in adult animals as well. Recent experiments by Marcus et al.
 The Development of Learning and Memory in Aplysia 49

(1987, 1988) show that they can. By examining a variety of intensities of tail
shock, at several test intervals after shock, they found (1) that dishabituation has
an early onset (within 90 sec), whereas sensitization has a very delayed onset
(20-30 min) and (2) that dishabituation is produced by weak stimuli, whereas
sensitization is produced by strong stimuli. Moreover, tail shock intensities that
produced significant facilitation of habituated responses produced depression of
nonhabituated responses.
A similar tail shock induced inhibitory process has recently been described by
Mackey and colleagues (1987) and by Krontiris-Litowitz, Erikson, and Walters
(1987). Moreover, Mackey et al. (1987) identified an interneuron in Aplysia that
produces presynaptic inhibition of siphon sensory neurons, which is thought to
contribute to behavioral inhibition in the siphon withdrawal reflex. Thus there
is clear behavioral and cellular evidence that the inhibitory process we have
found in juvenile animals persists in adult Aplysia as well.

5. A Dual-Process View of Nonassociatlve Learning Requires Reevaluation

As described earlier in this chapter, until recently, a commonly held view was
that habituation, dishabituation, and sensitization could be accounted for by
two opposing processes: a decrementing process that produces habituation and
a single facilitatory process that produces both dishabituation and sensitization
(Carew et al., 1971; Carew & Kandel, 1974; Groves & Thompson, 1970). Our
experiments show that this view requires revision. Specifically, a key prediction
of the dual-process view is that dishabituation and sensitization should invari-
ably occur together. However, we showed that dishabituation and sensitization
(as well as an additional process, inhibition) can be dissociated behaviorally in
the siphon withdrawal reflex of Aplysia in three independent ways: (1) on the
basis of their differential time of onset; (2) on the basis of their differential sen-
sitivity to stimulus intensity; and (3) on the basis of their differential develop-
mental timetables. Moreover, previous theoretical and behavioral work by
Whitlow (1975) and Wagner (1976) suggested that dishabituation might be dif-
ferent from sensitization, and more recent cellular evidence by Hochner and
colleagues (1986a, 1986b) shows that two cellular processes in Aplysia sensory
neurons (transmitter mobilization and spike broadening) may contribute differ-
entially to behavioral dishabituation and sensitization. Based on these collective
results, we suggest that a dual-process view involving a single decrementing and
a single facilitatory process requires revision, and a multiprocess view, perhaps
involving inhibitory as well as facilitatory interactions, is better suited to explain
the interactive mechanisms underlying nonassociative learning.

Acknowledgments
We are very grateful to the Howard Hughes Medical Institute for their generous
support in supplying juvenile Aplysia. This research has been supported by an
NSF predoctoral fellowship (to E.A.M.), by NIH NRSA grant 5-F32-NS-07480
(to T.G.N.), and by NSF grant BNS 8614961, NIMH grant MH 41083, and
ONR grant N0014-87-K-0381 (to T.J.C.).
50 FORMS OF MEMORY

REFERENCES

Byrne, J. H. (1987). Cellular analysis of associative learning. Physiological Review, 67(2), 329-
439.
Carew, T. J. (1987). Cellular and molecular advances in the study of learning in Aplysia. In J.
P. Changeaux & M. Konishi (Eds.), The neural and molecular basis of learning (pp.
177-204). New York: Wiley.
Carew, T. J., Castellucci, V. F., & Kandel, E. R. (1971). An analysis of dishabituation and sen-
sitization of the gill-withdrawal reflex in Aplysia. International Journal ofNeurosci-
ence, 2, 79-98.
Carew, T. J., & Kandel, E. R. (1974). A synaptic analysis of the interrelationship between dif-
ferent behavioral modifications in Aplysia. In E. R. Kandel & M. V. L. Bennett
(Eds.), Synaptic transmission and neuronal interaction (pp. 187-215). New York:
Raven.
Carew, T. J., & Sahley, C. L. (1986). Invertebrate learning and memory: From behavior to
molecules. Annual Review of Neuroscience, 9, 435-487.
Carew, T. J., Walters, E. T., & Kandel, E. R. (1981). Associative learning in Aplysia: Cellular
correlates supporting a conditioned fear hypothesis. Science, 211, 501-504.
Cash, D., & Carew, T. J. (1989). A quantitative analysis of the development of the CNS in
juvenile Aplysia. Journal of Neurobiology, 20, 25-47.
Groves, P. M., & Thompson, R. F. (1970). Habituation: A dual process theory. Psychological
Review, 77, 419-450.
Hawkins, R. D., Clark, G. A., & Kandel, E. R. (1987). Cell biological studies of learning in
simple vertebrate and invertebrate systems. In F. Plum (Ed.), Handbook of physiol-
ogy: Sec. 1. The nervous system: Vol. 5. Higher functions of the brain (pp. 25-83).
Bethesda, Md.: American Physiological Society.
Hening, W. A., Walters, E. T., Carew, T. J., & Kandel, E. R. (1979). Motorneuronal control of
locomotion in Aplysia. Brain Research, 179, 231-253.
Hochner, B., Klein, M., Schacher, S., & Kandel, E. R. (1986a). Action-potential duration and
the modulation of transmitter release from the sensory neurons of Aplysia in presyn-
aptic facilitation and behavioral sensitization. Proceedings of the National Academy
of Sciences USA, 83, 8410-8414.
Hochner, B., Klein, M., Schacher, S., & Kandle, E. R. (1986b). Additional component in the
cellular mechanism of presynaptic facilitation contributes to behavioral dishabitua-
tion in Aplysia. Proceedings of the National Academy of Sciences USA, 83, 8794-
8798.
Jahan-Parwar, B., & Fredman, S. M. (1978a). Control of pedal and parapodial movements in
Aplysia, I. Proprioceptive and tactile reflexes. Journal of Neurophysiology, 41, 600-
608.
Jahan-Parwar, B., & Fredman, S. M. (1978b). Control of pedal and parapodial movements in
Aplysia. II. Cerebral ganglion neurons. Journal of Neurophysiology, 41, 609-620.
Kandel, E. R., & Tauc, L. (1965). Mechanism of heterosynaptic facilitation in the giant cell of
the abdominal ganglion of Aplysia depilans. Journal of Physiology, 181, 28-47.
Kriegstein, A. R. (1977a). Stages in the post-hatching development of Aplysia californica. Jour-
nal of Experimental Zoology, 199, 275-288.
Kriegstein, A. R. (1977b). Development of the nervous system of Aplysia californica. Proceed-
ings of the National Academy of Sciences USA, 74, 375-378.
Krontiris-Litowitz, J. K., Erikson, M. T., & Walters, E. T. (1987). Central suppression of defen-
sive reflexes in Aplysia by noxious stimulation and by factors released from the body
wall. Society for Neuroscience Abstracts, 13, 815.
Mackey, S. L., Glanzman, D. L., Small, S. A., Dyke, A. M., Kandel, E. R., & Hawkins, R. D.
(1987). Tail shock produces inhibition as well as sensitization of the siphon-with-
drawal reflex of Aplysia: Possible behavioral role for presynaptic inhibition mediated
 The Development of Learning and Memory in Aplysia 51

by the peptide Phe-Met-Arg-Phe-NH2. Proceedings of the National Academy of Sci-

ences USA, 84, 8730-8734.
Marcus, E. A., Nolen, T. G., Rankin, C. H., & Carew, T. J. (1987). Behavioral dissociation of
dishabituation, sensitization and inhibition in the siphon withdrawal reflex of adult
Aplysia. Society for Neuroscience Abstracts, 13, 816.
Marcus, E. A., Nolen, T. G., Rankin, C. H., & Carew, T. J. (1988). Behavioral dissociation of
dishabituation, sensitization, and inhibition of Aplysia. Science, 421, 210-213.
Mpitsos, G. J., & Lukowiak, K. (1985). Learning in gastropod molluscs. In A. O. D. Willows
(Ed.), The mollusca (Vol. 8, pp. 96-267). Orlando, Fla.: Academic Press.
Nolen, T. G., & Carew, T. J. (1988). The cellular analog of sensitization in Aplysia emerges at
the same time in development as behavioral sensitization. Journal of Neuroscience.
8, 212-222.
Nolen, T. G. Marcus, E. A., & Carew, T. J. (1987). Development of learning and memory in
Aplysia. III. Central neuronal correlates. Journal of Neuroscience, 7, 144-153.
Rankin, C. H., & Carew, T. J. (1987a). Development of learning and memory in Aplysia. II.
Habituation and dishabituation. Journal of Neuroscience, 7, 133-143.
Rankin, C. H., & Carew, T. J. (1987b). Analysis of the development emergence of sensitization
in Aplysia reveals an inhibitory effect of a facilitatory stimulus. Society for Neurosci-
ence Abstracts, 13, 816.
Rankin, C. H., & Carew, T. J. (1988). Dishabituation and sensitization emerge as separate pro-
cesses during development in Aplysia. Journal of Neuroscience. 8, 197-211.
Rankin, C. H., & Carew, T. J. (1989). Developmental analysis in Aplysia reveals inhibitory as
well as facilitatory effects of tail shock. Behavioral Neuroscience, 103, 334-344.
Rankin, C. H., Stopfer, M., Marcus, E. A., & Carew, T. J. (1987). Development of learning and
memory in Aplysia. I. Functional assembly of gill and siphon withdrawal. Journal of
Neuroscience, 7, 120-132.
Rayport, S. G. (1981). Development of the functional and plastic capabilities of neurons medi-
ating a defensive behavior in Aplysia. Unpublished doctoral dissertation, Columbia
University, New York.
Rayport, S. G., & Camardo, J. S. (1984). Differential emergence of cellular mechanisms medi-
ating habituation and sensitization in the developing Aplysia nervous system. Jour-
nal of Neuroscience, 4, 2528-2532.
Stopfer, M., & Carew, T. J. (1988). Development of sensitization in the escape locomotion
system in Aplysia. Journal of Neuroscience, 8, 223-230.
Wagner, A. R. (1976). Priming in STM: An information-processing mechanism for self-gener-
ated or retrieval-generated depression of performance. In T. J. Tighe & R. N. Leaton
(Eds.), Habituation: Perspectives from child development, animal behavior, and neu-
rophysiology (pp. 95-128). Hillsdale, N.J.: Erlbaum.
Walters, E. T., Carew, T. J., & Kandel, E. R. (1979). Classical conditioning in Aplysia califor-
nica. Proceedings of the National Academy of Sciences USA, 76, 6675-6679.
Walters, E. T., Carew, T. J., & Kandel, E. R. (1981). Associative learning in Aplysia: Evidence
for conditioned fear in an invertebrate. Science, 211, 504-506.
Whitlow, J. W. (1975). Short-term memory in habituation and dishabituation. Journal of
Experimental Psychology, Animal Behavioral Processes, 1, 189-206.
3
Synaptic Plasticity, Neural Architecture,
and Forms of Memory
R I C H A R D G. M. M O R R I S

This chapter deals with the relationship between synaptic plasticity and learning,
with particular reference to the work of Kandel and his colleagues on Aplysia.
After summarizing several general principles that may be derived from this
work, I argue that the claim that the neural mechanisms of conditioning involve
altering the efficacy of existing synapses does not require that all forms of con-
ditioning elaborate existing reflexes. I argue further that a neural mechanism
which detects the conjunction of two events will allow the nervous system to
represent this temporal relationship only if the conjunction mechanism is
embedded into appropriate circuitry. These arguments are relevant to efforts to
understand the relationship between psychological and neurobiological accounts
of conditioning, and they offer a new way of thinking about the significance of
parallel distributed memory systems.
The idea that synaptic plasticity plays a role in the neural mechanisms of
learning and memory has a long history. Tanzi (1893, cited by Ramon y Cajal,
1911, p. 886) was the first to propose that learning involved alterations in the
efficacy of existing connections from one neuron to another. Ramon y Cajal
(1911), however, argued that although such changes would be sufficient to
account for the perfection of existing skills, they would be insufficient to account
for the acquisition of new skills: "En admettant tout d'abord que les voies orga-
niques prexistantes sont renforcees par Fexercice; en supposant ensuite que de
nouvelles voies s'etablissent, grace a une ramification et une croissance de plus
en plus grande des arborisations dendritiques et cylindre-axiles" (p. 897).
This passage might allow two interpretations, but the context makes it clear
that Ramon y Cajal believed that learning involved the formation of new con-
nections between hitherto unconnected neurons. In 1949 Hebb elaborated the
older of these ideas in his theory of cell assemblies and proposed his now well-
known rule concerning the conditions under which alterations of synaptic effi-
cacy might occur: "When an axon of cell A is near enough to excite a cell B and
repeatedly or persistently takes part in firing it, some growth process or meta-

52
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 53

bolic change takes place in one or both cells such that A's efficiency, as one of
the cells firing B, is increased" (p. 62). Hebb's (1949) monograph is usually taken
as the starting point for the modern assumption, widely held by many neuro-
scientists, that synaptic plasticity has a role in learning as a mechanism for stor-
ing information. However, the first definitive evidence that alterations of syn-
aptic efficacy play a causal role in learning did not come until 20 years later,
with the discovery, by Pinsker, Kupfermann, Castellucci, and Kandel (1970) and
Castellucci and Kandel (1974) that behavioral habituation of the gill and siphon
withdrawal reflex ofAplysia was mediated by a reduction in transmitter release
at a defined synaptic locus. Subsequently, Bailey and Chen (1983) showed that
habituation was accompanied by alterations in the morphology of electrophy-
siologically identified synapses. Greenough and Bailey (1988) offer an excellent
review of the current state of the art concerning the relationship between learn-
ing and morphological alterations of synapses in both vertebrates and
invertebrates.
Given that it is now feasible to study isolated synapses in some nervous sys-
tems through the use of a variety of ingenious techniques, it is tempting to
believe that the major remaining problems in understanding the relationship
between plasticity and learning are essentially cell biological. Implicit in the
arguments presented in this chapter is a note of caution about whether these are
the only major problems left. To be sure, understanding the physical basis of
memory—the sequence of cellular and molecular events that give rise to the
various forms of synaptic and neural plasticity—is both a formidable goal and
a major one for many investigators (Horn, 1985; Kandel & Schwartz, 1982;
Lynch & Baudry, 1984; Rose, 1984; Routtenberg, 1987). However, I shall argue
that a full understanding of the relationship between synaptic plasticity and
learning requires an appreciation of different forms of learning and of the role
of different kinds of neural architecture within which a given type of synaptic
plasticity is embedded.

TWO ARGUMENTS OF THIS CHAPTER

Specifically, I shall argue that while neurobiological data favor the view that the
neural mechanisms of conditioning involve altering the efficacy of existing syn-
apses, accepting this need not commit us to view all conditioning as a process
through which existing reflexes are modulated. Here I must emphasize the
phrase "need not commit us," because certain well-described forms of condi-
tioning do involve no more than the alteration of existing reflexes. However,
not all forms of conditioning are of this kind. Importantly, it is possible to imag-
ine neurobiologically realistic schemes in which two hitherto unrelated events
can be associated through a neural mechanism that involves changing the effi-
cacy of existing neural connections but that can in no way be construed as a
modulation of existing reflexes. This possibility may seem paradoxical, but it is
54 FORMS OF MEMORY

feasible provided the stimuli to be associated are represented in the nervous sys-
tem in a distributed code.
The second argument of this chapter is that the type of learning mediated by
a particular form of synaptic plasticity is a function of the neural architecture
within which it is embedded. Such an argument may at first seem trivial, since
few if any neuroscientists would deny that knowledge of neuroanatomy is essen-
tial for understanding the functioning of the nervous system. However, my
emphasis is not on the indubitable and uncontroversial fact that knowledge of
neural architecture is useful (e.g., for the technical purpose of doing neurophys-
iological experiments properly), but on the idea that different kinds of neural
architecture allow different logical functions to be computed by particular types
of plasticity. That is, a particular type of plasticity will enable one logical func-
tion to be computed in one circuit and another in a different circuit. Clearly, a
corollary of this second argument is that there need be no simple relationship—
no isomorphism—between a particular type of synaptic plasticity and a partic-
ular form of learning. Again, it may often turn out that nature's secrets are sim-
ple and that a one-to-one relationship between learning and plasticity generally
holds, but the principle I want to establish is that a simple relationship is not
logically demanded. Examples will be offered both from recent work on Aplysia
and from one type of distributed memory system.

THE NECESSITY FOR THREE LEVELS OF INQUIRY

IN THE NEUROSCIENCES

In this chapter I follow Marr (1982) in recognizing the necessity for three differ-
ent levels of explanation in the neurosciences. Marr refers to these as the com-
putational, the algorithmic, and the neural implementational levels (Table 3.1).
The computational level (which is hereafter referred to as level 1) is used for
describing the task to be performed. In the case of learning, this might be to
ignore irrelevant stimuli (habituation) or to associate two events (conditioning).
The algorithmic level (level 2) describes the processes (but not the mechanisms)

TABLE 3.1 Levels of Inquiry in the Neurosciences

Computational Theory Representation and Algorithm Hardware Implementation
What is the goal of the How can this computational How can the representation
computation, why is theory be implemented? and algorithm be
it appropriate, and In particular, what is the realized physically?
what is the logic of representation for the
the strategy by input and output, and
which it can be what is the algorithm for
carried out? the transformation?
Source: From Marr (1982).
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 55

with which these tasks are accomplished. Thus we might have a procedural algo-
rithm such as "reduce response amplitude to all stimuli not followed by biolog-
ically significant events" (the conventional view of habituation) or, to take a
different theory of habituation, "learn what stimuli are to be expected in a given
context and do not process predictable stimuli in short-term memory" (Wagner,
1978). The algorithmic level is the level at which most psychological theories are
couched. Finally, the neural implementation (level 3) is the tangible description
of the neurons, the essential circuitry, the mechanisms of plasticity, etc., which
are hypothesized to underlie a given form of learning. It is the level concerning
itself with the physical machinery of the brain.
Each of these different levels of inquiry is important, each is characterized by
a particular type of discourse, and each has its own terminology. In this chapter,
for reasons that I hope to make clear, I shall maintain a particularly sharp dis-
tinction between processes (level 2) and mechanisms (level 3).

SOME GENERAL PRINCIPLES ABOUT LEARNING

THAT HAVE EMERGED FROM RESEARCH ON APLYSIA

The idea that many—perhaps all—forms of conditioning are the modulation of

existing reflexes has its origins in research on the neurophysiological basis of
simple forms of learning in Aplysia. The "basic circuitry" for habituation, dis-
habituation, sensitization, and a particular form of classical conditioning is dis-
cussed by Carew and colleagues (Chapter 2, this volume) and is shown in Figure
3.1. An appreciation of this basic circuit is sufficient for understanding certain
general principles that are widely held to have emerged from the remarkable
studies of learning in Apylsia since the early 1960s. However, this is by no means
the complete circuit responsible for these different forms of learning; Hawkins,
Castellucci, and Kandel (1981) described several other interconnections now
known to be important. Moreover, this is not the complete circuit for explaining
the learning of different aspects of responsiveness (duration, amplitude, etc.) in
even one of these forms of learning, namely sensitization (see Frost, Clark, &
Kandel, 1988). In a later section, I shall refer to these more complex aspects of
circuitry but the discussion here is restricted to the "basic circuit."
Reductions in transmitter release at the terminals from the sensory neurons
onto both interneurons and motorneurons are held to be the basis for habitua-
tion. Sensitization may be explained in terms of the activity in facilitator neu-
rons synapsing onto these same presynaptic terminals, causing a cascade of
intracellular events resulting in increased transmitter release. Classical condi-
tioning involves an activity-dependent modulation of this presynaptic facilita-
tion, a modulation induced by the temporal contiguity of neural activity in the
sensory neuron with neural activity in the facilitator neuron (Kandel &
Schwartz, 1982).
In effect, there is a relatively simple isomorphism between the type of plastic-
56 FORMS OF MEMORY

FIGURE 3.1. Basic circuit diagram used by Kandel to explain the essential mechanisms of
habituation, dishabituation, sensitization, arid classical conditioning in Aplysia. This diagram is
not the complete circuit of even part of the abdominal ganglion. (From Kandel and Schwartz,
1985, Fig. 62-3)

ity and the form of learning it mediates. There are three different types of syn-
aptic plasticity mediating three different forms of learning with, as Kandel and
Schwartz (1982) emphasized, one of these types (activity-dependent modulation
of presynaptic facilitation) being closely related to and perhaps evolved from
another (presynaptic facilitation). To the extent that ontogeny recapitulates phy-
logeny (Haekel, 1864, cited in Boakes, 1984), the intriguing proposal that the
neural mechanism of classical conditioning may have evolved from the mech-
anism of sensitization is supported by Carew and his colleagues' (Chapter 2, this
volume) studies of their time course of emergence during development—an
ordered time course which they refer to as functional assembly. As each succes-
sive mechanism of plasticity emerges, so does a corresponding form of learning.
Moreover, in developing their concept of a cellular alphabet of conditioning,
Hawkins and Kandel (1984) described how these cellular mechanisms of plas-
ticity can, when considered together and in the context of this basic circuitry,
explain various aspects of acquisition, extinction, and discrimination in condi-
tioning. Reference to the additional aspects of known circuitry referred to pre-
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 57

viously (Hawkins et al., 1981) is, however, essential to explain various other
conditioning phenomena such as second-order conditioning and blocking.
Recently Gluck and Thompson (1987) and Hawkins (1989) have extended Haw-
kins and Kandel's (1984) analysis with a quantitative model that postulates an
S-shaped learning function to explain blocking.
What general principles have emerged from this work? First, it implies that
information storage is intrinsic to sensorimotor pathways mediating a particular
learned behavior. A second principle is the idea that information storage is an
alteration in the efficacy of existing neural pathways. Kandel repeatedly stressed
this second point in numerous papers and the evidence strongly supports his
view, subject to one qualification. Bailey and Chen (1986) found that long-term
memory for sensitization can involve new connections being formed de novo—
an increase in the number of varicosities per sensory neuron. That this sprouting
occurs can perhaps be best understood in relation to the quantal release equation
(m = n X p). Greater overall synaptic throughput may, statistically, be more
easily achieved by having more presynaptic terminals available to release trans-
mitter (an increase in n) than by a modest increase in the probability of release
(an increase in p). If this presumption is correct, the sprouting of new terminals
is better seen as a special case of increasing the efficacy of existing connections
than as the formation of new connections between hitherto unconnected neu-
rons. Subject to this qualification, the evidence from Aplysia (as well as from
several other preparations, e.g., Hermissenda [Alkon & Farley, 1984] and nicti-
tating membrane conditioning in rabbits [Thompson, 1986]) supports the prin-
ciple that the underlying mechanism of learning involves altering the efficacy of
preexisting neural pathways.
A third general principle of the work on Aplysia is the possibility that the
detection of contiguity in classical conditioning is a cell-biological property of
neurons, not a system property. Thus activity-dependent modulation of presyn-
aptic facilitation is a mechanism for detecting the contiguity of CS and US
events. Specifically, prior depolarization of the sensory neuron presynaptic ter-
minals (i.e., CS activity) coupled with activation of the 5-HT receptor approxi-
mately 500 msec later is detected as an allosteric modification of a membrane-
bound adenylate cyclase (Abrams & Kandel, 1988). This is believed to happen
because of the possibility of dual activation of the cyclase by calcium-calmo-
dulin (representing the CS) and a G-protein (representing the US). The conse-
quence of the allosteric modification of the adenylate cyclase is that its subse-
quent activation by USs will result in a greater production of cAMP and, in turn,
enhanced phosphorylation of presynaptic K+ channels. Later CSs will then
cause even greater transmitter release than occurs after presynaptic facilitation
on its own. The detection of contiguity of CS and US is therefore not a property
of circuitry except in the trivial sense that the two neurons carrying CS and US
information must have a synapse between them. Rather it is a cell-biological
property arising out of the capacity for dual activation of adenylate cyclase. This
third principle is a brilliant discovery. It is perhaps not entirely without prob-
58 FORMS OF MEMORY

lems, notably in dealing with long-delay learning in classical conditioning (Gar-

cia & Koelling, 1966) and effects of interstimulus interval (not CS-US interval)
on conditioning (Yeo, 1974). However, long-delay learning is explicitly held to
be outside the scope of the principle and I shall not dwell on the point further
(see Abrams & Kandel, 1988).
Several other principles have emerged from the work on Aplysia but the three
just summarized provide a sufficient basis for introducing the detailed argu-
ments of this paper.

NOT ALL FORMS OF CONDITIONING INVOLVE THE MODULATION

OF PREEXISTING REFLEXES

As stated earlier, the first argument of this chapter is that we need not view all
forms of conditioning as a process through which preexisting reflexes are mod-
ulated even if we accept the premise that the neural mechanisms of conditioning
involve altering the efficacy of preexisting neural connections.
It has often been noted that the classical conditioning displayed by the abdom-
inal ganglion circuitry of Aplysia is what is generally known as alpha-condition-
ing. The CS of tactile stimulation to the siphon is not a neutral CS at the start
of conditioning. It already elicits the CR of siphon withdrawal. The effect of
pairing this CS with a tail shock US is to increase the strength of this preexisting
CR.
Opinion is divided about whether alpha-conditioning is qualitatively different
from beta-conditioning (in which the learned CR is different from any responses
the CS elicits prior to conditioning). Kandel and his colleagues are well aware of
this division of opinion—tired of it even—because they have repeatedly
asserted that any differences between these types of conditioning are unimpor-
tant (e.g., Hawkins & Kandel, 1984, p. 379nl). The position I shall adopt is that
at the level of neural mechanism the issue is trivial, whereas at the level of psy-
chological process it is of some importance. Furthermore, I believe that despite
many years of argument about alpha-conditioning, extending back to Hull
(1934), the position I am presenting is a novel one.
The first step is to understand why Kandel looks on the issue of alpha-con-
ditioning as unimportant. One reason is his belief, given the current state of our
knowledge, that any insights into any forms of learning are useful. Even if alpha-
conditioning should later prove not to be a particularly interesting form of learn-
ing, profitable research begins with a tractable problem (Medawar, 1966). The
study of alpha-conditioning is just such a problem and it is therefore right to
study it. I concur with this argument and my concern about alpha-conditioning
lies elsewhere.
The second and main reason Kandel regards the alpha-conditioning issue as
unimportant is because he sees no qualitative difference between alpha- and
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 59

beta-conditioning. To understand why he takes this view, it is helpful to contrast

a long-abandoned psychological theory of conditioning, the S-R version of
"stimulus substitution" (Hull, 1934), with Kandel's (1965, 1976) differential
facilitation hypothesis. In a typical classical conditioning task (level 1 descrip-
tion), the procedure is to pair an initially neutral CS with a US which elicits a
UR. This results in the development of a CR, which in many cases resembles
the UR. According to S-R stimulus substitution theory (level 2 algorithm), the
new CR can be evoked by the CS because a "connection" is formed (dashed line
in Fig. 3.2) from CS to UR. This process account must, however, worry anyone
attempting to understand the neural mechanisms of learning (a level 3 imple-
mentation), because the new connection from CS to UR seems to require a new
neural pathway. As Kandel (1976) points out, "Presumably this involves the
development of a new neural pathway [dotted line in Fig. 3.2] through which
the CS (which previously did not have access to the UR) now produces
responses similar to the UR" p. 639, Fig. 13-20). Although not stated explicitly,
it is clear that Kandel is rightly skeptical about the possibility of such new con-
nections being formed. He therefore rejects stimulus substitution as neurally
implausible.
Kandel's differential facilitation hypothesis is radically different. Following
Marr's (1982) terminology, I shall refer to this hypothesis as the "differential
facilitation algorithm" in order to maintain the sharp distinction between pro-
cess accounts (level 2) and accounts in terms of mechanism (level 3). Strictly
speaking, differential facilitation is a process account (a level 2 account), which
could be implemented in neural hardware in any of a variety of ways. According
to the hypothesis, a CS covertly elicits a wide variety of CRs and the effects of
pairings of the CS with the US is to facilitate one (or more) of these CRs at the
expense of the others (Fig. 3.3). The implication of this hypothesis is absolutely
clear in Kandel's (1976) description: "Prior to pairing with a US, a CS may be
capable of eliciting, if only in subthreshold form, all the responses to which it
can subsequently become conditioned" (p. 638). This extraordinary feature of
the differential facilitation hypothesis raises various problems that I shall discuss
shortly. Its great virtue, of course, is that it can more easily be translated into a
plausible account at the level of neural mechanism. Connections from CS to CR
become pathways, and differential facilitation becomes activity-dependent mod-
ulation of presynaptic facilitation. Interestingly, the differential facilitation
hypothesis was not developed after the fact—it predated the discoveries about

FIGURE 3.2. Pairing of a CS with a US, which elicits a UR,

is held, according to Hull's S-R version of stimulus substitution,
to result in the formation of a new connection between the CS
and UR.
60 FORMS OF MEMORY

FIGURE 3.3. Kandel's (1965) differential-facilitation hypothesis of conditioning. According to

the model, the CS gives rise to a family of responses many of which will initially be below threshold
for behavioral expression. Pairing of the CS and US results in a differential facilitation of certain
CRs depending on the particular US with which a given CS is paired. As the strength of these
covert CRs increases, one CR may rise above threshold and so be expressed in the form of overt
behavior. (From Carew et al., 1984, Fig. 8.4)

the mechanisms of conditioning in Aplysia by several years (Kandel, personal

communication, 1987).
Kandel's idea that all conditioning is the elaboration of existing reflexes was
first presented in an unpublished paper dating from 1965, which he has kindly
made available to me. This paper outlines the relation between alpha-condition-
ing and other forms of learning. Table 3.2, adapted from that paper, summarizes
the differences between four different forms of learning: alpha-conditioning,
beta-conditioning, pseudoconditioning, and sensitization. Hull's (1934) distinc-
tion between alpha- and beta-conditioning is shown on the left. For Hull, "true"
conditioning is pairing-specific (and is therefore in the top row) and involves the
emergence of a new response (lefthand column). Hull saw alpha-conditioning as
akin to sensitization because the CR is an amplification of a preexisting response
(righthand column), but he failed to emphasize that they differed with respect to
pairing specificity. Grether (1938) introduced the distinction between condition-
ing and pseudoconditioning. In both of these forms of learning, the conditioned
response is a new one (lefthand column). For Grether, the important difference
between conditioning and pseudoconditioning lay in only the former being pair-
ing-specific. In pseudoconditioning, US presentations in the absence of a CS sub-
sequently result in the CS evoking a CR that resembles the UR, even though
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 61

TABLE 3.2 Differences Between Forms of Learning

Hull (1934)
Preexisting
Learning New Response Response
Specific Beta-conditioning
Nonspecific Alpha-conditioning

Grether(1938)
Preexisting
Learning New Response Response
Specific Beta-conditioning
Nonspecific Pseudo-
conditioning

Kandel(1965)
Preexisting
Learning New Response Response
Specific Beta-conditioning Alpha-conditioning
Nonspecific Pseudo- Sensitization
conditioning

there has been no pairing of the CS and US. However, Grether failed to empha-
size any distinction between sensitization and pseudoconditioning and in many
textbooks since that date, the two phenomena have frequently been confused.
Kandel's (1965) contribution was to see these two distinctions, for the first time,
as components of a 2X2 table in which specificity of pairing was one dimension
and newness of the conditioned response was the other. In such a 2 X 2 table, a
new box emerges (top right), specific alpha-conditioning, the type of condition-
ing displayed by Aplysia. It is true conditioning because it is pairing-specific, but
the CR that emerges is the same as that originally elicited by the CS (i.e., a pre-
existing response). For Kandel, the two rows of this 2X2 table are qualitatively
different: they reflect the cell-biological mechanism of activity-dependent mod-
ulation. But the two columns differ only quantitatively. New responses are
merely CRs that were below threshold before conditioning.
My purpose in comparing the S-R version of stimulus substitution with dif-
ferential facilitation is threefold. First, it makes clear that Kandel's theorizing
occupies two very different domains, level 2 and level 3: (1) he has an essentially
psychological hypothesis about the kind of algorithm being computed in a con-
ditioning task and (2) he has a wide range of ideas (and evidence!) about the
neural mechanisms involved (the general principles summarized above). Sec-
ond, the comparison suggests a plausible reason why Kandel developed the dif-
ferential facilitation hypothesis. I surmise that he wanted to avoid what he cor-
62 FORMS OF MEMORY

implausibility. In effect, he was accepting that aspect of the Ramon y Cajal her-
itage which supposed that synapses were important for information storage
while rejecting the "strong interpretation" of Ramon y Cajal's position—the
necessity for new connections between hitherto unconnected neurons. Kandel
frequently writes that he sees learning as "fundamentally different" from other
plastic neural processes in which new connections are formed (such as reactive
synaptogenesis; see Lynch & Baudry, 1984, who take the opposite view). Indeed,
in Chapter 62 of Principles of Neural Science, work on learning in Aplysia is
juxtaposed with work on the unmasking of silent synapses in cat spinal cord
(Wall, 1975) and primate somatosensory cortex (Merzenich et al., 1983). Thus
the common feature of activity-dependent modulation and the sectioning of
afferent nerves is that they both have functional consequences which reflect the
intricacy of prewiring in the nervous system. Kandel's perspective echoes that
of Sherrington (1951): the nervous system is an "enchanted loom," and we
should respect its intricate prewiring.
The third point served by the comparison of stimulus substitution and differ-
ential facilitation is that it makes clear why Kandel regards the alpha-condition-
ing issue as trivial. According to the differential facilitation hypothesis, there
never are any new responses. Everything is prewired.1 In his view, the behavioral
scientist is misled by the emergence of apparently new responses during certain
conditioning procedures (see Carew, Abrams, Hawkins, & Kandel, 1984, pp.
175-177) because the underlying process of conditioning is a gradual increase in
the efficacy of CRs to which the CS has always been wired up. Eventually these
CRs cross a behavioral threshold and emerge as apparently novel responses.
It is on this third point that I have to part company with Kandel, because this
feature of the differential facilitation hypothesis—reflex prewiring—is psycho-
logically implausible. It cannot be correct to say that all the responses a CS can
ever elicit already exist as reflexes to that CS, albeit in subthreshold form. My
reasons for rejecting such a possibility are (1) its intuitive implausibility and (2)
the fact that conditioning is a process through which animals learn about the
predictability of the environment.
Consider the issue of plausibility first. Is it plausible to suppose that all the
responses that can ever be conditioned to a CS are already reflexly wired up to
it? That Pavlov's dog salivated to the metronome because the metronome stim-
ulus was covertly wired up to salivation? Kandel's work on the underlying neu-
ral mechanisms of learning seems to require this conclusion. However, several
phenomena in the behavioral literature render "reflex prewiring" particularly
implausible. One is the emergence of US-relevant responses during autoshaping
(Jenkins & Moore, 1973). Approach and contact may well be prewired investi-
gatory responses to a focal light stimulus, such as a key light in an operant cham-
ber, but it is hard to accept that the "drinking" keypecks which birds elicit when
autoshaped for water are also reflexively prewired. Another is the phenomenon
of counterconditioning in which, for example, weak electric shock is used as a
CS for food. Initially the shock elicits unconditional responses of flinching, freez-
 Synoptic Plasticity, Neural Architecture, and Forms of Memory 63

ing, and defecation. However, as conditioning proceeds, these responses subside

(although they need not be totally obscured; see Dickinson & Dearing, 1979) and
come to be replaced by behavior reflecting the anticipation of food: general
activity and, in the case of the study by Erofeeva cited by Pavlov (1927), sali-
vation. It is hard to believe that salivation is reflexively prewired to electric
shock.
The second reason for being suspicious of the generality of differential facili-
tation is that nothing in the algorithm captures the sense that conditioning is a
process through which an animal comes to learn about the predictability of the
environment. Any sense that salivation happens because the CS predicts the
meat (Zener, 1937) is, strictly speaking, absent from the hypothesis. Here, how-
ever, I confess to being confused by Kandel's writings. Several passages in Kan-
del and Schwartz (1985) make it abundantly clear that Kandel fully appreciates
the revolution in our thinking about the processes of conditioning that began in
the 1960s (Dickinson, 1980; Mackintosh, 1983;Rescorla, 1988; Rescorla& Wag-
ner, 1987). Thus Kandel writes:

Classical conditioning is fascinating because, as we saw earlier, it represents the learn-

ing of a fundamental causal relationship. When an animal has been conditioned, it
has learned that the conditioned stimulus predicts the unconditioned stimulus. Thus,
through classical conditioning, an organism is able to discriminate cause and effect
relationships in the environment. (Kandel & Schwartz, 1985, pp. 823-824)

The problem is that the differential facilitation algorithm cannot realize the pro-
cess of predicting the environment. It is an algorithm for differentially increasing
the strength of existing CRs when the CS and a particular US occur in temporal
contiguity. In no sense does the learned CR predict the US. A particularly striking
way of seeing this problem is with reference to the phenomenon of sensory pre-
conditioning in which two initially neutral CSs are paired and one of them is
later paired with a US (Rescorla, 1980). In a standard conditioning procedure,
a CS is arranged to predict the US. In sensory preconditioning, this procedure
is broken down into two stages: CS1 is arranged to predict CS2, and CS2 subse-
quently is arranged to predict the US. In a final test phase, CS1 is presented, and
the question at issue is whether the animal behaves in a manner reflecting the
anticipation of the US. The problem for differential facilitation here is that the
pairing of CS1 and CS2 (stage 1) occurs before the CS2 is paired with either an
aversive or an appetitive stimulus (stage 2). As there is no way of the animal
knowing, in the first stage of the experiment, what kind of event CS, will be
paired with in stage 2, there is no obvious way, short of serendipity, of selectively
facilitating defensive or consummatory reflexes in stage 1. Although sensory pre-
conditioning might be put aside as an example of a particularly complex form
of conditioning and therefore outside the scope of differential facilitation, it is,
in my view, the sort of phenomenon that brings us closer to an understanding
of the kind of process conditioning really is. In a recent book largely devoted to
work on invertebrate simple systems, Rescorla (1984) drew particular attention
64 FORMS OF MEMORY

to sensory preconditioning: "My view is that neurobiologists studying learning

could profit considerably by including this paradigm in their armoury of pro-
cedures" (p. 43).
Thus far in this section I have endeavored to suggest (1) an explanation of
why Kandel was led to propose his differential facilitation theory of condition-
ing, and (2) why it is psychologically implausible. However, the argument is
incomplete in two very different ways. First, Hawkins and Kandel (1984) are
well aware that the "basic circuit" of the abdominal ganglion is incapable of
explaining several higher forms of conditioning and they point out that it is
essential to refer to additional aspects of known circuitry to explain these. I turn
to these arguments in a later section where I consider the relationship between
synaptic plasticity and neural architecture. Second, because I have no quarrel
with the idea that learning involves no more than changing the efficacy of pre-
existing neuronal connections, the final part of this section has to be the more
constructive task of outlining one way in which this principle can be made com-
patible with a more plausible process account of conditioning. One such way is
the distributed associative matrix (Willshaw, Buneman, & Longuet-Higgins,
1969), the earliest and most frequently "rediscovered" system for storing infor-
mation in a distributed manner. Although some of the newer networks of con-
temporary work on parallel distributed processing involve more sophisticated
algorithms (Rumelhart & McClelland, 1986), the principle I wish to establish in
this section can be made with reference to the associative matrix alone.
Consider a matrix of wires intersecting at nodes (analagous to synapses) that
can be either OFF or ON (Fig. 3.4). If a node is ON, activity in a horizontal wire
can induce activity in a vertical wire. If a node is OFF, activity in a horizontal
wire has no effect on activity in a vertical wire. In the initial state of the matrix,
all nodes are OFF (Fig. 3.4a). Thus no pattern of activity on the horizontal wire
will cause any activity whatsoever in the vertical wires. This matrix can be made
to store information according to the "Hebbian" rule that a node is turned per-
manently from OFF to ON when there is simultaneous (i.e., conjunctive) activity
in both a horizontal and a vertical wire. Information is stored as a change of
state of the corresponding node when such conjunctions are detected. Stimuli
are presented to this matrix as spatial patterns of activity on either the horizon-
tal or the vertical wires. Thus 000111 might be one stimulus of a pair and
110100 the other. Notice that these stimuli differ in their spatial pattern, but they
overlap in one of their elements (in this instance, the fourth digit). The associ-
ation between these two stimuli is stored as a distributed pattern of changed
nodes (Fig. 3.4b). After one pair of stimuli has been stored, the matrix assumes
a particular pattern, and as further pairs of stimuli are presented in succession,
the matrix alters progressively (Fig. 3.4c and 3.4d). Numerous nodes are
involved in each association, and there is substantial overlap between different
pairs of associations. Thus a given node may be involved in the storage of sev-
eral associations. However, if an attempt is made to store too much information,
the matrix will become saturated and cause errors in recall.
Storage of associations is of use only if the information can be retrieved. This
FIGURE 3.4. Associative matrix memory system in which storage depends on conjunctive activation of orthogonal wires intersecting at modifiable nodes
(after Willshaw et al., 1969). Information to be associated is presented simultaneously to the matrix during storage, and those nodes at which there is
simultaneous activation become irreversibly switched from OFF to ON (heavy dots). Several pairs of associations can be stored within the matrix without
saturation. Retrieval results in activation of one member of an associative pair upon presentation of the other. A thresholding operation analogous to integer
division is required for correct recall. Such a system might be realized in a real nervous system in which NMDA receptors detect the conjunction of afferent
activity and postsynaptic depolarization, the resulting intracellular calcium signal triggers an alteration of synaptic efficacy, and feedforward inhibition serves
the function of thresholding during recall (see McNaughton and Morris, 1987).
66 FORMS OF MEMORY

is achieved by presenting one member of any associative pair and attempting to

activate ("remember") the other. In Figure 3.4e, the retrieval cue 000111 is pre-
sented to the horizontal wires. The retrieval rule is to allow activity in a hori-
zontal wire to induce activity in a vertical wire only at switches that are ON.
Adding things up, 000111 will therefore induce the output pattern 331321. At
first sight, this pattern in no way resembles the stimulus with which 000111 was
initially presented. However, integer division by the number of active elements
in the recall cue (i.e., 3) gives the number 110100—the stimulus with which
000111 was originally associated.
Numerous fascinating properties of matrix memories have been explored for-
mally over many years (Kohonen, 1977; Palm, 1982; Rumelhart & McClelland,
1986; Willshaw et al., 1969). I shall not dwell on these, except to say that their
storage capacity is a function of (1) how stimuli are represented across the input
array, and (2) the "local synaptic rules" governing changes at the nodes com-
prising the matrix (Palm, 1982). The Hebbian rule is one effective synaptic rule,
but there are others (Levy, Andersen, & Lehmkule, 1985). In the present context,
three principles of matrix memory systems must be stressed:
1. The matrix concept is not an obscure invention of computer science of no
biological relevance. It is biologically feasible and McNaughton and Morris
(1987) and Morris (1988) have tentatively explored the possibility that the
mammalian hippocampus (or at least parts of it) is a type of associative
matrix memory.
2. If implemented in neural hardware, the mechanism of storage would involve
no more than changing the efficacy of preexisting synapses. Thus the matrix
concept shares with the neural mechanisms of learning displayed by Aplysia
the fact that no new neural connections have to be formed.
3. Any stimulus that can be represented as a pattern of activity over the hori-
zontal wires can be paired with any stimulus represented as activity on the
vertical wires. Of course, a potentiality for association must exist; it must be
possible to represent stimuli in the form of a pattern of activity across the
wires of the matrix in order for the stimuli to enter into associations (this
potentiality might be thought of as a "biological constraint" on learning in
exactly the same sense as Carew et al. (1984) proposed; see note 1). However,
subject to this qualification, any stimulus can be arbitrarily paired with any
other and, in recall, the "memory" of one stimulus will be activated by pre-
sentation of the other. Such a storage principle captures the essential notion
of "representation," which, as described above, is so central to modern
accounts of conditioning.

FORMS OF LEARNING ARE A FUNCTION OF NEURAL

ARCHITECTURE

The second argument of this chapter is that the form of learning mediated by a
particular type of synaptic plasticity is a function of the neural architecture
within which it is embedded.
 Synoptic Plasticity, Neural Architecture, and Forms of Memory 67

A mechanism for detecting the conjunction of two events is clearly an essen-

tial feature of any implementational theory of classical conditioning (level 3).
However, a synaptic mechanism that detects contiguity will allow the nervous
system to represent this temporal relationship only if the conjunction mecha-
nism is embedded into appropriate circuitry.
Consider the "basic circuit" for Aplysia seen in Figure 3.1. The facilitator neu-
ron synapses onto both the sensory neuron-interneuron pathway and the sen-
sory neuron-motor neuron pathway. Activity in the facilitator neuron can there-
fore modulate the effectiveness of these synapses (in both sensitization and
classical conditioning). However, the CS cannot evoke any neural representation
of the US because the circuitry apparently does not permit it. The basic circuit
will detect the contiguity of CS and US events during conditioning through
activity-dependent modulation of presynaptic facilitation, but it cannot repre-
sent this relationship, where "representation" is taken to mean the CS being
capable of evoking activity in the US pathway.
Hawkins and Kandel (1984) are well aware of the problem. In the second half
of their paper, they draw attention to various additional aspects of known cir-
cuitry in Aplysia and, in particular, to neural pathways from the sensory neurons
to the facilitator neurons (Fig. 3.5). These synapses are modifiable according to
a quasi-Hebbian principle by virtue of feedback modulation from the facilitator
neuron collaterals onto the presynaptic terminals of the sensory neuron input
(Fig. 3.5B). Accordingly, in the real circuit of the abdominal ganglion, if sensory
neuron activity impinges on the facilitator just before US activity, the synaptic
connection between the sensory neuron and the facilitator neuron will be subject
to activity-dependent modulation. As a result, subsequent activation of the sen-
sory neuron can drive the motor neuron via the US pathway. Hawkins and Kan-
del (1984) make reference to these additional aspects of circuitry to explain
higher order conditioning. However, this additional pathway—the pathway
from sensory neuron to facilitator neuron—has wider ramifications. Specifically,
the real circuit differs from the "basic circuit" in that it both detects and can, in
principle, represent contiguity. The reason is as follows. Activity in the US path-
way, as a consequence of facilitator neuron activation, can be regarded as an
elementary kind of representation of the US. The CS now has the capacity to
evoke this representation. The real circuit may therefore, as Hawkins and Kan-
del (1984) point out, be regarded as one way to implement a completely different
theory of stimulus substitution, namely Pavlov's (1927) and Konorski's (1948)
S-S version of stimulus substitution. However, in describing a circuit in Aplysia
compatible with a different version of stimulus substitution theory, a version
that does not require the formation of new neural pathways between hitherto
unconnected neural elements, Hawkins and Kandel have broken the ostensibly
intimate connection between sensitization and classical conditioning.
Consider what would happen in the real circuit if US presentations occurred
alone. Strong stimulation of the tail region would activate tail sensory neurons,
drive the facilitator neurons, and, by the feedback pathway on the presynaptic
terminals of the sensory neuron-facilitator neuron pathway, increase the
strength of these terminals by the process of presynaptic facilitation. That is, US
68 FORMS OF MEMORY

FIGURE 3.5. A particularly important property of distributed memories is their capacity to recall
stimuli accurately from fragmentary input. This property is a function of the neural architecture
and independent of the particular cell-biological mechanism through which nodes are modified.

rectly judged to be the fatal weakness of Hull's (1934) account—its neural

presentations alone could, at least in principle, increase the tendency of the CS
to evoke responses similar to the UR without any necessity for pairing of the CS
and US. This, as we saw earlier, is the process of "pseudoconditioning." That is,
presynaptic facilitation may be the underlying mechanism of pseudocondition-
ing as well as the underlying mechanism of the process with which it is usually
discussed, sensitization. Whether the behavioral outcome is sensitization or
pseudoconditioning will depend on the circuitry within which presynaptic facil-
itation is embedded. Further, given that there is no a priori reason to regard the
basic circuit as any more fundamental than the real circuit, it would surely be
as appropriate for Hawkins and Kandel (1984) to describe classical conditioning
as an elaboration of pseudoconditioning as to describe it as an elaboration of
sensitization. The nature of the response that could be conditioned to a CS by
activity-dependent modulation of presynaptic facilitation would then be deter-
mined by the efferent connections of the facilitator neuron, not the efferent con-
nections of the CS pathway.
The breaking of the connection between classical conditioning and sensitiza-
tion in favor of an equally plausible one between classical conditioning and
pseudoconditioning might be thought to be of little consequence. Indeed, Kan-
del argued in his unpublished paper of 1965 that the difference between the col-
umns of Table 3.2 may only be quantitative. In my opinion, such a position is
untenable, first, because the link between classical conditioning and sensitiza-
tion is axiomatic within differential facilitation—the whole point of the algo-
rithm being that conditioning is to be viewed as a process through which existing
reflexes to a CS are modulated by CS-US conjunctions. And it is untenable
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 69

because describing a critical change in the neural architecture within which

activity-dependent modulation is embedded as "quantitative" gives unreason-
able priority to the importance of cell-biological principles over those at the sys-
tem level.
The idea that the relationship between plasticity and learning depends criti-
cally on the neural architecture within which it is embedded can also be seen in
relation to matrix memories. I shall consider this principle with reference to one
positive feature and one weakness of matrix memories.
The positive feature is the capacity of matrix memories to achieve accurate
recall despite being given only an incomplete or inaccurate input. A corollary of
this feature is that matrix memories may continue to function quite well when
partially damaged, a feature which Kohonen (Chapter 16, this volume) refers to
as "graceful degradation" and which has been explored by Wood (1978). In the
description given earlier, I outlined how the association of 000111 with 110100
in the learning phase would allow the stimulus 000111 to reactivate a memory
representation of 110100 during retrieval even if several other associations are
overlaid in the same matrix. Consider now what happens if, after learning, the
retrieval cue 000110 was presented instead of 000111. The two stimuli are sim-
ilar, differing only in the last digit. Adding things up, as in Figure 3.6, 000110
causes a pattern of activity in the vertical wires amounting to 221211. Remem-
ber, however, that this number must be divided by the total number of active
elements in the retrieval cue to achieve accurate recall. Previously this number
was 3 (the total number of 1's in 000111). Now the total number is 2 because
the incomplete cue is being presented (000110). Integer division of 220211 by 2
yields 110100, the stimulus with which 000111 was originally associated. Thus
an important "emergent" property of matrix circuitry is its capacity to continue
to function effectively in recall despite incomplete or fragmentary input. This
property of matrix memories has received extensive analysis elsewhere (e.g.,
Willshaw, 1981).
The weakness of matrix memories, which is also a property of the architecture
rather than the particular cell-biological mechanism with which the constituent
Hebb synapses are implemented in the nervous system, is that they fail to offer
a satisfactory account of predictability or contingency. This may be seen most
easily in relation to the phenomenon of blocking first described by Kamin
(1968). The basic paradigm involves two groups of animals and a three-phase
experiment (Fig. 3.7). In phase 1, one group learns that a stimulus (A) is paired
with a US; the other group is given no training. In phase 2, a second stimulus
(B) is added and the compound A + B is paired with the same US that was used
in phase 1. Identical training procedures are used for both groups in phase 2. In
a final test phase, B is presented alone and the question at issue is whether B
evokes a memory of the US. Kamin's (1968) discovery was that learning of the
B-US association was "blocked" in group 1. Blocking is not an inevitable con-
sequence of compound training because satisfactory learning of the B-US asso-
ciation occurred in group 2. Kamin's original account of this phenomenon
emphasized the sense in which the US was already predicted in group 1 at the
A

C
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 1{

FIGURE 3.7. Basic stimulus arrangements for an experiment on blocking (after Kamin, 1968).

start of phase 2 and that, in some way, this prevented any new learning from
happening. Clearly the US is not predicted at the start of phase 2 for group 2
and thus learning proceeds normally. Rescorla and Wagner (1972) formalized
this intuition in a simple equation according to which conditioning occurs
whenever there is a discrepancy between the total associative strength of all
stimuli present on a trial and the total associative strength that a given US can
support. Other acocunts of blocking have also been offered (Mackintosh, 1975;
Pearce&Hall, 1980).
The distributed matrix cannot compute the essential difference algorithm of
Rescorla and Wagner (1972) in order to account for blocking. This is because
the Hebbian rule operates locally at nodes throughout the matrix in the absence
of global information about the extent to which one stimulus adequately pre-
dicts the other with which it is paired. Figure 3.8 shows how pairings of a stim-
ulus A (e.g., 10011) and the US (e.g., 11010) might produce a particular matrix
of altered nodes (in a 5 X 5 matrix), how the compound of A + B would result
in the matrix changing at those nodes where the B stimulus interacts with the
US and which had not already been modified, and how subsequent presenta-
tions of B would therefore result in a memory representation of the US being
evoked. Thus, as described, the matrix concept fails to account for blocking. The
central problem is that there appears to be no way to prevent the additional B-
US nodes from being altered because of the prior association of A with the US.

FIGURE 3.6. Synaptic connections of L29, one of the facilitator neurons. (A) A spike elicited by
intracellular stimulation in an LE neuron produces an EPSP in L29. (From Hawkins et al., 1981,
Fig. 7A) (B) More complete circuit for gill- and siphon-withdrawal reflex and its modification by
tail stimulation and differential conditioning. Note feedback innervation from the facilitator inter-
neuron onto terminals from the LE neurons onto the facilitator neuron itself. (From Hawkins and
Kandel, 1984, Fig. 5A) (C) More complete summary diagram of the synaptic connections of
interneurons involved in the gill-withdrawal reflex. (From Hawkins et al., 1981, Fig. 9, where
details of the interconnections are given) The recent paper by Frost et al. (1988) confirms that the
siphon-withdrawal reflex is also mediated by highly complex circuitry and discusses the idea that
different components of responsiveness are represented at different points of this circuitry.
72 FORMS OF MEMORY

There may be a number of ways to get around the problem and these fall natu-
rally into two classes: (1) solutions that add extra circuitry, which "gates" input
to the matrix by solving what is known as the "exclusive-or" problem (see
Rumelhart & McClelland, 1986, Chapter 2); and (2) solutions that alter the local
synaptic rule operating at individual nodes (e.g., Bear, Cooper, & Ebner, 1987).
I shall not pursue these detailed matters here.
Some readers may find it confusing to have a discussion of conditioning in an
animal whose circuitry has been worked out through painstaking anatomical
and electrophysiological experimentation together with discussion of an ideal-
ized system whose implementation within the nervous system has never been
established. However, my aim is to establish a principle that applies equally to
both systems: the form of learning mediated by a particular type of synaptic
plasticity is a function of the neural architecture within which it is embedded.
In this final section, I have attempted to show how the basic circuit ofAplysia,
widely used by Kandel and his colleagues for the pedagogical purpose of explain-
ing their ideas about synaptic plasticity, is a circuit that cannot represent the
relationship between the CS and US in classical conditioning. The real circuitry
ofAplysia can represent contiguity (and predictability), but it does so awkwardly
and at the cost of abandoning the relation between classical conditioning and
sensitization that has been central to Kandel's approach for many years. In com-
parison, the matrix concept provides a much more straightforward way to
explain how any stimulus can evoke a memory representation of any other stim-
ulus, provided both stimuli can be presented to the matrix in a distributed code.
However, the matrix concept fails to provide a satisfactory account of predict-

FIGURE 3.8. The associative matrix operating according to a Hebbian conjunction rule is unable
to explain blocking because the alteration of nodes from an OFF to an ON state is determined locally
and unaffected by the global computation of the extent to which the US is already predicted by
stimulus A.
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 73

ability because the local synaptic rule determining information storage is not
error-correcting.

CONCLUSIONS

This chapter has reconsidered an old issue, the relation of alpha-conditioning to

classical conditioning, with the aim of revealing insights into (1) the relationship
between psychological and neurobiological accounts of conditioning and (2) the
problem of representation by the nervous system. I argued that the algorithmic
computations performed during learning are realized by particular mechanisms
of synaptic plasticity embedded into particular types of neural architecture such
that the nervous system's capacity to represent the association between events
is an emergent property of circuitry. Thus there need be no simple mapping
between cell biology and learning. A cellular alphabet of plasticity exists; a cel-
lular alphabet of learning does not (Table 3.3).
Finally, as parts of this chapter might be wrongly construed as critical of the
work on Aplysia, it may be worth clarifying the essential point of disagreement.
It is not with respect to the experimental work itself, which has justly earned the
respect of neuroscientists all over the world. Equally, I have no criticisms of
Kandel's account of the mechanisms of those forms of conditioning that are dis-
played by the abdominal ganglion of Aplysia. Rather, in trying to think about
how to apply the principles often drawn from this work, particularly in relation
to the vertebrate nervous system, I find that it may be possible to accept the idea
that learning involves the modulation of existing neural pathways without
accepting the idea of reflex prewiring. The matrix idea is but one of a range of
distributed information processing systems with the capacity to learn and, like
others, it is a system for which it makes no sense to speak of a particular CR
covertly prewired to a particular CS.

TABLE 3.3 Summary of Arguments

Changes in synaptic efficacy during learning can store information.
Presynaptic facilitation
Long-term potentiation
Cell-biological mechanisms may be responsible for detecting associations between temporally
coincident events.
Activity-dependent modulation
NMDA receptors
Computations performed during learning depend on the anatomical circuitry within which cell-
biological mechanisms are embedded.
Activity-dependent modulation detects but need not always represent contiguity.
Matrix memories represent contiguity but may not represent predictability.
74 FORMS OF MEMORY

Acknowledgments
I should like to thank E. R. Kandel, N. J. Mackintosh, J. N. P. Rawlins, I. C.
Reid, and D. J. Willshaw for their detailed comments on an earlier draft of this
chapter. I should particularly like to thank Eric Kandel for sending me the
unpublished manuscript of 1965 referred to in the text and for his permission
to reproduce certain previously published figures. This work was supported by
an MRC Programme Grant to R. G. M. Morris and D. J. Willshaw.

Note

1. A careful reading of Carewetal. (1984) points to a qualification of this claim: "We should
emphasize that this does not imply that every potential stimulus is pre-wired to every potential
motor response in the nervous system" (pp. 175-177). At first I wondered if this implied that
they had in mind some other mechanism of learning than alpha-conditioning. However, the
succeeding sentence makes it clear they do not, and that the qualification refers only to the fact
that stimuli and responses with no neural connection between them cannot be learned.

REFERENCES

Abrams, T. W., & Kandel, E. R. (1988). Is contiguity detection in classical conditioning a sys-
tems or a cellular property? Learning in Aplysia suggests a possible molecular site.
Trends in Neurosciences, 11, 128-135.
Alkon, D. L., & Farley, J. (1984). Primary neural substrates of learning and behavioral change.
Cambridge: Cambridge University Press.
Bailey, C. H., & Chen, M. C. (1983). Morphological basis of long-term habituation and sensi-
tization in Aplysia. Science, 220, 91-93.
Bailey, C. H., & Chen, M. C. (1986). Long-term sensitization in Aplysia increases the total
number of varicosities of single identified sensory neurons. Society for Neuroscience
Abstracts, 12, 860.
Bear, M. F., Cooper, L. N., & Ebner, F. F. (1987). A physiological basis for a theory of synapse
modification. Science, 237, 42-48.
Boakes, R. A. (1984). From Darwin to behaviourism: Psychology and the minds of animals,
Cambridge: Cambridge University Press.
Carew, T. J., Abrams, T. W., Hawkins, R. D., & Kandel, E. R. (1984). The use of simple inver-
tebrate systems to explore psychological issues related to associative learning. In D.
L. Alkon & J. Farley (Eds.), Primary neural substrates and behavioural change (pp.
169-183). Cambridge: Cambridge University Press.
Castellucci, V. F., & Kandel, E. R. (1974). A quantal analysis of the synaptic depression under-
lying habituation of the gill-withdrawal reflex in Aplysia. Proceedings of the National
Academy of Sciences, 71, 5004-5008.
Castellucci, V. F., & Kandel, E. R. (1976). Presynaptic facilitation as a mechanism for behav-
ioural sensitization in Aplysia. Science, 194, 1176-1178.
Dickinson, A. (1980). Contemporary animal learning theory. Cambridge: Cambridge Univer-
sity Press.
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 75

Dickinson, A., & Dearing, M. (1979). Appetitive-aversive interactions and inhibitory processes.
In A. Dickinson & R. A. Boakes (Eds.), Mechanisms of learning and motivation: A
memorial volume to Jerzy Konorski (pp. 203-231). Hillsdale, N.J.: Erlbaum.
Frost, W. N., Clark, G. A., & Kandel, E. R. (1988). Parallel processing of short-term memory
for sensitization in Aplysia. Journal of Neurobiology, 19, 297-334.
Garcia, J., & Koelling, R. A. (1966). Relation of cue to consequence in avoidance learning.
Psychonomic Science, 4, 123-124.
Gluck, M. A., & Thompson, R. F. (1987). Modeling the neural substrates of associative learning
and memory: A computational approach. Psychological Review, 94, 176-191.
Greenough, W. T., & Bailey, C. H. (1988). The anatomy of a memory: Convergence of results
across a diversity of tests. Trends in Neurosciences, 11, 142-146.
Grether, W. F. (1938). Pseudoconditioning without paired stimulation encountered in
attempted backward conditioning. Journal of Comparative Psychology, 25, 91-96.
Hawkins, R. D. (1989). A biologically realistic neural network model for higher order features
of classical conditioning. In R. G. M. Morris (Ed.), Parallel distributed processing:
Implications for psychology and neurobiology (pp. 214-247). Oxford: Oxford Uni-
versity Press.
Hawkins, R. D., Castellucci, V. F., & Kandel, E. R. (1981). Interneurons involved in mediation
and modulation of gill-withdrawal reflex in Aplysia. I. Identification and character-
ization. Journal of Neurophysiology, 45, 304-314.
Hawkins, R. D., & Kandel, E. R. (1984). Is there a cell-biological alphabet for simple forms of
learning"! Psychological Review, 91, 375-391.
Hebb, D. O. (1949). The organization of behaviour. New York: Wiley.
Horn, G. (1985). Memory, imprinting and the brain. Oxford: Clarendon Press.
Hull, C. L. (1934). Learning II. The factor of the conditioned reflex. In D. Murchison (Ed.), A
handbook of'general experimental psychology (pp. 392-455). Worcester, Mass.: Clark
University Press.
Jenkins, H. M., & Moore, B. R. (1973). The form of the auto-shaped response with food or
water reinforcers. Journal of Experimental Analysis of Behavior, 20, 163-181.
Kamin, L. J. (1968). Attention-like processes in classical conditioning. In M. R. Jones (Ed.),
Miami symposium on prediction of behavior; aversive stimulation (pp. 9-33). Coral
Gables, Fla.: University of Miami Press.
Kandel, E. R. (1965). The relation of alpha conditioning to classical conditioning. Unpublished
manuscript.
Kandel, E. R. (1976). Cellular basis of behavior: An introduction to behavioral neurobiology.
San Francisco: Freeman.
Kandel, E. R., & Schwartz, J. H. (1982). Molecular biology of learning: Modulation of trans-
mitter release. Science, 218, 433-443.
Kandel, E. R., & Schwartz, J. H. (1985). Principles of neural science (2nd ed.). New York:
Elsevier.
Kohonen, T. (1977). Associative-memory. Heidelberg: Springer-Verlag.
Konorski, J. (1948). Conditioned reflexes and neuron organisation. Cambridge: Cambridge
University Press.
Levy, W. B., Andersen, J. A., & Lehmkule, S. (1985). Synaptic modification, neuron selectivity
and nervous system organization. Hillsdale, N.J.: Erlbaum.
Lynch, G., & Baudry, M. (1984). The biochemistry of memory: A new and specific hypothesis.
Science, 224, 1057-1063.
Mackintosh, N. J. (1975). A theory of attention: Variations in the associability of stimuli with
reinforcement. Psychological Review, 82, 276-298.
Mackintosh, N. J. (1983). Conditioning and associative learning. Oxford: Clarendon Press.
Marr, D. (1982). Vision. San Francisco: Freeman.
McNaughton, B. L., & Morris, R. G. M. (1987). Hippocampal synaptic enhancement and infor-
mation storage within a distributed memory system. Trends in Neuroscience, 10,
408-415.
76 FORMS OF MEMORY

Medawar, P. M. B. (1966). The art of the soluble. London: Methuen.

Merzenich, M. M., Kaas, J. H., Wall, J. T., Sur, M., Nelson, R. J., & Felleman, D. J. (1983).
Progression of change following median nerve section in the cortical representation
of the hand in areas 3b and 1 in adult owl and squirrel monkey. Neuroscience, 10,
639-665.
Morris, R. G. M. (1988). Elements of a hypothesis concerning participation of hippocampal
NMDA receptors in learning. In D. Lodge (Ed.), Excitatory amino acids in health
and disease (pp. 297-320). Chichester: Wiley.
Palm, G. (1982). Neural assemblies: An alternative approach to artificial intelligence. Heidel-
berg: Springer-Verlag.
Pavlov, I. P. (1927). Conditioned reflexes. Oxford: Oxford University Press.
Pearce, J. M., & Hall, G. (1980). A model for Pavlovian learning: Variations in the effectiveness
of conditioned but not of unconditioned stimuli. Psychological Review, 87, 532-552.
Pinsker, H., Kupfermann, I., Castellucci, V. F., & Kandel, E. R. (1970). Habituation and dis-
habituation of the gill-withdrawal reflex in Aplysia. Science, 167, 1740-1742.
Ramon y Cajal, S. (1911). Histologie du systeme nerveux de I'homme et des vertebrates (Vol.
2). Paris: Maloine.
Rescorla, R. A. (1980). Simultaneous and successive associations in sensory preconditioning.
Journal of Experimental Psychology: Animal Behavior Processes, 4, 267-275.
Rescorla, R. A. (1984). Comments on three Pavlovian paradigms. In D. L. Alkon and J. Farley
(Eds.), Primary neural substrates of learning and behavioural change. Cambridge:
Cambridge University Press.
Rescorla, R. A. (1988). Behavioural studies of Pavlovian conditioning. Annual Review of Neu-
roscience, 11, 329-352.
Rescorla, R. A., & Wagner, A. R. (1972). A theory of Pavlovian conditioning: Variations in the
effectiveness of reinforcement and nonreinforcement. In A. H. Black & W. F. Prokasy
(Eds.), Classical conditioning (Vol. 2, pp. 64-99). New York: Appleton-Century-
Crofts.
Rose, S. P. R. (1984). Strategies in studying the cell biology of learning and memory. In L. R.
Square & N. Butters (Eds.), Neuropsychology of memory (pp. 547-554). New York:
Guilford.
Routtenberg, A. (1987). Synaptic plasticity and protein kinase C. In W. H. Gispen & A. Rout-
tenberg (Eds.), Progress in brain research (Vol. 69, pp. 221-234). Amsterdam:
Elsevier.
Rumelhart, D. E., & McClelland, J. (1986). Parallel distributed processing (Vols. 1, 2). Cam-
bridge, Mass.: MIT Press.
Sherrington, C. S. (1951). Man on his nature. (The Gifford Lectures). (2nd ed.). Cambridge:
Cambridge University Press.
Thompson, R. F. (1986). The neurobiology of learning and memory. Science, 233, 941-947.
Wagner, A. R. (1978). Expectancies and the priming of STM. In S. H. Hulse, H. Fowler, & W.
K. Honig (Eds.), Cognitive processes in animal behavior (pp. 177-209). Hillsdale,
N.J.: Erlbaum.
Wall, P. D. (1975). Signs of plasticity and reconnection in spinal cord damage. In Outcome of
severe damage to the central nervous system. London: Ciba Foundation Symposium,
34.
Willshaw, D. J. (1981). Holography, associative-memory and inductive generalisation. In G. A.
Hinton & J. A. Anderson (Eds.), Parallel models of associative memory (pp. 83-104).
London: Academic Press.
Willshaw, D. J., Buneman, O., & Longuet-Higgins, H. C. (1969). Non-holographic associative
memory. Nature, 222, 960-962.
Wood, C. (1978). Variations on a theme by Lashley: Lesion experiments on the neural model
of Andersen, Silverstein, Ritz and Jones. Psychological Review, 85, 582-591.
 Synaptic Plasticity, Neural Architecture, and Forms of Memory 11

Yeo, A. G. (1974). The acquisition of conditioned suppression as a function of interstimulus

interval duration. Quarterly Journal of Experimental Psychology, 26, 405-416.
Zener, K. (1937). The significance of behaviour accompanying conditioned salivary secretion
for theories of the conditioned response. American Journal of Psychology, 50, 384-
403.
4
Forms of Memory
in Pavlovian Conditioning
P E T E R C. H O L L A N D

Psychologists have made few compelling generalizations about memory as a

whole, but a variety of satisfying statements about performances on particular
memory tasks. The conventional psychological approach to this state of affairs
was to view memory as a single system and each task as a way of illuminating
a small set of its attributes. More recently psychologists have taken the view that
there are several memory systems, which may serve different purposes and fol-
low different rules of operation. Furthermore, it has been tempting to suggest
that those systems have separate neural mechanisms and evolutionary histories.
A plethora of "divisions of memory" has been proposed, many based on excit-
ing new evidence from neuropsychology and animal conditioning. These forms
of memory are distinguished along many attributes, such as content (e.g., seman-
tic vs. episodic), type of processing (e.g., working vs. reference), capacity (limited
or unlimited), and endurance (transient vs. permanent). Considerable experi-
mentation shows that memory tasks thought to engage these different systems
are differentially sensitive to certain variables and manipulations.
I do not intend to examine any of the currently popular memory-form dis-
tinctions explicitly, but rather to describe a functional view of memory in ani-
mal conditioning, which links the study of memory to that of perception, motor
skills, and other psychological and neurobiological processes. Within this view,
much memory is an elaboration of sensory and motor processes, the conse-
quence of plasticity and convergence at many steps in the pathways of percep-
tion and action. Thus at some level it begs a multisystems view; it would not be
surprising if tasks that engaged different perceptual and/or motor systems, or
engaged the same systems in different ways or at different loci, would be differ-
entially responsive to certain variables or exhibit different functional character-
istics. At another level, however, such a view demands a more integrating, uni-
tary perspective. It is unlikely that any "memory system" would be completely
independent: memory itself is likely to be both the consequence and the heart
of integration of neural systems.

78
 Forms of Memory in Pavlovian Conditioning 79

PAVLOVIAN CONDITIONING AS A MEMORY TASK

Many claims about memory are limited in applicability to certain tasks. Unfor-
tunately, understanding of what many tasks in vogue are tapping, or what influ-
ences them, is vague. Theories of memory based on research with poorly under-
stood tasks are on shaky ground at best. But limiting one's study to a single task,
no matter how completely understood it may be, is unlikely to reveal much of
the range of memory's functions. For me, the Pavlovian conditioning prepara-
tions described in this chapter are a reasonable compromise between under-
standing and diversity.
Considerable progress has been made since the late sixties in describing the
nature of Pavlovian conditioning, at both psychological and neurobiological lev-
els. Its determinants and effects are at least as well understood as any other psy-
chological phenomenon. Furthermore, it is a rich phenomenon, with the diver-
sity needed to permit a broad view of memory. Although psychologists and
neurobiologists typically focus their analytic efforts on single behavioral conse-
quences of Pavlovian learning experiences, it must be remembered that those
experiences engage multiple behavior systems, often drastically changing the
organism's commerce with its environment. Thus, although Pavlovian condi-
tioning is often caricatured as a trivial learning process, lord only of spit and
twitches, I believe Pavlovian learning is a major basis for the organization and
integration of behavior. In this chapter I describe three lines of research in my
laboratory, with intact adult rats, which I hope capture some of that richness in
ways that are amenable to neural analogies. In the first section I discuss some
traditional and nontraditional views of substitution and event representation,
which emphasize perceptual processing in the memorial consequences of con-
ditioning. Next I describe research that focuses on conditioning as an integrator
of organized behavior systems. Finally I discuss recent work concerning a con-
ditioning function that I call "occasion setting" and the formation of more com-
plex, higher order units in conditioning.

EVENT REPRESENTATION AND MEMORY

In a prototypical Pavlovian conditioning experiment, two stimuli are presented

contiguously in time and independently of the subject's behavior. Usually, one
of those events (the unconditioned stimulus, US) spontaneously produces an
easily measurable change in the subject's behavior (the unconditioned response,
UR), and the other (the conditioned stimulus, CS) does not. Learning is inferred
if pairing the CS and US generates change in the behavior during the CS (the
conditioned response, CR).
In traditional theoretical accounts, Pavlovian conditioning is described as the
substitution of a previously neutral stimulus into an existing reflex system, the
transfer of control of an unconditioned reflex from the US to the CS. This view
80 FORMS OF MEMORY

was encouraged by early observations suggesting that the conditioned response

was a replica of the unconditioned response. Furthermore, this substitution view
is amenable to neurobiological analysis: the underpinnings of learning could be
charted by tracing the activation of the unconditioned reflex system upstream
(e.g., Thompson, 1986) toward the receptor pathways activated by the CS, or by
working downstream from the receptor surface (e.g., Cohen, 1985) until sites of
plasticity are discovered.
The notion that the CS acts as a substitute elicitor of the UR was also perva-
sive in early psychological accounts because it is consistent with both of the two
major theoretical approaches to conditioning debated in the early days of psy-
chology, S-R and S-S. Unconditioned response-like CRs might be evoked
either directly by associations of CS and UR or indirectly by the CS's activation
of a "US center," which in turn elicited the UR. To the modern neurobiologist,
this distinction often seems to have a mystical ring to it, because the S's, R's,
and centers of the psychologist seldom bear much resemblance to recognizable
neural events or systems. Clearly, describing a single "food representation" or
"center" is illusory: food, or any other US, is a multimodal experience. It is hard
to imagine that the rat's nervous system partitions the world to match my con-
ceptual frame. Nevertheless, I believe discussions based on such simple notions
of event representation and on distinctions between S-S and S-R links may help
illuminate the neural organization of learned behavior.
In its simplest form, the S-R versus S-S distinction concerns whether a site
of plastic, functional convergence of CS and US pathways is near the sensory or
motor end of the US-UR reflex. Loosely speaking, how much of the US-UR
machinery is engaged by the CS alone after conditioning?
Figure 4.1 is a fanciful sketch of two reflex pathways, that of a US, say food,
which elicits a reflex response, say salivation, and of a CS, say tone. For the
moment, we are concerned with the US pathway, because most views of Pav-
lovian conditioning concentrate on this particular behavior change (I will cor-
rect this oversimplification in the next section of this chapter). The pathway
imagined here should be viewed as containing whatever sensory input and medi-
ating systems are necessary to produce untrained, reflexive responding. I pre-
sume that these CS and US pathways converge at many levels and loci. For the
learning theorist, the question is, what is the locus of plastic change that relates
these two pathways? Does the CS acquire the ability to directly activate late por-
tions of the US pathway, or does it act early in the US pathway?
Consider two extreme representations. The site labeled S-S reflects a changed
potential for the CS to excite the first US-path unit. Essentially all of the US-
path machinery is thus engaged by CS presentation alone. Thus the rat not only
will salivate when the tone comes on, but also will taste food, feel it, and oth-
erwise experience it in all the ways a rat would experience the food itself. In
short, this path generates a hallucination of food; the tone is a perfect surrogate.
Conversely, the site labeled S-R reflects a changed potential for the CS to excite
the output unit that controls the UR. But the final common path between CS
and US pathways includes only that unit: no other US-patch machinery is
 Forms of Memory in Pavlovian Conditioning 81

FIGURE 4.1. Simple representation of S-S and S-

R associations between CS and US paths.

engaged. The rat that has gained this link because of tone-food pairing salivates
when the tone is presented, but experiences nothing else of US-path activity.
It is unlikely that any significant Pavlovian CR is generated entirely by plas-
ticity at either one of these loci, in all but the simplest systems. Rats probably
do not hallucinate perfect copies of absent USs, nor are final output neurons
likely to be the culprit. It is possible that some plasticity occurs at all levels (e.g.,
Cohen, 1985), but the changes most closely correlated with the emergence of
CRs are likely to be a bit more central. Nevertheless, the question of how much
US-path machinery is engaged by the CS remains a useful one. Considerable
recent data from the laboratories of learning theorists suggest (1) there are mul-
tiple sites of plasticity along the US path, (2) the relative importance of those
various sites is influenced greatly by the nature of the learning experience, and
(3) in many situations, considerably more of the US-path machinery is engaged
by the CS after learning than we were previously willing to admit.
Today, the performance of simple Pavlovian conditioned responses is almost
universally described as mediated by internal representations of the conditioned
and unconditioned stimuli: associations formed between those representations
endow the CS alone with the ability to activate the representation of the US (e.g.,
Rescorla, 1974). For present purposes, a "US representation" refers to a path of
activation of the US-UR, or a portion of it. Within the context of Figure 4.1, if
a CS is said to evoke a US representation, plastic changes more like the one
labeled S-S than the one labeled S-R are presumed to predominate. Thus pair-
ing a tone with food gives the tone the ability to activate a representation of
food, in the absence of food itself. Conditioned responses (CRs) then are not
82 FORMS OF MEMORY

evoked directly by the CS (as would be the case if the site labeled S-R predom-
inated), but rather are mediated by higher portions of the US pathway, the "US
representation." This simple notion has proved useful in describing a number
of features of Pavlovian conditioning, two of which I will summarize next.

Mediated Performance of Conditioned Responses

The ability of previously trained CSs to evoke CRs often depends on the current
status of the US. For example, Holland (1989b) first presented hungry rats with
two kinds of conditioning trials. Tone 1 (T1) was followed by delivery of food 1
(F 1 ) and tone 2 (T2) was followed by food 2 (F2). (The two foods were winter-
green- and peppermint-flavored sucrose, selected so that the rats showed iden-
tical URs and no preferences between them.) Identical conditioned food-related
behavior was established to both tones (left side of Fig. 4.2). Then, in the absence
of either tone, one food was paired with a toxin (lithium chloride). Thus one
former US (F2) now served as a CS in a new Pavlovian relation. The rats rapidly
formed an aversion to F2: as a consequence of F2-toxin pairings, F2 no longer
evoked many aspects of its previous UR, nor would it serve as an effective US
for Pavlovian conditioning of a new CS. Finally, the original tone CSs were pre-

FIGURE 4.2. Results of mediated performance experiment (after Holland, 1989b). The curves
in the left panel show acquisition of conditioned responding during Tt and T2, two tone CSs that
were each paired with distinctive foods, F1 and F2, respectively; the bars show test responding
during those tones, after F2 was paired with a toxin. The center and right panels display data from
a subsequent test session in which the rats received presentations of T, and T2 while consuming
plain water. The center panel shows consumption, and the right panel shows the frequence of
disgust responses in that test.
 Forms of Memory in Pavlovian Conditioning 83

FIGURE 4.3. Representation of possible S-S and S-R associations established in the mediated
performance experiments. —>-= excitatory action; ( = inhibitory action.

sented to the rats, in the absence of the foods or toxins. In this test, T2, whose
F2 associate had been paired with toxin, showed a substantially reduced CR, but
T1,, whose Ft associate had not been devalued, continued to display substantial
CRs (center portion of Fig. 4.2).
This outcome is consistent with the S-S view of conditioning described above,
that is, a fair amount of the US-path activity is engaged by the CS. The poten-
tiated S-S link between T2 and F2 pathways must be upstream from whatever
link is established by F2-toxin conditioning (Fig. 4.3). If instead the T2-F2 link
was downstream from the F2-toxin link (the S-R path), then the T2 alone would
not activate the F2-toxin link, and hence would still evoke the intact CR. Fur-
thermore, because the response evoked by US1 (and T1,) was unaffected, the F2~
toxin plasticity must have been upstream from a final common output path.

Perceptual Nature of Substitution

Apparently the information coded by these CS-evoked representations is quite
detailed. In the experiment just described, the two foods were delivered to the
same food cup and differed only in the flavoring added. Furthermore, the con-
centrations of flavorings were selected so that there were no preferences for one
or the other in groups of rats or for most individual rats. Thus rats interacted
with the two foods in identical ways. Consequently, the rats' very selective use
of the representations of the foods makes it likely that the substitution of tones
for foods is more easily labeled "perceptual" than "motor." However, it is dif-
84 FORMS OF MEMORY

ficult conceptually to distinguish between "stimulus" and "response" informa-

tion in representations or between "sensory" and "motor" interneurons in path-
ways. At what point in processing do we class the response of cell assemblies as
specifying properties of the organism's response rather than its sensory process-
ing? I have made several attempts to deal with this issue at an operational-
behavioral level.
Perhaps the most dramatic method of distinguishing between stimulus- and
response-based mediation in these experiments involved slow-motion video
monitoring of the form of the rats' consummatory responses. Grill and Norgren
(1978) noted that identifiable patterns of responses, which reflect active "accep-
tance" and "rejection," occur to naturally occurring positive (e.g., sweet) and
negative (e.g., bitter) flavors. Furthermore, Pelchat, Grill, Ro/in, and Jacobs
(1983) noted that pairing a formerly accepted flavor with toxin results in the
rejection of that flavor in a characteristic rejection or "disgust" pattern; con-
versely, pairing it with shock results in passive spilling of the flavor but no
acquistion of the disgust response. Pelchat et al. (1983) suggested that these out-
comes indicate that flavor-toxin pairings reduce consumption by making the
flavor taste bad, but flavor-shock pairings reduce consumption less directly by
signaling an upcoming aversive event (see also Garcia, Kovner, & Green, 1970).
Observation of the response forms in my experiments (Holland, 1989b) sup-
ports claims that taste information can be conveyed in these CS-evoked repre-
sentations. In the "mediated performance" experiment just described, I
observed the form of the rats' consumption of a plain sucrose solution (which
contained neither peppermint nor wintergreen), in the presence of tone 2 or tone
1, after food 2 had been devalued. During tone 2, the rats showed more char-
acteristic disgust responses and fewer active acceptance responses than during
tone 1, whose food 1 partner had not been devalued (right side of Fig. 4.2).
Note that in the first phase of the experiment, when the tones were originally
paired with the flavored solutions, no disgust responses were observed during
either tone or food presentations. Furthermore, control experiments (Holland,
1989b) show that tones paired with toxins do not acquire the ability to evoke
disgust responses, unless those tones were previously paired with a flavored food
substance. Thus the tone-associated rejection responses portrayed in Figure 4.2
were likely to be mediated by the tone's activation of perceptual processes nor-
mally evoked by the food itself.
A related experiment (Holland, 1989b) pitted the perceptual and response
consequences of the evoked representations against each other. Rats first
received pairings of tone 1 with sucrose (A) and tone 2 with salt (B). Both of
these solutions evoked acceptance but not disgust responses; thus if consum-
matory responses were directly conditioned to the tones, they would be accep-
tance responses. Next, in the absence of the tones, the rats received pairings of
a compound of A and B with toxin, but unpoisoned presentations of A and B
separately. The rats learned this discrimination easily, so that they readily con-
sumed either A and B when presented separately, but they rejected the AB com-
 Forms of Memory in Pavlovian Conditioning 85

pound. Finally, consumption of plain water was examined in the presence of

tone 1, tone 2, and a compound of tone 1 and tone 2.
If the tones' effects on consumption were mediated by consummatory
responses conditioned directly to the tones, then the compound of the two tones
should evoke even greater acceptance than elicited by either tone alone. How-
ever, if the tones mediated performance by activating perceptual processing of
the sugar and salt, then the tone 1 + tone 2 compound would activate a com-
pound sugar + salt unit. Because in the second phase the rats had learned to
reject the sugar + salt compound, they would accept less water and show more
rejection responses during the compound than during either tone 1 or tone 2.
This latter outcome was observed, indicating that the rats' performance was con-
trolled more by the perceptual (flavor) features evoked by the tones than by any
responses that may have been conditioned to those tones. (A companion exper-
iment, in which in the second phase the AB compound was unpoisoned, but A
and B were poisoned when presented separately, supported these conclusions:
in the test, the tone 1 + tone 2 compound evoked fewer rejection responses but
more acceptance responses than either tone 1 or tone 2 alone.)

Mediated Learning
Consider further the nature of the plasticity represented in the link labeled S-S
in Figure 4.1. By virtue of this link, the CS acquires the ability to activate some
of the upstream processing machinery of the US-UR system. How much? Is the
CS's substitutability for the US limited to its response-evoking role, or does the
CS-activated US-path activity bear closer resemblance to the activity produced
by the US itself?
I have claimed that in some circumstances, activation of a US path or repre-
sentation by a CS may not only influence the performance of learned responses,
but also may permit new learning about that absent US. The basic observation
I have elaborated on is that an aversion to a food substance may be acquired if
a CS-activated representation of that food, rather than the food itself, is paired
with toxin. For example, Holland (198la, 1989b) paired one tone with a win-
tergreen-flavored food and another tone with a peppermint-flavored food. Then,
in the absence of any of the foods, one of the tones was paired with a toxin.
Consistent with the observations of Garcia and Koelling (1966), we found no
evidence of the formation of associations between the tone and toxin, but asso-
ciations apparently were formed between the absent food and toxin. Later tests
of food consumption in the absence of the tones showed the establishment of an
aversion to the food whose tone signal had been paired with toxin. That food
provoked less consumption, less acceptance responding, and more disgust
responses than the other food (Fig. 4.4). Apparently the occurrence of illness
while the representation of one of the foods was activated by the tone was suf-
ficient to establish an aversion to that food, even though the food was never
paired with toxin or a substitute for toxin. In terms of Figure 4.3, activation of
86 FORMS OF MEMORY

FIGURE 4.4. Responding in the con-

sumption test of the mediated acquisi-
tion experiment (after Holland, 1989b).
F, and F2 were distinctive foods which
had been signaled by two tones, T, and
T2, respectively. Before the consump-
tion test, T2 was paired with a toxin. The
bars labeled C refer to consumption,
those labeled A refer to the frequency of
acceptance responses, and those labeled
D refer to the frequency of disgust re-
sponses.

unit US2b by the CS allowed the establishment of the food 2-toxin link, just as
if that unit had been activated by presentation of food 2 itself.
Similarly, Holland and Forbes (1982b) found that the nonoccurrence of illness
in the presence of a CS-evoked representation of an avoided food produced par-
tial extinction of the previously established aversion to that food. As in Hol-
land's (1981a) experiment just described, two tones were first paired with two
foods, presumably endowing each tone with the ability to activate the represen-
tation of its food partner. Each of the two foods next was paired with a toxin, to
condition aversions to them. One of the tones then was presented in the absence
of foods or toxin. Subsequent tests of consumption of the two foods showed a
reduced aversion to the food whose tone partner had been presented in the
absence of toxin in the previous phase. Apparently nonreinforcement of a CS-
evoked representation of a food substance was sufficient to at least partially
extinguish a previously trained aversion to that substance, just as nonreinforce-
ment of that food itself would have done. (Note that a similar technique is often
used in human antiphobic treatment: patients are taught to relax while imagin-
ing scenes assoicated with fear, with the hope that the replacement of fear by
relaxation in the presence of imaginary surrogates of fearful situations will trans-
fer to the real situations.)
Clearly, the activation of an event representation or pathway by an associate
can do more than just mediate conditioned responding: the experiments just
described show that both the establishment of new associations to an event and
the extinction of old associations can occur in the absence of the event itself, if
a previously established surrogate for that event is present during experimental
manipulations. Other experiments from my laboratory have shown that evoked
representations of events can substitute for those events in a variety of other
condition functions. I mention three examples briefly.

Mediated Overshadowing
Typically, a stimulus, A, acquires less conditioned responding if it is reinforced
in the presence of another cue (e.g., AX) than if it is reinforced alone. This inter-
ference by X of acquisition to A is pervasive (but see below) and has played an
important role in the construction of recent conditioning theories (e.g., Mack-
 Forms of Memory in Pavlovian Conditioning 87

intosh, 1975; Pearce & Hall, 1980; Rescorla & Wagner, 1972). A recent series of
experiments (Holland, 1983b) showed that conditioning of a flavor aversion to
a food substance can be overshadowed by a CS-evoked representation of another
flavor. In three groups of rats, a tone was first endowed with the ability to evoke
a representation of a flavor, X, by presenting several tone-X pairings. Then, in
group A, another flavor, A, was paired with toxin; in group AX, a compound of
the flavors A and X was paired with toxin, and in group AT + , a compound of
the flavor A and the tone (which presumably evoked a representation of the
flavor X) was paired with toxin. Finally, consumption of flavor A alone was
examined. There was less aversion to A in the groups that received either the X
+ A compound (AX) or the tone + A compound (AT + ) in conditioning. Thus
conditioning of A was overshadowed by both the flavor X and its tone-evoked
representation. Another comparison showed that the overshadowing in group
AT+ was the consequence of the flavor representation evoked by the tone, not
simply the presence of the tone: there was less aversion to A in group AT+ than
in another group (AT—), which received tone + A paired with toxin in phase
2, but for which the tone did not activate a representation of flavor X, because
it did not receive tone-X pairings in the first phase.

Mediated Potentiation
Although overshadowing is the typical outcome when a compound stimulus is
reinforced, in some circumstances the opposite occurs. For example, although
conditioning of an aversion to a flavor A is overshadowed when it is com-
pounded with another flavor, X, conditioning to an odor A is potentiated when
it is compounded with a flavor. That is, more conditioning of the A odor is
observed if it is accompanied by a flavor in conditioning than if it is conditioned
alone. Holland (1983b) duplicated the experiment just described, but substituted
an odor A for the flavor A. In that experiment, presence of the flavor X in group
AX augmented the aversion established to A rather than reducing it. More
important, in group AT + , the aversion to the odor A was also potentiated by
the presentation of the tone, which presumably evoked a representation of the
flavor X. Thus, once again, the tone-evoked representation of a flavor substi-
tuted for the flavor itself: in circumstances in which the flavor X overshadowed
conditioning to A, the evoked representation of X also overshadowed, but in
circumstances where X potentiated conditioning to A, so did the evoked repre-
sentation of X.

Mediated Occasion Setting

Finally, evoked representations of events can substitute for those events in mod-
ulating or setting the occasion for other associations, as will be described in a
later section of this chapter. To anticipate a bit, cues presented in particular tem-
poral and logical relations to other pairs of events can acquire a conditional con-
trol function that seems relatively independent of simple associative functions.
Thus a cue's power to signal when a relation between two other cues is in effect
is separable from its power to signal one or the other of those events.
88 FORMS OF MEMORY

Holland and Forbes (1982a) first endowed two visual cues with the ability to
evoke representations of two flavored foods. Then the rats were trained to use
those flavors to indicate when a tone would be followed by water: when a tone
was preceded by one flavor, water was delivered, but when the tone was pre-
ceded by the other flavor, water was not delivered. In a final test, the tone was
preceded by the visual cues but not the flavored foods. Performance to the tone
after each visual cue mimicked the performance observed when the tone was
preceded by the corresponding flavored foods. Thus the occasion-setting func-
tion of those flavored foods was also controlled by CS-evoked representations
of those foods.

Limits on Event Representation

The previous discussion implies an interchangeability of evoked representations
and their referents, in at least five functions: (1) evocation of CRs by CSs, (2)
acquisition and extinction of new learning to cues, (3) interference with condi-
tioning between other, coincident CSs and USs (overshadowing), (4) catalyzing
the formation of associations between other CSs and USs (potentiation), and (5)
setting the occasion for the action of other associations. But it seems unreason-
able to assume that organisms do not distinguish between events and evoked
representations of those events. In fact, most of my data suggest that evoked
representations of events and the events themselves are not processed
identically.
First, the effects produced by event surrogates are considerably smaller than
those produced by the events themselves. Second, several experiments (e.g.,
Holland, 198la, 1983b, 1989b) indicate that the stimulus that activates a rep-
resentation is not affected by the activated representation. For example, Holland
(1983b, 1989b) established an odor cue X as an evoker of a representation of a
flavor cue F by presenting an FX compound. Later, a compound of odor X and
another odor, Y, was paired with toxin. The aversion to Y was potentiated (rel-
ative to subjects for whom odor X had not been paired with the flavor), presum-
ably by the flavor representation evoked by X. However, despite X's also being
present during activation of the flavor representation, the aversion to X was not
potentiated. In another experiment, although a flavor representation evoked by
A overshadowed conditioning to another flavor, B, when an AB compound was
reinforced, it did not overshadow conditioning to A. Finally, after A-X pairings,
such that A evoked a representation of X, presentation of an A-B compound,
under some circumstances, established B-X associations but extinguished the
A-X associations (Holland, 1989b).
In all cases, presentation of X itself (rather than its evoked representation)
coincident with the AB compound had similar effects on both A and B. Thus X
and its evoked representation are not treated identically: the evoker of a repre-
sentation is uniquely immune to modification by the representation. In some
fashion, a CS's current activation of a US path apparently prevents further plas-
 Forms of Memory in Pavlovian Conditioning 89

ticity between those two paths at that time, except for whatever changes lead to
extinction. Elsewhere (Holland, 1989b), I have described how this limitation
may be diagnostic as to the nature of action of evoked representations.
Finally, despite the many instances of substitution of signals for their referents
documented above, this substitution is by no means universal. The experiments
described were carefully designed to demonstrate such effects, to assure activa-
tion of upstream portions of the US circuitry. Many variations in procedure can
reduce the mediational effects I have described. Two variations are especially
noteworthy. The first concerns the salience or associability of the evoked rep-
resentation relative to that of events that are physically present. Typically, a real
event, being more salient than an evoked representation of an absent event, is
likely to overwhelm. My success at demonstrating a variety of mediational
effects in the experiments described in this section is due in part to selection of
a situation in which the real events (tones, lights, and odors) that evoke the fla-
vor representations are themselves only poorly associable with the toxin,
whereas the represented flavors are well associated (e.g., Garcia & Koelling,
1966). Elsewhere (Holland, 1985), I showed that several manipulations that
increase the relative salience of other aspects of the training context (e.g., real
events, response patterns) can eliminate evidence for mediational effects, even
for the normally robust case of mediated performance of CRs. Apparently these
manipulations encourage plastic convergence of CS and US pathways more
downstream, that is, more S-R links.
Second, I have found that at least some of these mediational effects vanish if
the original training designed to permit one stimulus to evoke a representation
of another is prolonged (Holland, 1989b). For instance, with extended tone-food
training, the tone loses the ability to substitute for that food in learning a food
aversion if the tone is paired with a toxin. Although the response to the tone CS
is maintained, the tone seems to lose its ability to activate a functionally effec-
tive representation of the food. This loss of plasticity with more experience has
a variety of parallels in psychology (e.g., Allport, 1937; Kimble & Perlmuter,
1970) and neurobiology (e.g., Held & Hein, 1963), and implies functional differ-
ences between the effects of newly altered and extensively altered pathways.
Quite simply, with repeated execution of the CS-UR complex, more down-
stream portions of the circuit become more highly tuned and come to
predominate.
These outcomes suggest that plasticity may occur at many levels within a
response system, depending on the precise nature of the conditioning experi-
ence, and the resultant behavior may have different characteristics and sensitiv-
ities to various manipulations, depending on the locus of plasticity. Although
the mediational effects just described may not be a major determinant of behav-
ior in all traditional conditioning paradigms, they have important implications
for the mechanisms of conditioning and integration of behavior systems. At the
very least, these mediational processes permit the integration of behavior sys-
tems that many have claimed are normally distinct, both functionally and ana-
tomically (e.g., Garcia, Lasiter, Bermudez-Rattoni, & Deems, 1985).
90 FORMS OF MEMORY

MULTIPLE BEHAVIOR SYSTEMS IN PAVLOVIAN CONDITIONING

The experiments discussed in the previous section were concerned primarily

with conditioned behaviors that were elicited by the US prior to conditioning.
Perhaps because these behaviors are most consistent with the simplest "reflex
transfer" accounts of Pavlovian conditioning, most psychological and neurobi-
ological research has focused on them. But Pavlovian conditioning procedures
usually engender more complex behavioral adaptations than a simple substitu-
tive replica of the UR. Very early in the history of conditioning experiments it
was noted that the CR often lacked obvious features of the UR; for instance,
swallowing and jaw movements were typically not observed in canine salivary
conditioning, although these behaviors form a prominent part of the UR (Zener,
1937). Conversely, frequently the CR included behaviors that were not part of
the UR, for example, motor activity often accompanied signals for food even
though that activity is typically not evoked by food delivery itself (e.g., Zener,
1937). And some CRs even involve behavior changes that seem opposite to
those produced by the US. For example, although foot shock usually elicits a
jump and heart rate acceleration, a signal for that shock generally evokes freez-
ing and heart rate deceleration (e.g., DeToledo & Black, 1966; see Cohen, 1985).
The occurrence of these so-called compensatory responses is especially prevalent
when the US is the administration of toxic drugs. For instance, Siegel (1977)
reports that although the rat's UR to morphine includes hyperthermia and hypo-
algesia, the CR to signals for that US include hypothermia and hyperalgesia.
Deviations from perfect substitutability of CS and US may help illuminate
conditioning mechanisms. Any "US" is likely to be a multifaceted event, gen-
erating various behaviors through many paths. The lack of some component of
the UR in the CR might reflect a lack of plasticity in the circuitry involved in
generating that particular response. Although that rigidity might be intrinsic to
that path, it could well be that the conditioning parameters ideal for condition-
ing one component are not ideal for another component. For example, Kiery-
lowicz, Soltysik, and Divac (1968) suggested that the absence of swallowing in
many salivary conditioning experiments was not the consequence of a general
nonplastic nature of that response, but an artifact of delivering food to a dish;
swallowing was simply too temporally separated from the CS to be easily con-
ditioned to it. They found that if food was instead delivered to the back of the
mouth, robust swallowing responses occurred to the CS. Even CRs opposite to
URs might be explained in this manner: many USs evoke both immediate pri-
mary responses and delayed compensatory responses that might be viewed as
serving a homeostatic function (e.g., Eikelboom & Stewart, 1982). It would not
be unreasonable to anticipate greater plasticity of the secondary response in
some circumstances, particularly those in which anticipating and countering
homeostatic disruptions confer adaptive advantage.
An especially striking influence on the nature of the behavior obtained in Pav-
lovian conditioning experiments is the nature of the conditioned stimulus. CS
parameters such as modality, localization, duration, and CS-US interval have
 Forms of Memory in Pavlovian Conditioning 91

been reported to affect the form of a variety of conditioned responses. Although

it is inconsistent with simple reflex transfer schemes and was initially regarded
as a somewhat unusual outcome, I have claimed that CS determination of CR
form is a basic and important feature of many conditioning paradigms (Holland,
1977, 1984).
The most systematically studied example of the influence of CSs on the form
of the CR comes from my direct observations of rats' behavior anticipatory to
food (summarized in Holland, 1984). Most of those experiments compared the
responses evoked by various auditory and visual CSs paired with food delivery.
Typically, localizable visual cues evoked first rearing on the hind legs and vari-
ous attempts to contact the CS source followed by passive standing at the food
magazine, but auditory cues evoked a startle response followed by short, rapid
movements of the head (head jerk) and some passive standing at the food
magazine.
Some of these responses, like magazine behavior, were never observed to
occur during the CSs alone prior to conditioning. However, other behaviors,
especially the rear and startle responses, were observed to occur to the CS prior
to conditioning, although at relatively low levels. I call these behaviors "orient-
ing responses" (ORs). If the CS was not paired with the US, then these orienting
responses rapidly habituated. But if the CSs were paired with food, then the ORs
increased in frequency. These increases were found to be the consequence of the
Pavlovian CS-US relation, rather than the mere presentation of the US (sensi-
tization) or operant response-reinforcer relations.
Consequently, I (Holland, 1977, 1984) distinguished between two broad
classes of CR components, one determined by and appropriate to the US, and
the other an enhancement of the response elicited by the CS prior to condition-
ing. Although both classes of response were found to be the consequence of the
same Pavlovian CS-US relation, considerable evidence suggested they were
often quite separable. First, the ORs occurred primarily early in the CS-US
interval, and the UR-like CRs near the end. Second, if CS onset (which was the
reliable generator of an OR) was made less relevant to delivery of the US than
were later portions of the CS, by extending the CS-US interval or by intermixing
short, nonreinforced CSs with longer, reinforced ones, the early, OR-like behav-
iors were substantially reduced in frequency. Third, changes in the nature of the
CSs affected the nature of responding early in the CS intervals but had relatively
little effect on late CS behavior, thought to be determined by the US. Conversely,
changes in the nature of the unconditioned stimulus affected the nature of
behavior late in the CS-US intervals but had little effect on responding during
early portions of the CS, though to reflect an OR. Fourth, several experimental
manipulations, such as varying US magnitude or probability (Fig. 4.5), the
imposition of omission contingencies, the use of partial reinforcement contin-
gencies, and the postconditioning devaluation of the reinforcer (to be discussed
in detail later) differentially affected the putative CS- and US-generated
behaviors.
The scheme outlined in Figures 4.1 and 4.3 concerned only US-based behav-
92 FORMS OF MEMORY

FIGURE 4.5. Effects of varying the

probability (left panel) and magnitude
(right panel) of reinforcement on CRs
and conditioned ORs. P = partial re-
inforcement; C = consistent rein-
forcement; L = large; S = small.

iors, that is, changes that permitted the CS to engage some aspects of the US-
UR machinery. The recognition of an independent CS-generated component to
the CR demands an addition to those schemes. That addition distinguishes the
US's role in providing the machinery for response production from its role in
potentiating links: for the US-generated behavior class, the US both potentiates
the link (between CS and US path) and provides the machinery for response
production. But for the CS-generated behavior class, the US serves only as a
potentiator for responding that is determined by another pathway. The simplest
possibility would be to imagine that contiguity of CS and US potentiates some
level(s) of the CS-OR path (Fig. 4.6, heavy arrow); this possibility makes obvi-
ous the independence of OR and UR components of the CR.
A more reasonable possibility is that the potentiation of the OR is mediated

FIGURE 4.6. Representation of associations between CS and US

pathways that could underlie the acquisition of CRs and conditioned
ORs. The symbols are explained in the text.
 Forms of Memory in Pavlovian Conditioning 93

FIGURE 4.7. The frequency of CRs

and conditioned ORs during the tone
conditioned stimulus, after its food as-
sociated was paired with lithium chlo-
ride (bars labeled L), rotation (R), or
neither event (C). (After Holland and
Straub, 1979)

by activity in the US path (the modulatory link in Fig. 4.6). That is, potentiation
of the CS-OR path is conditional on activity of both CS and the unit USb. Thus
changes in the activation of unit USb would affect not only US-path behaviors,
but also CS-OR behaviors. I suggest this latter path because the bulk of our
evidence indicates that the performance of OR behaviors is also mediated in
this fashion: postconditioning devaluation of the US either by pairing it with
rotation (Holland & Straub, 1979; Fig. 4.7) or by satiation on that particular
foodstuff (Holland, 1981b) lowered the probability of those responses as well as
US-generated ones.
Holland and Straub (1979) noted that devaluation of the US by pairing it with
one toxin, lithium chloride, eliminated US-based behaviors but left CS-deter-
mined behaviors unscathed (see Fig. 4.7). This outcome suggests that devalua-
tion by food-LiCl pairing either modulated activity of the US pathway at some
point after that pathway potentiates CS-path activity (shown in Fig. 4.8) or
affected a parallel pathway that mediates the UR-like CR component but does

FIGURE 4.8. Representation of associations established in Hol-

land and Straub's (1979) experiment. OR = conditioned orienting
response; ROT = rotation; -> = excitatory action; —( = in-
hibitory action; —• = facilitory action.
94 FORMS OF MEMORY

not potentiate CS-OR relations. It further points out that whether a given site
of plasticity is labeled S-R or S-S by the psychologist depends on the methods
and responses used to assess mediation: with OR-like CRs, tone-food plasticity
at site USa would be labeled S-S if food was devalued by association with rota-
tion but S-R if food was devalued by pairings with toxin.
Other evidence also supports the view that different CRs evoked by various
CSs nonetheless have a common mediator within the US pathway. Briefly, the
data show that despite their different CRs, these CSs are functionally inter-
changeable in conditioning experiments. For example, Holland (1977) found
that pretraining a CS that evoked one response enabled that CS to block con-
ditioning of a very different response to another CS when the pretrained and
new CSs were reinforced in compound. Because considerable data suggest that
the pretrained stimulus blocks conditioning involving only the US with which
it was pretrained, the occurrence of blocking despite different CRs suggests the
various CSs are associated with a common US. Similarly, the power of a CS to
serve as a reinforcer for second-order conditioning to another cue seems unre-
lated to the precise response that it evokes: Holland (1977) noted similar forms
of responding to an auditory cue, whether it was paired with another auditory
cue that evoked the same conditioned behaviors or with a visual CS that evoked
entirely different behaviors. Furthermore, the reinforcing powers of two CSs
sum when those CSs are combined, regardless of whether the CRs they evoke
are similar or dissimilar. Thus considerable commonality of learning must exist,
despite the different CRs.

Specificity of CS- and US-Path Changes

The plasticities described in the last two sections have all been specific to a CS-
US pair. That is, conditioning operations were said to potentiate particular links
between a CS and a US path, the associations of the learning theorist. The
changes were not general enhancements or inhibitions of CS or US pathway
activity. For example, the enhanced orienting response to a visual stimulus
paired with food depends on specific learning: it is not enough that the US path
be activated (say, by another cue that had been previously paired with food)
while a stimulus is presented (Holland, 1977, 1981b).
Most modern learning theorists claim that conditioning operations also may
modulate the effectiveness of a CS and/or US pathway in general. For example,
many hypothesize that conditioning operations modulate processing of the CS
in some way, such as enhancing attention to it. This enhanced attention to the
CS might be described as a potentiation of one or more links within the CS path,
which in turn could serve to enhance the OR (alpha-conditioning), as well as
any responses it controls by virtue of links with the US path. This modulation
in attention to the CS, while perhaps established concomitantly with CS-US
associations, is independent of that link (much like the heavy arrow in Fig. 4.6).
Consequently, action of that CS is modulated, regardless of changes in the US.
 Forms of Memory in Pavlovian Conditioning 95

For example, that CS might be more effective as a CS even if it were retrained

with another US.
Other theorists (e.g., Rescorla & Wagner, 1972; Wagner, 1981) have suggested
that conditioning operations modulate activity of the US path in general, as well
as establishing particular associations. For example, a frequent claim is that
"expected" USs are less effective than surprising ones, in terms of their rein-
forcing power, response-generating capacity, and endurance in memory. In
terms of the pathway jargon, if the US pathway is already active, by virtue either
of immediately prior presentation of that US or by presentation of a CS that is
capable of activating that path because of past associations, then presentation of
that US will have less effect than if that path was not already active (Wagner,
1981). Thus a US that is anticipated because it was preceded by a previously
paired CS is unlikely to be an effective reinforcer for any cue that may occur
contiguously. This phenomenon, called blocking (Kamin, 1968), has been dem-
onstrated hundreds of times, even in simple systems.
Another example of hypothesized modulation of the US path comes from the
study of conditioned inhibition. Consider a procedure in which a CS, A, is
paired with the US, and a compound of A and another stimulus, B, is nonrein-
forced. Rats rapidly learn the discrimination, responding on A alone trials and
not on AB trials. Rescorla and Holland (1977) suggested that B acquired the
power to inhibit the action of the US path. If so, then B should inhibit the CR
to any CS whose responding was mediated by that US path, but not the CR to
CSs whose responding was mediated by some other US. Rescorla and Holland
found those outcomes.

Other Behavior Systems

I have emphasized distinctions between CS and US path behaviors, mainly
because I have investigated them extensively behaviorally and because they
seem amenable to the type of analysis I propose here. It should be recognized
that this dichotomy is just the tip of the iceberg. Historically, different classes of
conditioned behavior (e.g., emotional vs. skeletal) have been posited as having
separate conditioning loci (e.g., Rescorla & Solomon, 1967; Thompson, 1986).
In addition, it seems likely that the circuits of the US paths are well integrated,
so that the use of any particular US will engender not only the precise response
evoked by that particular event, but also a more complex set of naturally occur-
ring behavior patterns appropriate to that US.
For instance, feeding is a complex activity containing many separate, sequen-
tially organized action patterns such as individual foraging, social approach and
food sharing, investigation, predation, food handling, hoarding, and ingestion
or rejection. Conditioning may occur at many levels and points within a behav-
ior system; the precise conditioned behaviors observed might depend on the
nature of the cues used and their resemblance to cues naturally used in the feed-
ing situation. In the language of Tolman (1932), the CR is more appropriate to
96 FORMS OF MEMORY

the expectancy of food engendered by the CS than to food per se. For example,
the use of short CSs would be likely to encourage behavior appropriate to immi-
nent food, such as mouthing and search of the food cup, but longer CSs would
encourage more early chain behaviors, like searching. Those CSs, by virtue of
their timing, tap into different portions of natural feeding sequences. The form
of the CR would be further filtered and tuned by other features of the cues. For
example, the use of static cues like tones and lights might encourage investiga-
tion and consummatory responses, but using small moving objects as CSs might
encourage predatory activity, and use of conspecifics as cues might lead to social
food-sharing activities. Experiments conducted in my laboratory (Holland,
1984) and, more extensively, by Timberlake (1983) support these claims. In
essence, Pavlovian conditioning often acts as a mechanism of modulating nat-
urally organized behavior patterns.
Of course at this point I have left far behind the plausibility of neurobiological
analysis. But more feasible investigations of simpler interacting systems, for
example CS- and US-generated behaviors, may provide a reasonable prototype
of the more complex interactions that comprise adaptive behavior.

HIGHER ORDER UNITS

In the previous sections I have discussed very simple conditioning arrange-

ments, which usually involved only one CS and one US. In contrast, most the-
oretical and empirical investigation of Pavlovian conditioning has concerned
compound CSs (e.g., AB). Nevertheless, most of that work focused on associa-
tions between each individual event and the US, or more recently (e.g., Rescorla,
1985a), on associations between the elements (A and B). This work thus con-
cerns the same event-event relations I discussed earlier.
Interest has been growing in other potential conditioning functions that might
be involved when compound cues are used, especially in solving conditional
discriminations in which the meaning of one cue depends on the value of
another cue. In this section I briefly discuss two such functions, occasion setting
and configuring. They are of special interest to the topic of this book because
they are directly related to some of the more articulate memory distinctions in
animal learning.

Occasion Setting
One of the simpler conditional discriminations is the feature positive discrimi-
nation, in which a compound stimulus (XA) is reinforced, but one of its ele-
ments (A) is separately nonreinforced. If X and A are presented simultaneously
on compound trials, X comes to evoke a CR based on its association with the
US, and A controls little or no responding. However, if the onset of X precedes
that of A, X acquires the ability to modulate the action of an association that is
 Forms of Memory in Pavlovian Conditioning 97

formed between A and the US. Thus X "sets the occasion" for responding to A
(Holland, 1983a). After serial training, X gates or enables the A-US link, per-
mitting A to activate the US path, whereas after simultaneous training, X
directly activates the US path and A does not.
The results of several kinds of experiments support this distinction. First,
Ross and Holland (1981) found that the form of the CR differed in serial (S ->
A+, A-) and simultaneous (XA + , A-) feature positive discriminations. With
simultaneous compounds, the form of the CR was determined by X, as would
be anticipated if the CR were the consequence of X-US associations. But with
serial compounds, A determined the nature of the CR, consistent with the view
that X acted by modulating the effectiveness of an A-US association.
Second, several manipulations differentially affect the acquisition and extinc-
tion of these occasion-setting and simple associative functions. For example,
increasing the interval or inserting a stimulus-free trace between X and the US
reduces the strength of X's association with the US but (up to a point) increases
its ability to set the occasion for responding based on the A-US association
(Holland, 1986a; Ross & Holland, 1981). Similarly, Ross (1983) and Ross and
LoLordo (1986) found that A -> US pairings prior to feature-positive discrimi-
nation training blocked the acquisition of X-US associations but not X's ability
to modulate responding based on A-US associations. Similarly, Ross (1983)
found that enhancing X-US associations by prior X-US training interfered with
X's acquisition of occasion setting (see Rescorla, 1986b). Finally, nonreinforced
presentations of X alone quickly abolish its ability to elicit CRs by virtue of its
simple association with the US (or A) but have relatively little effect on X's abil-
ity to set the occasion for responding to A (Rescorla, 1986a).
Third, cues trained in serial and simultaneous discriminations acquire differ-
ent properties. For instance, Ross and LoLordo (1987) found that an X that was
previously trained in a serial discrimination blocked the acquisition of operant
stimulus control to another cue when a compound of that cue and X were
trained as a discriminative stimulus, but an X trained in a simultaneous dis-
crimination had no such blocking effect. Finally, Holland (1986b) found that
whereas the simple associative powers of an X trained in a simultaneous dis-
crimination were manifest regardless of whether it was accompanied by the orig-
inal A cue or some other cue, the occasion-setting powers of an X trained in a
serial discrimination were greatly diminished if X was combined with cues other
than A.
These observations, together with analogous ones from feature-negative dis-
criminations, in which an XA compound is nonreinforced while A is separately
reinforced, suggest that simple associative and occasion-setting functions are
quite dissociable. Interestingly, these functions may be anatomically distinct as
well. Ross, Orr, Holland, and Berger (1984) trained rats on serial-feature positive
discriminations of the form light -> trace -> tone ->food; tone -> no food, which
endowed the light with the ability to set the occasion for responding to the tone
and with the ability to evoke CRs based on direct light-US associations and on
98 FORMS OF MEMORY

within-compound light-tone associations. Additionally, the same rats received

training on a nonconditional discrimination between a reinforced clicker and a
nonreinforced noise.
Rats that received bilateral hippocampectomy prior to training were unable
to acquire the serial feature-positive discrimination (occasion setting), and rats
that were trained intact and then hippocampectomized lost the ability to per-
form that discrimination (Fig. 4.9). Conversely, relative to unoperated controls,
hippocampectomy completely spared learning and performance of CRs based
on the light-food association and on the clicker-noise discrimination, and it had
no more deleterious effect on CRs based on the light-tone association than had
lesions of the neocortex overlaying hippocampus (which was also destroyed in
hippocampectomized rats). Thus the behavioral-functional independence of
simple associative and occasion-setting powers of CSs was complemented by a
degree of anatomical independence.
The nature of the behavioral deficit observed in these experiments deserves
further attention. First, it was not due to a simple short-term memory deficit—
an inability to learn across a trace interval—because both the light-food and
light-tone learning, which required bridging the same trace interval, were unaf-
fected. Second, the occasion-setting deficit cannot be attributed merely to its
being a more difficult task, because there was little difference in the performance
levels of the conditional (serial feature positive) and nonconditional discrimi-
nations. Furthermore, Orr and I (unpublished) subsequently found that hippo-
campectomy had no effect on the acquisition of a difficult nonconditional dis-
crimination (between two tones of similar pitches) but prevented the acquisition
of a much easier (for unoperated controls) conditional discrimination.

FIGURE 4.9. Effects of hippocampal lesions on performance in conditional (closed bars, labeled
C) or simple (open bars, labeled S) discriminations. The bars labeled PRE refer to performance
during the last 12 training sessions when the lesions were made prior to training, and those labeled
POST refer to performance during the last 12 retraining sessions when the lesions were made after
training.
 Forms of Memory in Pavlovian Conditioning 99

Finally, hippocampectomy had relatively little impact on performance during

the serial light -> trace -> tone compound: the rats acquired the CRs appropriate
to that compound in a practically normal fashion. The deficit came in their being
unable to withhold responding during the nonreinforced tone-alone presenta-
tions. Note, however, that there was no global deficit in inhibition: the rats' abil-
ity to withhold responding to the nonreinforced cue in the nonconditional dis-
crimination was unimpaired. This data pattern implies that hippocampal
systems do not involve only processing of the compound, for example, forming
of a higher order light + tone configuration. If these systems involved only pro-
cessing of the compound, then hippocampectomy after training would have
been expected to abolish responding to the compound.
Our findings are compatible with a variety of theoretical languages that have
been used to describe hippocampal memory function (see Ross et al., 1984).
Common to most of these languages is the notion that the simple pairings of CSs
and USs engage a nonhippocampal memory system that stores only information
about the strength of that association, but that certain more complex problems
engage a hippocampal memory system that makes use of some other kind of
information or process. One possibility that appeals to me is that hippocampus
acts as a selector (e.g., Teyler & DiScenna, 1985) which disambiguates cues that
have multiple histories. That is, the tone may have been represented as both
reinforced and nonreinforced, and the light served to select the reinforced path.
Supporting this claim is recent evidence (Holland, 1986b; Lamarre & Holland,
1987; Rescorla, 1985b) that shows particular sorts of ambiguity in the treatment
of the target (tone in the previous example) to be critical to the occurrence of
occasion setting, as opposed to simple association.
This view of hippocampus as modulator or selector is a very Jacksonian one:
the main purpose of higher (or at least synaptically distant) structures like hip-
pocampus is to improve the precision of processes that can be performed else-
where. It implies that the memory for the conditional discrimination, despite
requiring intact hippocampal circuitry to be acquired or displayed, is probably
not really localized there. Although hippocampal systems may enable formation
of separate representations for reinforced and nonreinforced trials, which may
or may not importantly involve hippocampal circuitry itself, essential memory
processing goes on in other portions of the perceptuai-motor paths, as always.

Configurations
The simplest strategy for solving the feature-positive discriminations just
described (XA + , A-) is to associate X and the US. A more complex strategy is
to use X to signal when A is reinforced, which I called occasion setting (above).
In a sense, the occasion-setting strategy for solving those discriminations
involves the construction of an A-US unit, which in turn is signaled or selected
by another cue, X. A third way to solve the problem is to treat the compound
as a single unit or configuration, light + tone, which is discriminated from the
tone alone. Considerable evidence suggests that rats use the first strategy when
100 FORMS OF MEMORY

X and A are presented simultaneously but use the second when X precedes A
on compound trials. There is no evidence that rats ever use the configurational
strategy to solve this problem.
However, consider a discrimination in which an XA compound is reinforced,
but neither X nor A alone is reinforced. This positive-patterning discrimination
cannot be mastered with the simple associative strategy. If X precedes A on
compound trials, rats use the occasion-setting strategy described earlier and find
the task no more difficult than the feature-positive discrimination (Holland,
1989a). But if X and A are presented simultaneously, the positive-patterning
discrimination is far more difficult than the feature-positive discrimination.
Considerable evidence indicates that to solve the former discrimination, rats
associate a light + tone configuration with the US (Holland & Block, 1983; Res-
corla, 1972). Although the rats may simply attend to a weak feature unique to
that compound (Rescorla, 1972), it is tempting to suggest that they may actively
form a higher order XA unit or superfeature, which is then associated with the
US (Razran, 1971). Either way, this strategy is quite different from the occasion-
setting strategy in that no special "selector" function seems to accrue to the cues
involved. Furthermore, I (Holland, 1989a) have found that several manipula-
tions that make the solution of simultaneous patterned discriminations easier
(and by implication, encourage configuring), such as increasing X-A similarity,
make the solution of serial patterned discriminations more difficult (and hence
discourage occasion setting).
Studies of configurating in learning and memory (Lamb & Riley, 1981) share
much with the study of perception. My guess is that the sort of higher level unit
formed in the experiments just described is perceptual in nature, an elaboration
of early path activity, whereas the unit formed in the occasion-setting experi-
ments is an elaboration of later stream activity. Thus, although the serial and
simultaneous positive patterning discriminations are logically equivalent con-
ditional discriminations, I would anticipate important hippocampal involve-
ment in solving the former, but not the latter.

CONCLUSIONS

What do these data have to do with "forms of memory"? It could be argued that
they distinguish at least three pairs of subsystems. In the first section I distin-
guished between S-S and S-R links, which some psychologists (e.g., Dickinson,
1980) have characterized as declarative ("knowing what") and procedural
("knowing how"). Then I contrasted CS-OR and CS-UR paths to conditioned
behavior, and finally I distinguished simple associative and occasion-setting
functions.
In each case I showed considerable evidence for the independence of subsys-
tems. That evidence was in the form of functional dissociations of the perfor-
mance indicators of those memory skills. Several variables affected performance
based on S-S but not behavior based on S-R links, enhanced CS-OR respond-
 Forms of Memory in Pavlovian Conditioning 101

ing but reduced CS-UR responding, or vice-versa, or differentially affected sim-

ple associative and occasion-setting functions of CSs. Just these types of disso-
ciations have led others to propose the existence of multiple independent
memory systems.
However, not everyone agrees that such data patterns demand major memory
system distinctions. For example, Roediger (1984) argued for the insufficiency
of dissociation outcomes in inferring independent memory systems by claiming
that some task demands may enourage the use of one subset of processes (say,
A and B) from a memory system, whereas another task may demand the use of
a different subset of processes (say, B and C) from the same memory system. But
should this be viewed as a single memory system, two separate systems that
share a subsystem, or three separate systems that occasionally work together,
depending on the demands of the moment? To a large extent, the adequacy of
any methodology in inferring multiple memory systems depends on how "mem-
ory systems" are viewed.
Tulving (1985) argued that stochastic rather than functional independence of
two memory task measures provides better evidence for independence of two
memory systems. Briefly, if a given memory task engages two completely sepa-
rate memory systems, then the product (measure) of one system should be
uncorrelated with that of the other. Thus if one views recall and recognition as
involving separate memory systems, then the probability of recalling an item
that was recognized should be identical to the probability of recalling an item
that was not recognized. Statistically, the probability that an item is both rec-
ognized and recalled should be equal to the product of the overall probabilities
of recall and recognition. There is no necessary logical connection between sto-
chastic and functional independence; Tulving (1985) described how it is per-
fectly possible to find one with no evidence of the other.
Although stochastic independence may be sufficient to infer separate forms of
memory, it is probably misleading to consider it necessary. First, consider the
problem of measurement. Imagine two memory systems that follow completely
different rules, respond to every manipulation in opposite ways, and are local-
ized in completely different parts of the nervous system. The output of memory
system 1 is the performance of behavior 1 and the output of system 2 is behavior
2. If a given task taps both systems, then we will find stochastic dependence in
our memory task to the extent that behaviors 1 and 2 compete at the output
level. Although the 2 measures are dependent, it may be unreasonable to claim
that the memories are. Second, is it reasonable to demand, as did Roediger
(1984), that two memory systems share no common path or process in order to
describe them as "separate systems"? One could just as easily claim that one
system is separate from another if at least one path or process is unique to one
of them. The more practical truth must lie somewhere between these extremes.
Again, the problem reduces to disagreement about what is meant by a mem-
ory system. One resort is to turn the tables on the psychologist and define them
in terms of neural processes or localization. Even psychologists often take this
approach (e.g., Sherry & Schacter, 1987). But this cannot be the only answer. To
102 FORMS OF MEMORY

the psychologist who is interested in how memory works, the knowledge of neu-
ral localization is not particularly valuable, unless systems in different loci fol-
low different rules and systems with similar loci follow similar rules. That would
be convenient, but it is not logically necessary. Moreover, at our current level of
localizing memory function, there is no reason why several different processes
cannot occur at the same locus. We are comfortable distinguishing unit plasticity
that is the consequence of intrinsic presynaptic or postsynaptic changes, or
changes in a modulator, which then act on that unit presynaptically or postsyn-
aptically. It may not be absurd to think that these various types of plasticity at
more or less the same locus might underlie different memory functions (e.g.,
Hawkins & Kandel, 1984).
Fortunately, there may be a way out. Evolution tends to be conservative. Once
a "memory system" emerges as a solution to a problem of adaptation, it is likely
to come to serve other functions and solve other problems, too—even if that
system might not be the best imaginable solution to any of the other problems.
Memory may reflect a sort of adaptive compromise among the needs of a variety
of behavioral systems that require plasticity with experience. Only when that
compromise is incompatible with a major demand of some behavioral system
is a new memory system likely to evolve. Sherry and Schacter (1987) have
described many popular memory system distinctions in terms of their functional
incompatibility, suggesting that separate learning systems are most profitably
sought by considering the purposes served by those systems. Perhaps we should
follow evolution's lead and be conservative in postulating separate memory
systems.

REFERENCES

Allport, G. W. (1937). Personality: A psychological interpretation. New York: Holt.

Cohen, D. H. (1985). Some organizational principles of a vertebrate conditioning pathway: Is
memory a distributed property? In N. M. Weinberger, J. L. McGaugh, & G. Lynch,
(Eds.), Memory systems of the brain (pp. 27-48). New York: Guilford.
DeToldeo, L., & Blach, A. H. (1966). Heart rate: Changes during conditioned suppression in
rats. Science, 152, 1404-1406.
Dickinson, A. (1980). Contemporary animal learning theory. Cambridge: Cambridge Univer-
sity Press.
Eikelboom, R., & Stewart, J. (1982). Conditioning of drug-induced physiological responses.
Psychological Review, 89, 507-528.
Garcia, J., & Koelling, R. A. (1966). Relation of cue to consequence in avoidance learning.
Psychonomic Science, 4, 123-124.
Garcia, J., Kovner, R., & Green, K. S. (1970). Cue properties versus palatability of flavors in
avoidance learning. Psychonomic Science, 20, 313-314.
Garcia, J., Lasiter, P. S., Bermudez-Rattoni, F., & Deems, D. A. (1985). A general theory of
aversion learning. Annals of the New York Academy of Sciences, 443, 8-21.
Grill, H. J., & Norgren, R. (1978). The taste reactivity test: I. Mimetic responses to gustatory
stimuli in neurologically normal rats. Brain Research, 143, 263-279.
Hawkins, R. D., & Kandel, E. R. (1984). Is there a cell-biologist alphabet for simple forms of
learning? Psychological Review, 91, 375-391.
 Forms of Memory in Pavlovian Conditioning \ 03

Held, R., & Hein, A. (1963). Movement-produced stimulation in the development of visually-
guided behavior. Journal of Comparative and Physiological Psychology, 56, 872-
876.
Holland, P. C. (1977). Conditioned stimulus as a determinant of the form of the Pavlovian
conditioned response. Journal of Experimental Psychology: Animal Behavior Pro-
cesses, 3, 77-104.
Holland, P. C. (198la). Acquisition of representation-mediated conditioned food aversions.
Learning and Motivation, 12, 1-18.
Holland, P. C. (1981b). The effects of satiation after first- and second-order appetitive condi-
tioning. Pavlovian Journal of Biological Science, 16, 18-24.
Holland, P. C. (1983a). Occasion-setting in Pavlovian feature positive discriminations. In M.
L. Commons, R. J. Herrnstein, & A. R. Wagner (Eds.), Quantitative analyses of
behavior: Discrimination processes (Vol. 4, pp. 183-206). New York: Ballinger.
Holland, P. C. (1983b). Representation-mediated overshadowing and potentiation of condi-
tioned aversions. Journal of Experimental Psychology: Animal Behavior Processes, 9,
1-13.
Holland, P. C. (1984). The origins of Pavlovian conditioned behavior. In G. Bower (Ed.), The
psychology of learning and motivation (Vol. 18, pp. 129-173). Englewood Cliffs, N.J.:
Prentice-Hall.
Holland, P. C. (1985). Element pretraining influences the content of appetitive serial compound
conditioning in rats. Journal of Experimental Psychology: Animal Behavior Pro-
cesses, 11, 367-387.
Holland, P. C. (1986a). Temporal determinants of occasion setting in feature positive discrim-
inations. Animal Learning and Behavior, 14, 111-120.
Holland, P. C. (1986b). Transfer after serial feature positive discrimination training. Learning
and Motivation, 17, 243-268.
Holland, P. C. (1989a). Acquisition and transfer of conditional discrimination performance.
Journal of Experimental Psychology: Animal Behavior Processes, 15, 154-165.
Holland, P. C. (1989b). Event representation in Pavlovian conditioning: Image and action.
Manuscript submitted for publication.
Holland, P. C., & Block, H. (1983). Evidence for a unique cue in positive patterning. Bulletin
of the Psychonomic Society, 21, 297-300.
Holland, P. C., & Forbes, D. T. (1982a). Control of conditional discrimination performance by
CS-evoked event representations. Animal Learning and Behavior, 10, 249-256.
Holland, P. C., & Forbes, D. T. (1982b). Representation-mediated extinction of flavor aver-
sions. Learning and Motivation, 13, 454-471.
Holland, P. C., & Straub, J. J. (1979). Differential effects of two ways of devaluing the uncon-
ditioned stimulus after Pavlovian appetitive conditioning. Journal of Experimental
Psychology: Animal Behavior Processes, 5, 65-78.
Kamin, L. J. (1968). Attention-like processes in classical conditioning. In M. R. Jones (Ed.),
Miami symposium on the prediction of behavior: Aversive stimulation (pp. 9-32).
Coral Gables, Fla.: University of Miami Press.
Kierylowicz, H., Soltysik, S., & Divac, I. (1968). Conditioned reflexes reinforced by direct and
indirect food presentation. Ada Biologica Experimentalis, 28, 1-10.
Kimble, G. A., & Perlmuter, L. C. (1970). The problem of volition. Psychological Review, 77,
361-384.
Lamarre, J., & Holland, P. C. (1987). Acquisition and transfer of serial feature negative dis-
criminations. Learning and Motivation. 18, 319-342.
Lamb, M. R., & Riley, D. A. (1981). Effects of element arrangement on the processing of com-
pound stimuli in pigeons (Columba livia). Journal of Experimental Psychology: Ani-
mal Behavior Processes, 7, 45-58.
Mackintosh, N. J. (1975). A theory of attention: Variations in the associability of stimuli with
reinforcement. Psychological Review, 82, 276-298.
104 FORMS OF MEMORY

Pearce, J. M., & Hall, G. (1980). A model for Pavlovian learning: Variations in the effectiveness
of conditioned but not of unconditioned stimuli. Psychological Review, 106, 532-
552.
Pelchat, M., Grill, H. J., Rozin, P., & Jacobs, J. (1983). Quality of acquired responses to tastes
by rattus norvegicus depends on type of associated discomfort. Comparative Psy-
chology and Behavior, 97, 140-153.
Razran, G. H. S. (1971). Mind in evolution. Boston: Houghton Mifflin.
Rescorla, R. A. (1972). "Configural" conditioning in discrete-trial bar pressing. Journal oj'Com-
parative and Physiological Psychology, 79, 307-317.
Rescorla, R. A. (1974). A model of Pavlovian conditioning. In V. S. Rusinov (Ed.), Mecha-
nisms of formation and inhibition of conditional reflex (pp. 25-39). Moscow: Acad-
emy of Sciences of the USSR.
Rescorla, R. A. (1985a). Associative learning: Some consequences of contiguity. In N. M. Wein-
berger, J. L. McGaugh, & G. Lynch (Eds.), Memory systems of the brain (pp. 211-
230). New York: Guilford.
Rescorla, R. A. (1985b). Inhibition and facilitation. In R. R. Miller & N. E. Spear (Eds.), Infor-
mation processing in animals: Conditioned inhibition (pp. 299-326). Hillsdale, N.J.:
Erlbaum.
Rescorla, R. A. (1986a). Extinction of facilitation. Journal of Experimental Psychology: Animal
Behavior Processes, 12, 16-24.
Rescorla, R. A. (1986b). Facilitation and excitation. Journal of Experimental Psychology: Ani-
mal Behavior Processes, 12, 325-332.
Rescorla, R. A., & Holland, P. C. (1977). Associations in Pavlovian conditioned inhibition.
Learning and Motivation, 8, 429-447.
Rescorla, R. A., & Solomon, R. L. (1967). Two process learning theory: Relationships between
classical conditioning and instrumental learning. Psychological Review, 74, 151-182.
Rescorla, R. A., & Wagner, A. R. (1972). A theory of Pavlovian conditioning: Variations in the
effectiveness of reinforcement and nonreinforcement. In A. H. Black & W. F. Prokasy
(Eds.), Classical conditioning (Vol. 2, pp. 64-99). New York: Appleton-Century-
Crofts.
Riley, A. L., Jacobs, W. J., & LoLordo, V. M. (1976). Drug exposure and the acquisition and
retention of a taste aversion. Journal of Comparative and Physiological Psychology,
90, 799-807.
Roediger, H. L. III (1984). Does current evidence from dissociation experiments favor the epi-
sodic/semantic distinction? Behavioral and Brain Sciences, 7, 252-254.
Ross, R. T. (1983). Relationships between the determinants of performance in serial feature
positive discriminations. Journal of Experimental Psychology: Animal Behavior Pro-
cesses, 9, 349-373.
Ross, R. T., & Holland, P. C. (1981). Conditioning of simultaneous and serial feature-positive
discriminations. Animal Learning and Behavior, 9, 293-303.
Ross, R. T., & LoLordo, V. M. (1986). Blocking during serial feature-positive discriminations:
Associative versus occasion setting functions. Journal of Experimental Psychology:
Animal Behavior Processes, 12, 315-324.
Ross, R. T., & LoLordo, V. M. (1987). Evaluation of the relation between Pavlovian occasion
setting and instrumental discriminative stimuli: A blocking analysis. Journal of
Experimental Psychology: Animal Behavior Processes, 13, 3-16.
Ross, R. T., Orr, W. B., Holland, P. C., & Berger, T. W. (1984). Hippocampectomy disrupts
acquisition and retention of learned conditional responding. Behavioral Neurosci-
ence, 98, 211-225.
Sherry, D. F., & Schacter, D. L. (1987). The evolution of multiple memory systems. Psycholog-
ical Review, 94, 439-454.
Siegel, S. (1977). Morphine tolerance acquisition as an associative process. Journal of Experi-
mental Psychology: Animal Behavior Processes, 3, 1-13.
 Forms of Memory in Pavlovian Conditioning 105

Teyler, T. J., & DiScenna, P. (1985). The role of hippocampus in memory: A hypothesis. Neu-
rosdence and Biobehavioral Reviews, 9, 377-389.
Thompson, R. F. (1986). The neurobiology of learning and memory. Science, 233, 941-947.
Timberlake, W. (1983). The functional organization of appetitive behavior: Behavior systems
and learning. In M. D. Zeiler & P. Harzem (Eds.), Advances in the analysis of behav-
ior: Vol. 3. Biological factors in learning (pp. 177-221). Chichester: Wiley.
Tolman, E. C. (1932). Purposive behavior in animals and men. New York: Appleton-Century.
Tulving, E. (1985). How many memory systems are there? American Psychologist, 40, 385-398.
Wagner, A. R. (1981). SOP: A model of automatic memory processing in animal behavior. In
N. E. Spear & R. R. Miller (Eds.), Information processing in animals: Memory mech-
anisms (pp. 5-47). Hillsdale, N.J.: Erlbaum.
Zellner, D. A., Berridge, K. C, Grill, H. J., & Ternes, J. W. (1985). Rats learn to like the taste
of morphine. Behavioral Neuroscience, 99, 290-300.
Zener, K. (1937). The significance of behavior accompanying conditioned salivary secretion for
theories of the conditioned response. American Journal of Psychology, 50, 384-403.
5
Functional Forms of Human Memory
MARCIA K. JOHNSON

Several recent theories of memory can be classed as "functional subsystems"

approaches. They share the ideas that human memory is composed of distin-
guishable subsystems that serve different behavioral and cognitive functions,
operate according to different laws, and are represented by different neural struc-
tures or mechanisms. These subsystems may develop at different rates from
infancy to adulthood, and they may be differentially susceptible to disruption
from alcohol and other drugs, aging, disease, or injury (e.g., Cohen, 1984; John-
son, 1983; Squire, 1982; Tulving, 1983; Warrington & Weiskrantz, 1982).
Although the field is still struggling to characterize these subsystems, progress
does not depend on consensus; many schemes have been proposed that have
generated valuable ideas and research. I haven't space here to discuss various
viewpoints in detail, but one way they differ is in the aspects of the meaning of
"function" they seem to emphasize by the way subsystems are labeled. Terms
such as episodic versus semantic memory (Cermak, 1984; Schacter & Tulving,
1982; Tulving, 1983), procedural versus declarative memory (Cohen & Squire,
1980), and memories versus habits (Mishkin, Malamut, & Bachevalier, 1984)
emphasize the purpose/result/consequence meaning of function. Terms such as
horizontal versus vertical processes (Wickelgren, 1979), mediated versus stim-
ulus-response learning (Warrington & Weiskrantz, 1982), and sensory versus
perceptual versus reflective subsystems (Johnson, 1983) emphasize the activity/
process meaning of function. Of course, most investigators who have tackled the
problem of defining subsystems have considered both aspects of function—pur-
pose and process—to some degree, but the terms we use reflect certain tacit
assumptions, I think.
In any event, one major goal of a psychological analysis of memory is to
account for memory in terms of purposes such as skill learning, autobiographical
recall, generalization of concepts, and learned emotional reactions. It seems
unlikely, however, that the processes that serve these goals are uniquely dedi-
cated to particular purposes of this sort. For example, some of the same types
of memory processes might, under some circumstances, contribute to skill learn-
ing, autobiographical recall, generalization of concepts, and learned emotional

106
 Functional Forms of Human Memory 107

reactions. In fact, as processes were added (and/or modified) during evolution,

purposes probably multiplied. Thus the framework I describe in this chapter
assumes as a point of departure that the "functional forms" of human memory
(i.e., subsystems) yet to be specified are sets of processes that contribute in dif-
ferent combinations to the many purposes memory serves.

A MULTIPLE-ENTRY MODULAR MEMORY SYSTEM

The framework that I have found useful for thinking about memory is called a
multiple-entry modular memory system (MEM) (Johnson, 1983). I propose that
memory is composed of at least three interacting but distinguishable subsys-
tems: the sensory system, the perceptual system, and the reflection system (Fig.
5.1). Each subsystem is really a set of processes; thinking in terms of subsystems
allows us to group processes that have something in common and to highlight
apparent differences in processes that may imply interesting directions for
research.
The sensory and perceptual subsystems record the consequences of perceptual
processing. They differ in the type of perceptual information to which they are
most sensitive.1 The sensory system records the type of information that typi-
cally is not in itself the major object of perception but that operates as perceptual

FIGURE 5.1. A particular event creates entries in three subsystems of memory: sensory, per-
ceptual, and reflection. The x's indicate activation, the heavier the x's, the greater the likelihood
that the activation will recruit attention. Various memory tasks are listed near the subsystem(s)
that they are particularly likely to draw upon. (From Johnson, 1983, by permission)
108 FORMS OF MEMORY

cues. The sensory system develops associations involving such aspects of per-
ception as brightness, localization, and direction of movement. The sensory rec-
ord is important in many skills, such as developing hand-eye coordination,
learning to adjust one's posture to changes in external cues, improving perfor-
mance in tracking tasks, and other largely stimulus-driven tasks. The perceptual
subsystem records phenomenally experienced perceptual events, that is, external
objects in relation to one another. The reflection system records the active think-
ing, judging, and comparing that we do. It records our attempts to organize and
control what happens to us, and our commentary on events. Thus it records
memories that are internally generated.
In Figure 5.1 the subsystems are depicted as overlapping. If you look at the
figure for a few moments, it will begin to reverse like a Necker cube; each sub-
system may appear in the front, middle, or back. This visually represents the
interactive quality of the entire system. No subsystem is primary, and the order
in which processes are engaged is not fixed. As Kolers (1975) suggested, entries
in memory are the records of the specific processes that created them. Activation
(indicated by X's in Fig. 5.1) indicates that processing is taking place. The heav-
ier X 's indicate that we are aware of only a subset of activated information at
any particular time. Nevertheless, activation, not awareness, is the necessary
and sufficient condition for changing memory.
Most experiences will generate entries in all subsystems. For instance, in a
complex activity like playing tennis, various components are probably largely
mediated by different subsystems. Learning to get to the right place at the right
time involves sensory information that itself is not the object of perception,
especially visual information about trajectories such as the rate of change in the
size of the ball as a function of time. Learning to respond to configurational
aspects in the stimulus array is a perceptual function, for example, learning to
see relations among an opponent's position on the court, posture, and racket
orientation that signal what shot she is likely to make. Reflection is critically
involved in learning to understand an opponent's strategy, or to plan or initiate
one of your own. The memory for playing tennis is not a single type of repre-
sentation in a unitary memory system, but is multiply represented in various
subsystems of memory. Furthermore, most experiences have this same complex
character and are thus multiply represented.
I do not think of MEM as finished, but as a general framework for proceeding.
It may be necessary to add subsystems, and those already proposed need to be
specified more precisely (for example, several ideas for clarifying some of what
is involved in reflection are suggested in the section on recall below). One pos-
sibility is that subsystems should be defined in terms of relations that certain
neural networks are prepared (through either evolution or past experience) to
handle. For example, certain stimulus information may more easily be associ-
ated with body adjustments than with a complex plan. On the other hand, cen-
trally generated representations or plans may more easily serve as cues for the
revival of other ideas than as cues for certain body adjustments. If so, in MEM,
 Functional Forms of Human Memory 109

the question of the interaction of subsystems might be framed in terms of

whether and how sensory information contributes to the activation of reflections
and whether and how self-generated information such as plans affect actions
(such as eye movements) that are predisposed to respond to external cues.
The MEM framework helps organize a number of empirical findings from
normal human subjects. For example, memory can be measured in various
ways: with recall tests, recognition tests, or tests that show the effect of experi-
ence in indirect ways (or "implicitly"; Schacter, 1987b), such as improved ability
to identify a stimulus under degraded perceptual conditions. In normal human
subjects, these various measures are often not correlated or necessarily affected
in the same way by experimental manipulations (e.g., Jacoby & Dallas, 1981;
Tulving, 1983; see Schacter, 1987b for a recent review and Spear, 1984 for a
discussion of related issues). This lack of correlation among measures can be
understood if we assume that recall draws heavily on the reflection system, rec-
ognition draws on the perceptual system, and perceptual-threshold tasks like
perceptual identification depend mainly on the sensory system. In Figure 5.1,
tasks are listed near the subsystems that are most implicated, but there is no
"pure" one-to-one correspondence between tasks and subsystems.
This model can also be applied to amnesia.2 Quite a bit of evidence is consis-
tent with the idea that anterograde amnesia results from a deficit primarily in
the reflection subsystem, with the sensory and perceptual memory subsystems
relatively intact (for reviews see Parkin, 1982; Cermak, 1982; Hirst, 1982). For
example, amnesics show dramatic deficits in recall tasks. At the same time, they
show relatively normal performance in certain perceptually based tasks such as
learning mirror drawing (Milner, 1966), learning to read mirror text (Cohen &
Squire, 1980), and perceptual identification of previously exposed words (Cer-
mak, Talbot, Chandler, & Wolbarst, 1985). In the MEM framework, the more
an activity requires reflection—that is, the more it requires self-generated cues—
the more difficult it should be for amnesics to master. The more perceptual sup-
port for an activity there is (i.e., the more it is externally guided), the easier it
should be for amnesics to learn.
Current research in areas such as skill learning (e.g., Nissen & Bullemer, 1987)
and priming (e.g., Graf, Squire, & Mandler, 1984; Schacter, 1987b) is clarifying
our picture of some of the memory functions that appear to be preserved in
anterograde amnesia (e.g., development of sensory-motor associations and per-
ceptual analysis processes, activation of semantic concepts). Here I focus on
amnesic performance in tasks that are relevant to understanding memory func-
tions that are not preserved. This chapter applies the MEM framework to the
analysis of amnesia in three areas of research: recognition memory, acquisition
of affective responses, and free recall. These areas are particularly interesting
because amnesia's most profound consequence is loss of memory for recent per-
sonal experience and often what makes a memory seem personal is that it is
recognized as familiar from a particular context or source, evokes affective
responses, and can be revived or reconstructed voluntarily. Thus understanding
110 FORMS OF MEMORY

processes involved in recognition, memory for affect, and recall eventually

should help us better understand the loss of personal memory that is so char-
acteristic of amnesia.

RECOGNITION

One specific prediction generated from MEM is that amnesics should show less
disruption in recognition than in recall, because unlike recall, recognition pre-
sumably does not always require reflection but can be based on perceptual rec-
ords alone (e.g., Jacoby, 1982; Mandler, 1980). Direct support for this idea
comes from several experiments. In one study (Hirst et al., 1986) we tested two
groups of amnesics, patients with Korsakoff's syndrome (a memory disorder
associated with thiamine deficiency and chronic alcoholism) and a group of non-
alcoholic amnesics of mixed etiology (e.g., hypoxic ischemia, anterior commu-
nicating artery aneurysm). Each amnesic group was compared to an appropriate
age- and education-matched control group. The subjects studied word lists and
we equated the performance of amnesics and controls on forced-choice recog-
nition by giving the amnesics more time to study the words—amnesics took 8
sec per item of study time to reach the same level of recognition performance as
normals given only 0.5 sec per item. Then we looked at recall for the same lists.
Because the same patterns of results were obtained for the two amnesic groups,
they have been combined in Table 5.1 A. As you can see, even though their rec-
ognition performance was equal to that of controls, amnesics still showed a
marked decrement in their ability to recall the words.
We extended these results in subsequent experiments with the mixed-etiology
amnesics. In one experiment (Hirst, Johnson, Phelps, & Volpe, 1988, experi-
ment 1) amnesics and normals both had 8 sec to study each item, but we equated
amnesic and normal recognition by lengthening the retention interval for the
normal subjects. That is, the recognition performance of normals tested after a
1-day delay was about the same as that of amnesics tested immediately. Again,

TABLE 5.1 Mean Proportion Correct Recognition and Recall for

Three Experiments
Group Time n Recognition Recall
A. Amnesics (8 sec) 13 .78 .07
Controls (0.5 sec) 13 .77 .25
B. Amnesics (30-sec delay) 6 .85 .06
Controls (1-day delay) 6 .86 .22
C. Amnesics ( 2 X 5 sec) 6 .85 .07
Controls (1 X 0.5 sec) 6 .77 .16
Source: Data from (A) Hirst, Johnson, Kim, Phelps, Risse, and Volpe (1986; blocked,
categorized lists) and (B) and (C) Hirst, Johnson, Phelps, and Volpe (1988, experi-
ments 1 and 2; unrelated words).
 Functional Forms of Human Memory 111

with recognition equated, amnesics were at a substantial disadvantage in recall

(Table 5.IB). In another experiment (Hirst et al., 1988, experiment 2) the list
was presented to the amnesics twice at a 5 sec rate and only once for 0.5 sec to
normals. Under these conditions the amnesics actually scored significantly bet-
ter on recognition than the normals. But even with this recognition advantage,
amnesics scored much poorer on the recall test (Table 5.1C).
Although these experiments show that amnesic recognition is less disrupted
than recall, the amnesics still had a substantial recognition deficit that could be
overcome only with long or repeated exposures. One explanation is that word
stimuli are relatively "impoverished" perceptually. In addition, words are
extremely familiar and probably do not create discriminable memory records
without some reflection (e.g., Eysenck, 1979; Jacoby & Craik, 1979). It is possi-
ble that if we could make the perceptual qualities of the stimulus more impor-
tant, the amnesics would more closely approximate normal performance. This
in fact sometimes seems to be the case when the stimuli are unfamiliar pictures
rather than words.
In one experiment we showed Korsakoffs and controls pictures of abstract
"paintings" that we had made up (Johnson & Kim, 1985). Subjects saw each
item 1, 5, or 10 times; then they were given a forced-choice recognition test in
which they had to discriminate the more familiar picture from a new distractor
and rate their confidence in their choice. Korsakoif patients were mildly, but not
significantly, impaired in recognition memory (for both forced choice and choice
weighted by confidence) (Table 5.2A). Furthermore, we retested subjects 20 days
later, and even after 20 days, Korsakoff patients' recognition performance for

TABLE 5.2 Recognition Scores for "Paintings" after 5-

Minute and 20-Day Retention Intervals
Number of Exposures
Test 1 5 10
A. 5-minute delay
Mean proportion correct
Korsakoff (n = 9) .67 .81 1.00
Control (n = 9) .78 .94 1.00
Choice weighted by confidence"
Korsakoff 4.03 4.72 5.33
Control 4.31 5.31 5.64
B. 20-day delay
Mean proportion correct
Korsakoff .56 .54 .78
Control .53 .71 .86
Choice weighted by confidence
Korsakoff 3.65 3.82 4.46
Control 3.60 4.19 4.72
"Correct choice: 6 = very sure; 5 = quite sure; 4 = guessing. Incorrect
choice: 3 = guessing; 2 = quite sure; 1 = very sure.
Source: From Johnson and Kim (1985).
112 FORMS OF MEMORY

pictures seen 10 times was a remarkably good 78%, compared with 86% for con-
trols (Table 5.2B).
To put these findings in perspective, we compared them with previously
reported studies of picture recognition in Korsakoff patients. Figure 5.2 shows
the relative performance of Korsakoff and control subjects who were tested
under comparable conditions, where controls were below maximum perfor-
mance (not at ceiling). It includes 11 data points from six different experiments
including ours using abstract paintings (Johnson & Kim, 1985) and other exper-
iments using faces, miscellaneous magazine pictures, magazine covers, and pat-
terns (Biber, Butters, Rosen, Gertsman, & Mattis, 1981; Cutting, 1981; Huppert
& Piercy, 1976, 1977; Talland, 1965). If KorsakofFs and normals performed
exactly alike, the points would all be on a 45° line. In some of the conditions,
the amnesics showed a significant deficit; in others the amnesics did not differ
significantly from normals. More important, the correlation between the perfor-
mance of Korsakoff patients and that of normals is quite clear (r = .80). Items
or conditions that were difficult for controls were difficult for Korsakoffs; items
or conditions that were easy for controls were easy for Korsakoffs (also see
Mayes, Meudell, & Neary, 1980). This general pattern supports the idea that
picture recognition involves some memory processes that are relatively intact in
amnesia.
A dissertation by Weinstein (1987) helps clarify the conditions under which
we might expect to find comparable recognition for amnesics and controls, and
those conditions under which amnesics might show a deficit. All subjects saw a
series of pictures that were line drawings of familiar objects, each colored in a
single color. The type of orienting task was varied. Half of each group were given
a perceptual task in which they were directed to keep track of the number of
black objects. This task is perceptual because it requires subjects to attend to
only a physical feature of each object, its color. The other half of the subjects
were given a reflective task; they were to decide whether each object was pre-
sented in a common color or a novel color. For example, a yellow lemon would

FIGURE 5.2. Relative performance

of Korsakoff patients and control sub-
jects on picture recognition. Data are
from the following studies: Biber et
al., 1981 (A); Cutting, 1981 (A); Hup-
pert and Piercy, 1976 (O); Huppert and
Piercy, 1977 (•); Talland, 1965 O;
Johnson and Kim, 1985 (•). (Re-
printed from Johnson and Kim, 1985,
by permission)
 Functional Forms of Human Memory 113

TABLE 5.3 Mean Proportion Correct for Yes/No Picture Recognition

Perceptual Reflective Orienting
Orienting Task Task
Distraction Amnesic Control Amnesic Control
Condition (n = 12) (n = 12) (n = 12) (n = 12)

A. New-novel .69 .76 .76 .90

B. New-common .51 .79 .80 .93
C. Old-common .44 .68 .58 .87
Source: From Weinstein (1987).

be common; a purple camel would be novel. This task is reflective in that it

requires subjects to consider the ongoing perceptual product and information
activated about the object from general knowledge and then to compare the
information from these two sources and make a decision. After all the pictures
were presented, subjects were given a Yes/No recognition task, consisting of
novel-colored objects from the acquisition list mixed with new novel colored
objects.
The results are shown in Table 5.3A, which combines Korsakoffs and nonal-
coholic amnesics, who showed the same patterns. After the perceptual task, rec-
ognition was not significantly different for amnesics and controls. After the
reflective task, the controls were significantly better than the amnesics. These
results suggest that if controls and amnesics process information perceptually,
their recognition performance will be similar. If they process information reflec-
tively, either controls engage in more embellished reflection or they are better
able to reinstate memory for this reflection at the time of the test, or (most likely)
both.
Normals and amnesics appeared to be equally good at using perceptual rec-
ords for discriminating familiar old pictures from unfamiliar new pictures. Per-
haps this was because no conflicting familiarity response was evoked by the dis-
tractor items. If the test involved some ambiguity about the source of
familiarity, the importance of reflective processes should show up (e.g., Johnson,
1988; Johnson & Raye, 1981; Lindsay & Johnson, 1987). Thus Weinstein (1987)
tested these same subjects on two other picture series in which the recognition
decisions should have required reflection. In these conditions the acquisition
series were of the same type as before (involving, of course, new pictures). But
the distractors in the second condition were new objects in common colors and
the distractors in the third condition were old objects from the acquisition list
that had previously appeared in novel colors but now appeared in common col-
ors. In both these cases we expected the distractors to evoke some sort of famil-
iarity response (based on either semantic knowledge or semantic knowledge plus
recent experience). Discriminating the targets from the distractors under these
conditions should be more difficult because familiarity alone should not be a
sufficient cue. Thus the benefit from reflective activity should be greater.
The results for these more difficult distractor conditions are shown in Table
114 FORMS OF MEMORY

5.3B and C). When the test involved familiar distractors, the amnesics per-
formed considerably worse than the controls even in the perceptual orienting
task condition. These results are consistent with the idea that reflection helps
specify the source of a familiarity response and that amnesics perform more
poorly with increasing reflective demands.
To summarize these experiments on word and picture recognition, the fact
that the amnesics' recall deficit is greater than we would expect from their rec-
ognition performance indicates that regardless of what processes are disrupted
in amnesia, they are more important in recall tasks than in recognition tasks.3
We think the disrupted processes are controlled, self-generated mental activities
of the sort I have been calling reflection. The fact that amnesic recognition can
actually equal normal recognition when the recognition task is largely percep-
tual, and that amnesics begin to show marked deficits as recognition requires
more reflection, makes the point even more strongly. A sense of familiarity does
not necessarily depend on reflection, but further specifying the source of the
familiarity does (Huppert & Piercy, 1978; Schacter, Harbluk, & McLachlan,
1984).

AFFECT

There has not been much work on amnesics' acquisition of affective reactions,
but scattered reports suggest that amnesics do acquire affective responses (e.g.,
Claparede, 1911, cited in Baddeley, 1982), although some suggest that they do
not (Redington, Volpe, & Gazzaniga, 1984). As with recognition, a closer anal-
ysis of the problem suggests that we might be able to characterize the conditions
under which amnesics will and will not acquire affective responses.
According to the MEM model, emotion may originate with experiences that
are perceptual or with experiences that are more self-generated or reflective. For
example, after a traffic accident, the squeal of brakes may be associated in the
perceptual system with fear. Frustration or anger may be another part of the
affective response associated in the reflection system with thoughts after the acci-
dent about its consequences (e.g., missed appointments, lost work time, the
inconvenience of getting the car repaired). Later, hearing the squeal of brakes
may directly revive some components of the total affective response (e.g., the
fear); the revival of other components (e.g., the frustration) will be more depen-
dent on recalling earlier reflective activity.
If this characterization of emotion is correct, amnesics should still be able to
develop affective responses—emotions, preferences, etc.—in situations that do
not depend on reflection but in which the affect is tied to perceptual features of
a situation. In one experiment exploring this idea (Johnson, Kim, & Risse, 1985,
experiment 1), KorsakofFand control subjects heard unfamiliar Korean melo-
dies. Then these melodies were mixed with a number of new melodies and the
subjects were asked to rate how much they liked each one. We chose this situa-
tion because we reasoned that the affective response to melodies is largely based
 Functional Forms of Human Memory 115

on perceptual characteristics of the melody; if so, the preferences of amnesics

should be affected by the same variables as the preferences of normals. Research
by Zajonc (1980) and others has shown that normal subjects often prefer stimuli
they have previously experienced, and this is the result we obtained. The control
subjects liked the old melodies better than the new melodies. The interesting
new finding was that the Korsakoff subjects also gave higher ratings to the pre-
viously heard melodies; furthermore, the magnitude of the effect was the same
as for normals (Table 5.4A). On the other hand, even though the amnesics pre-
ferred the old melodies to new ones, on a recognition test they showed the usual
deficit in ability to say what they had and had not heard before (Table 5.4B).
These results suggest that affective reactions develop normally in amnesics if
the situation is largely perceptual. In a second experiment (Johnson et al., 1985,
experiment 2) we used a situation that we thought would be much more likely
to involve reflection. The subjects were the same as in the melody experiment.
They were shown pictures of two young men, Bill and John, and were asked to
give their impression of each by rating him on several characteristics, such as
honesty, politeness, and intelligence. Then subjects heard some facts about Bill
and some facts about John. John was depicted as a "good guy" (he helped his
father, got a Navy commendation for saving someone's life, etc.), and Bill was
depicted as a "bad guy" (he stole things, broke someone's arm in a fight, etc.).
After an interval, the subjects were shown the pictures and were asked about
their impressions again.
The subjects heard the biographical information a total of three times over 3
or 4 days. We brought them back 20 days later and asked for their impressions
again. The results are shown in Figure 5.3; the higher the score, the more posi-
tively the man in the picture was rated on such traits as honesty, politeness, and
intelligence. Look first at the control subjects (squares and triangles). The pic-
tures were rated first before the subjects heard any biographical information,
and, as you would expect, Bill and John were rated about equally. After the con-
trol subjects heard the biographical information for the first time, the ratings
changed dramatically with the good guy rated more favorably and the bad guy
less favorably. Ratings did not change much with repetitions of the biographical
information, and the effect of the biographical information persisted over the
20-day retention interval (day 3 in Fig. 5.3). Now look at the Korsakoff patients

TABLE 5.4 Preference Ratings and Recognition Scores for Melodies

A. Preference Ratings"
Group Old New B. Proportion Correct Recognition

Amnesic 4.10 3.74 .59

(n = 9)
Control 3.77 3.46 .83
(n = 9)
"Higher ratings indicate more preferred.
Source: From Johnson, Kim, and Risse (1985, experiment 1).
FIGURE 5.3. Mean impression ratings for Korsakoffs (N = 9), nonalcoholic controls (N = 9),
and alcoholic controls (N = 6). (From Johnson, Kim, and Risse, 1985, experiment 2)
 Functional Forms of Human Memory 117

(circles). Again, the ratings for Bill and John started out the same and diverged
significantly. Even after a 20-day retention interval, the amnesics gave the good
guy higher ratings than the bad guy. After 20 days, the control subjects could
recall about 35% of the biographical information, whereas the Korsakoffs
recalled virtually nothing. Therefore, although the Korsakoffs could not recall
the biographical information, it still affected their judgments about which of the
two men was nicer.
In the Bill and John study, Korsakoff patients clearly developed less extreme
impressions than did controls. In contrast, these same Korsakoff patients
showed the same development of preference for melodies as controls did. Dif-
ferences in the results of the two studies are interpretable within the MEM
framework. Compared to developing preferences for melodies, when we develop
preferences for people there is much more room for reflection to operate. The
MEM framework assumes that some affective responses are tied to the percep-
tual features of the pictures of the two men, whereas other affective responses
are tied to the reflective activity the subjects engaged in while hearing the bio-
graphical information (e.g., comparing the men to other people they have
known, evaluating the severity of misdeeds). Later, reinstating perceptual cues
from the pictures should revive some affective components, but other aspects of
the total affective response should depend on reinstating previous reflection.
Normal subjects could cue themselves by recalling specific biographical details
and should therefore have a more embellished affective response.
In summary, our two experiments on affect indicate that Korsakoff patients
retain the capacity for developing affective reactions. But the degree to which
we can expect amnesics to retain affective responses in any particular case
depends on the relative involvement of different memory subsystems in sup-
porting affective responses of nonamnesic subjects.

RECALL

The most salient and central symptom of amnesia is profoundly poor free recall
(e.g., Butters & Cermak, 1980; Milner, Corkin, & Teuber, 1968). Precisely
because amnesic recall is so poor, it is difficult to study systematically. By con-
sidering the various mental activities required by recall, however, we should be
able to break the problem down into tractable component parts. This will also
help further clarify the concept of reflection.

Basic Reflective Subprocesses

Even in its simplest form, normal recall clearly depends on a number of basic
reflective subprocesses, the most important of which have to do with establish-
ing relations among elements to be recalled (e.g., Bower, 1970; Mandler, 1967;
Tulving, 1962). Consider a hypothetical organizational problem involving a free
recall list that includes the words pig, dog, weed, and dinner (Fig. 5.4A). Each
118 FORMS OF MEMORY

FIGURE 5.4. Hypothetical activation patterns and noted relations for four items (pig, dog, weed,
dinner) from a free recall list. (A) Initial activation and noted relations. (B) Activation and noted
relations after a shift in perspective.

word activates some set of relations based on prior experience and current con-
text. One primary reflective function is noting the relations that are given in this
current activation pattern. For example, a subject might note that pig and dog
are both animals, and that flower (which is activated by weed) can be used as a
centerpiece on a table (which is activated by dinner). This noting activity would
establish two small units.
Another critical reflective function is shifting activation to other information
in order to change what is given. The change is accomplished by changing per-
spective or shifting attention from the currently activated aspects of a stimulus
to other potentially useful aspects (see Fig. 5.4B). Suppose that for our hypo-
thetical subject, attention is shifted from the idea of animals to the fact that pigs
and dogs are found on farms. At the same time, suppose the subject thinks about
aspects of weeds and dinner that in turn suggest the idea of garden. Garden and
farm animals are both related through farm. Now the subject has a single inter-
related set of relations including the four items.
 Functional Forms of Human Memory 119

A third reflective process is refreshing. Clearly, various ideas and relations

must remain active during shifting and noting until a stable or cohesive set of
relations has been noted. Keeping them active requires some sort of continuous
scan of current activation in order to keep it refreshed.
A fourth reflective function is delayed reactivation of information that has
disappeared from consciousness. Through internally generated remindings, sets
of relations become more cohesive.
To simplify, I have illustrated these subprocesses applied to learning and
remembering a word list, but all these basic reflective activities are central in
processing more complex, naturally occurring events.

Layers of Reflection
Reflective processes differ not only in type, but also along a continuum of "plan-
fulness," which may affect the characteristics of these basic reflective activities.
At the minimum level of planfulness, basic reflective activities may consist
largely of allowing the activation consequences of successive events to settle and
noting whatever cohesive sets of relations emerge.4 Even such relatively passive
mental activity receives some direction from ongoing goals or agendas (e.g., the
goal to read words aloud). Agendas vary in the degree of deliberation they
require to execute and monitor. Thus some agendas call up well-learned percep-
tual and reflective schemata that may organize component processes relatively
automatically, while others bring basic subprocesses under strategic control in
order to organize them to meet new, unusual, or complex demands.
Figure 5.5 shows how we might label reflective processes differently, depend-
ing on the degree of deliberation or planfulness involved. The cube represents
basic reflective processes that occur spontaneously or under control of relatively
simple agendas on the bottom and reflective processes under control of more
strategic planning on the top. Look first at the lower front corner. Shifting is a
change in perspective as a consequence of overlapping spreading activation pat-
terns. Initiating is a change in perspective via strategically controlled activities
(such as listing all the properties you can think of for two objects that are to be
related). Refreshing is activation prolonged by simple attention; rehearsal is acti-
vation prolonged by systematic, strategically controlled recycling. Reactivating
refers to the revival of inactive information via spontaneous or accidental men-
tal events that provide a reminder; retrieval refers to reviving information
through conscious attempts to get back to it. Noting involves seeing relations
that are relatively direct, and discovering involves seeing relations that are less
direct. Thus there are at least two layers of internal control, or layers of reflection
involved in learning and remembering. The idea that these two layers corre-
spond to two functional subsystems of reflection, R-l and R-2, is developed else-
where (Johnson, in press; Johnson & Hirst, in press). (Furthermore, some men-
tal activities could be thought of as combinations of these component processes.
For example, the combination of initiating and discovering, represented along
the top edge of the cube, would produce "elaboration" [e.g., Stein & Bransford,
120 FORMS OF MEMORY

FIGURE 5.5. Reflective processes differ in degree of "planfulness." Basic reflective processes
are represented at the bottom corners of the cube; corresponding but more strategic, goal-directed
reflective functions are represented at the top corners of the cube.

1979]). To understand amnesia, we need to specify exactly which reflective sub-

processes are disrupted and which are not, and among those that can vary in
strategic control, which layer of control is disrupted.

Amnesia and Reflective Subprocesses

Amnesia could cause problems with all reflective subprocesses or with only
some of them. What is our best guess about amnesia's effect on these basic reflec-
tive subprocesses? Several lines of evidence suggest that noting given relations
is intact. First, amnesics understand ordinary conversation. Also, amnesics do
fairly well on easy pairs on the Wechsler Memory Scale (hot-cold). Despite early
evidence to the contrary (Cermak & Butters, 1972), it appears that amnesics
spontaneously categorize items from taxonomic categories (McDowell, 1979).
What we need is more detailed or analytic information about this noting func-
tion—for example, what kinds of relationships can be noted, and under what
conditions?
A recent study from our lab (Johnson, Hirst, Phelps, & Volpe, unpublished
 Functional Forms of Human Memory 121

data) illustrates one potential approach. Nonalcoholic amnesics and normal

controls were presented with some difficult-to-understand sentences (e.g., Birn-
baum, Johnson, Hartley, & Taylor, 1980; Johnson, Doll, Bransford, & Lapinski,
1974). Half the sentences were preceded by an appropriate context (e.g., Bag-
pipes: The notes went sour when the seams split) and half were preceded by an
inappropriate context (e.g., New car: The house turned to water when the fire got
too hot). As each sentence was presented, subjects were asked to rate how much
sense it made. The points over ACQ (for acquisition) in Figure 5.6 show the
mean ratings; higher ratings reflect greater comprehension. As you can see, both
amnesics and controls showed a large context effect.
Comprehension ratings were also taken later for the sentences alone (without
the contexts) at three retention intervals, 2 min, 11/2 hr, and 1 week; these ratings
are also shown in Figure 5.6. Notice that the effect of having heard a relevant
context is relatively long-lasting for amnesics. This is an example of new learn-

FIGURE 5.6. Mean comprehension ratings for sentences preceded by either an appropriate or
inappropriate context at acquisition (ACQ) and later (2 min, 1/2 hr, 1 week) presented without
contexts (amnesic N = 4; control N = 4). (Johnson, Hirst, Phelps, and Volpe, unpublished data)
FIGURE 5.7. Mean comprehension ratings for sentences preceded by an appropriate context at acquisition (ACQ) and later (2 min,
l1/2 hr, 1 week) presented without contexts. Ratings are shown separately for high- and low-related context-sentence pairs and for
controls (A) and amnesics (B). (Johnson, Hirst, Phelps, and Volpe, unpublished data)
 Functional Forms of Human Memory 123

ing that cannot simply be attributed to short-term priming (see McAndrews,

Glisky, & Schacter, 1987, for a similar rinding).
We also looked at comprehension ratings separately for sentences in which
the context and sentence were highly related and for sentences in which the con-
text-sentence relation was low. Figure 5.7A shows the comprehension ratings
for the control subjects; these verify the ratings of the subjects who gave us the
normative data for high and low. In contrast to the controls, the amnesics
showed a relatedness-by-delay interaction; the initial advantage of highly related
items disappeared over the retention interval (Figure 5.7B).
At the 2-min, 11/2-hr, and 1-week retention intervals (when sentences were pre-
sented without any context), we also asked subjects to tell us what they thought
each sentence was about. Table 5.5 shows the mean number of appropriate con-
texts given in this context-generation task, collapsed across retention interval.
There was a subject-by-relatedness interaction; the difference between amnesics
and controls was greater on low than on highly related items. The less obvious
the relation between the context and sentence was, the more difficulty amnesics
had in later describing what the sentence was about (see also Warrington &
Weiskrantz, 1982). The initial comprehension ratings taken as the context-sen-
tence pairs were first presented showed a similar (though not significant)
pattern.3
One reasonable hypothesis supported by the results of our comprehension
study is that amnesics are able to note relations that are a relatively direct con-
sequence of activation patterns set up by incoming stimuli. The more distant
the relation—the more seeing, or discovering, a connection depends on shifting
attention from initially activated but not useful relations to new possibilities—
the more problems the amnesic may have.
At the one-week retention interval, we also gave subjects a recognition test on
the sentences presented without contexts. Amnesics' recognition of these sen-
tences, although clearly worse than normals, was quite good after a week (Table
5.6). More important, amnesics recognized more of the low-related sentences
than the high-related sentences. This recognition advantage of the low-related
sentences is interesting because it indicates that amnesics based their recognition
judgments on information somewhat different from that used for their compre-

TABLE 5.5 Mean Number of Contexts Given for Sentences

Initially Presented with an Appropriate Context (averaged over 2-
min, 11/2-hr,and 2-week retention intervals)
Relation Between
Context and
Sentence
Group High Low
Amnesic (n = 4) 2.33 2.00
Control (n = 4) 2.83 3.25
Source: From Johnson, Hirst, Phelps, and Volpe (unpublished data).
124 FORMS OF MEMORY

TABLE 5.6 Mean Proportion C orrect Yes/No Recognition (1-week

delay)
Appropriate Inappropriate
Context Context
Group High Low High Low New Items

Amnesic .75 1.00 .69 .88 .72

(n = 4)
Control .94 1.00 1.00 1.00 1.00
(n = 4)
Source: From Johnson, Hi rst, Phelps, ar d Volpe (unpjblished data).

hension judgments and their attempts to describe what the sentences were
about.
One way contexts help comprehension is by activating a schema that is used
in interpreting further information and in retrieving it later (e.g., Bransford &
Johnson, 1973; Schank & Abelson, 1977). Thus our context study indicates that
amnesics can use schemas. Another piece of evidence that amnesics can apply
schematic knowledge to incoming information comes from an experiment con-
ducted by Phelps (Phelps, Hirst, Johnson, & Volpe, 1988). Mixed-etiology
amnesics and controls were asked to remember two stories. All subjects heard a
general story that involved a topic with which most people are familiar (a shop-
ping trip). In addition, each amnesic and his or her control heard a story tailored
to the individual interests of each amnesic. For example, one amnesic knew a
lot about basketball, so he (and his control, who was not particularly interested
in basketball) received a story about a basketball game. As Figure 5.8 shows, the
controls recalled a larger percentage of the general story than the subject-specific
stories. In contrast, the amnesics recalled a larger percentage of the subject-spe-
cific stories than the general story. Thus, although for normal subjects the sub-
ject-specific stories were more difficult, they were relatively easy for amnesics
because the amnesics were able to draw on their special interest in and knowl-
edge of these areas to aid recall.6 It would be valuable to have more specific
information about the types of schemas that amnesics can and cannot use.7
We also do not know much about the subprocess of refreshing activated infor-
mation, but verbal rehearsal seems to be relatively intact. Available evidence
from short-term memory experiments with amnesics suggests that when they are
not distracted, amnesics are relatively good at rehearsing currently activated
verbal information (but perhaps not nonverbal information; Milner, 1966).
Later recall, however, depends on more than rehearsal (e.g., Glenberg, Smith, &
Green, 1977). Noting and, if necessary, shifting must take place as well. Amne-
sics may have more trouble than control subjects simultaneously engaging in
these various subprocesses (which may require coordinating layers of reflective
processes). Thus not only should amnesics show greater disruption when unre-
lated distraction is introduced, as in typical STM experiments (Cermak, Butters,
 Functional Forms of Human Memory 125

FIGURE 5.8. Mean percentage of idea units recalled from a story about a shopping trip (general)
and stories tailored to the interest of amnesics (subject-specific) (amnesic N = 6; control N = 6).
(Data from Phelps, Hirst, Johnson, & Volpe, 1988, experiment 1)

& Goodglass, 1971; Kinsbourne & Wood, 1975), but they should also show
greater disruption when additional task-relevant demands such as noting and
initiating are introduced. For example, in one experiment (Warrington, 1982)
subjects were quickly read three words, distracted for 15 seconds, and then
tested for recall. Whereas amnesics performed as well as normals if the words
were unrelated, they showed a deficit on related-word trials (e.g., drink-coffee-
cold).
Clearly, amnesics are impaired on delayed strategic retrieval of information
that has dropped out of consciousness. But what should be emphasized is the
critical role that such delayed reactivation plays in normal memory (Johnson,
1987; Levin et al., 1985; Ribot, 1882/1977). Without subsequent rehearsals,
many autobiographical memories do not remain clear or vivid (Suengas & John-
son, 1988). Some investigators suggest that a consolidationlike process in which
126 FORMS OF MEMORY

memories gradually become more resistant to disruption takes place over years
(Squire, 1982; Wickelgren, 1979). Whether or not such long-lasting processes are
initiated when a memory is first established, it seems clear that the accessibility
of memories is related to their subsequent reactivation (also see Squire, 1986, p.
1616) and that the loss of the consequences of reactivation contributes to the
amnesic deficit.

TYPES AND DEGREES OF AMNESIA

Investigators have tried to isolate memory deficits from other cognitive deficits,
but this is easier said than done (see, for example, discussions by Baddeley, 1982;
Kinsbourne & Wood, 1982; Moscovitch, 1982; Warrington, 1982). In practice,
isolating memory amounts to either (1) defining amnesia as deficits on certain
simple word, picture, and prose recall and recognition tasks (along with clini-
cally diagnosed impaired memory for recent personal events), and considering
normal performance on these tasks but disrupted performance on more complex
memory tasks to be due to secondary cognitive deficits, or (2) defining amnesia
as whatever is disrupted by lesions confined to certain areas of the temporal
lobes and diencephalon and attributing memory deficits from lesions in other
areas (e.g., prefrontal cortex) to secondary cognitive deficits. Both alternatives
imply that the major objective is to understand the "core amnesia" thus defined.
From this "core amnesia" point of view, many reflective processes would be
considered secondary cognitive processes.
Focusing on "core amnesia," although one reasonable strategy, is based on a
perhaps too simple view of memory (also see Morton, 1985). Many cognitive
psychologists would argue against the idea that memory can be isolated from
other cognitive processes such as attention, comprehension, thinking, planning,
and problem solving.8 Memory is the result of many processes, from those
engaged by "simple" perception to those engaged by highly organized plans and
schemas. In fact, remembering itself can be thought of as a skill (e.g., Ericsson,
Chase, & Faloon, 1980) or a type of problem solving. Thus the idea of "core
amnesia" may be misleading. Rather, there may be many types and degrees of
amnesia, depending on which combination of memory subprocesses is required
for particular memory tasks, and on which memory processes are disrupted.9
But the fact that few differences among amnesics as a function of etiology have
been reported (e.g., Corkin et al, 1985) suggests that reflection is a highly inte-
grated system.
If we adopt a relatively complex view of memory, we should be more likely
to attempt to integrate the results from a wider range of patterns of memory
deficits, and this (in combination with ongoing efforts in cognition to analyze
normal memory functioning) could be quite informative. One goal might be to
refine and embellish a preliminary categorization scheme for reflective processes
such as the following:
 Functional Forms of Human Memory 127

Strategic Reflective Subprocesses

Intact Disrupted
Intact Normal Deficits in complex
Basic memory information and
Reflective sequences
Subprocesses
Disrupted "Core Severe amnesia
amnesia"

According to the scheme, even if certain strategic reflective processes were

intact, major disruption in one or more basic reflective processes would produce
profound deficits in even the easiest memory tasks ("core amnesia"). If strategic
reflective processes were disrupted but basic reflective processes were intact, the
patient might show considerable memory for some things but have difficulty
with others requiring more sophisticated reflective processing (e.g., complex
information, sequences, or temporal orderings). Disruption of both levels of
reflection—basic subprocesses and those under more strategic control—would
produce the most severe amnesia. It is likely that different types of reflective
subprocesses depend on different neurological structures and thus different
lesion sites should produce amnesias differing in type or severity. For example,
certain temporal and diencephalic lesions may disturb basic reflective subpro-
cesses and certain frontal lesions may affect strategic reflective planning subpro-
cesses (Goldman-Rakic, 1987; Milner & Petrides, 1984; Schacter, 1987a; Squire,
1987; Warrington & Weiskrantz, 1982). This categorization is undoubtedly too
simple, but it illustrates the approach of assuming memory depends on a rich
repertoire of processes and exploring the relation between specific reflective sub-
processes and memory, including the effects of various lesions on these specific
subprocesses.

CONCLUSIONS

A fundamental idea embodied in MEM is that it is useful to consider as separate

classes those processes initiated and maintained by external stimuli and those
initiated and maintained by internally directed reflective processes (e.g., John-
son & Raye, 1981). This idea can help in understanding findings from amnesia
and in fact has also figured in recent analyses of prefrontal cortex function
(Goldman-Rakic, 1987) and aging effects in memory (Craik, 1986).
Available evidence about preserved learning abilities in amnesics indicates
that the deficit is much more specific than was once thought. In terms of MEM,
the sensory and perceptual systems appear to be largely intact, whereas the
reflective system is disrupted. Furthermore, the reflective system may be only
partially disrupted. (I have concentrated on disruptions in reflective functions
128 FORMS OF MEMORY

because these seem to underlie the amnesias that have been most salient in the
cognitive literature, but according to the MEM framework there could be mem-
ory deficits associated with disruptions in the sensory and the perceptual sub-
systems as well.)
Clearly, amnesics are affected by events in important ways in spite of their
profound memory deficit. Their comprehension of the meaning of stimuli may
change, as indicated by comprehension ratings, context generation, and affective
responses. They may experience a feeling of familiarity and, under some con-
ditions, show good recognition after long intervals. If appropriately cued, they
may show surprising levels of recall and be able to use prior knowledge to aid
new learning.
These facts are not consistent with the idea that amnesics show no new "epi-
sodic" learning but can only reinforce or "prime" what is in "semantic memory"
already (Cermak, 1986). These results are also not easily accounted for by the
procedural-declarative distinction (Cohen & Squire, 1980). Ideas that seem
closer to describing amnesic deficits have been around for some time: that amne-
sia reflects a "premature closure of function" (Talland, 1965), failure of consol-
idation (Milner, 1966; Squire, 1982), disruption of vertical processes (Wickel-
gren, 1979) or mediated learning (Warrington & Weiskrantz, 1982), an encoding
deficit (Butters & Cermak, 1980; Cermak, 1979), a deficit in initial learning
(Huppert & Piercy, 1982), or a contextual encoding deficit (Hirst, 1982). These
ideas, although somewhat vague or unsatisfactory for various reasons, all focus
attention on the fact that amnesic processing is somehow attenuated. Such atten-
uated processing would disrupt both acquisition and retrieval.
MEM provides one potentially useful framework for further clarifying the
nature of amnesic processing. Within this framework, what amnesics seem to
have lost is internally guided access to more information than is directly acti-
vated by external stimuli. If more information is not needed, as in certain per-
ceptual/motor tasks, priming tasks, understanding ordinary conversation, and
recognition of novel stimuli, then amnesics do quite well. But if further reflec-
tion is required, an amnesic is at a severe disadvantage. Situations in which fur-
ther information (and thus reflection) is required include recognition tests in
which familiarity alone is not a sufficient cue but in which the source of the
familiarity must be specified, the reinstatement of certain types of affect, and, of
course, free recall of events. I have also suggested that progress in understanding
the role of reflection in amnesia might be made by further decomposing reflec-
tion into component subprocesses and then attempting to match these subpro-
cesses with neurological findings (e.g., specific lesion sites).

Acknowledgments
This chapter was prepared while I was a fellow at the Center for Advanced Study
in the Behavioral Sciences. I am grateful for financial support provided by the
John D. and Catherine T. MacArthur Foundation, Princeton University, and
National Science Foundation Grant BNS-8510633.1 would also like to acknowl-
 Functional Forms of Human Memory 129

edge the excellent assistance of Sharon Ray, Deanna Knickerbocker, Kathleen

Much, and Rosanne Torre (members of the CASES staff) in preparing the manu-
script, and the comments from Bill Hirst, Amy Weinstein, Liz Phelps, and
Shahin Hashtroudi on an earlier draft of this chapter. Special thanks to Bill Hirst
for our many spirited discussions about memory and amnesia.

Notes
1. Sensory and perceptual subsystems are called P-l and P-2, respectively, in subsequent
papers (Johnson, in press; Johnson & Hirst, in press).
2. In this chapter I do not discuss retrograde amnesia (impaired memory for events occurring
before the onset of amnesia), but focus on anterograde amnesia (impaired learning and memory
for events and information occurring after the onset of amnesia).
3. Recall and recognition are sometimes treated as measures of "declarative" memory that
should be equally disrupted by amnesia (e.g., Squire, 1982). Our comparison of recall and rec-
ognition in amnesics indicates that this is not always the case and, at the least, requires that the
concept of declarative memory be further analyzed into component subprocesses that may be
differentially disrupted.
4. These basic reflective subprocesses are typically fast-acting and perhaps they are largely
controlled by variations in activation level in the memory system (e.g., Johnson, 1983). For
example, amnesic deficits could be produced by turning down the amount of potential activa-
tion level contributed by experience (McClelland, 1985), by reducing the probability that a
given level of activation will recruit attention (Johnson, 1983), or by attenuating the spread of
activation.
5. With a greater range of difficulty level (or simply more subjects), comprehension ratings
(as well as later context generation) would very likely show an increasing deficit for amnesics
as difficulty is increased. Comprehension ratings are an example of a task that taps reflective
processes without requiring the subject to remember new information (see also Cermak, Reale,
& Baker, 1978). Such tasks should help clarify amnesic deficits in reflective processes.
6. Although the focus here is on amnesics' ability to use meaningful schemas, it should be
noted that amnesics also showed a marked improvement in recall of unrelated words when the
experimenter embedded the words in a bizarre story with high imagery value and gave amne-
sics spaced practice and cues when needed (Kovner, Mattis, & Goldmeier, 1983).
7. In Phelps et al. (1988) the relevant prior knowledge was presumably acquired before the
onset of amnesia. A second experiment (Phelps et al., experiment 2), however, demonstrated
that under certain conditions amnesics can also use information acquired after the onset of
amnesia to help in learning new information. (Thus they did not show the "hyperspecific"
learning reported by Glisky, Schacter, & Tulving, 1986. Also see Hirst, Phelps, Johnson, &
Volpe, 1988, and Shimamura & Squire, 1988, for other reports of "flexible" memory in
amnesics.)
8. Memory is also often presumed to be independent of something called "intellectual capac-
ity" and thus intelligence is said to be intact in pure amnesia uncomplicated by other problems.
But memory is critically involved in most intellectual tasks, from seeing previously unnoted
relationships to solving extremely complex problems. Standard IQ tests are probably not the
most sensitive way to look for deficits in intellectual functioning that may result from disrup-
tion of reflective processes.
9. It has in fact been suggested that there are two types of amnesia: bitemporal (patient H.M.
and ECT patients) and diencephalic (patient N.A. and Korsakoff patients) (Squire, 1982).
Although it is consistent with the present view of memory that "these two brain regions con-
tribute in different ways to normal memory functions" (Squire, 1982, p. 246), current evidence
is quite weak (Corkin, Cohen, Sullivan, Clegg, & Rosen, 1985). The case rests largely on studies
of forgetting rates in picture recognition (Huppert & Piercy, 1979b, Squire, 1981b), and one
central finding did not replicate (Freed, Corkin, & Cohen, 1984). More sensitive tests, however,
developed to discriminate among the various reflective subprocesses outlined here, might be
useful in discriminating bitemporal and diencephalic amnesics.
130 FORMS OF MEMORY

REFERENCES

Baddeley, A. (1982). Amnesia: A minimal model and an interpretation. In L. S. Cermak (Ed.),

Human memory and amnesia (pp. 305-336). Hillsdale, N.J.: Erlbaum.
Biber, C, Butters, N., Rosen, J., Gertsman, L., & Mattis, S. (1981). Encoding strategies and
recognition of faces by alcoholic Korsakoffand other brain-damaged patients. Jour-
nal of Clinical Neuropsychology, 3, 315-330.
Birnbaum, I. ML, Johnson, M. K., Hartley, J. T., & Taylor, T. H. (1980). Alcohol and elabora-
tive schemas for sentences. Journal of Experimental Psychology: Human Learning
and Memory, 6, 293-300.
Bower, G. H. (1970). Organizational factors in memory. Cognitive Psychology, 1, 18-46.
Bransford, J. D., & Johnson, M. K. (1973). Considerations of some problems of comprehen-
sion. In W. Chase (Ed.), Visual information processing. New York: Academic Press.
Butters, N., & Cermak, L. S. (1980). AlcoholicKorsakoffssyndrome: An information-processing
approach to amnesia. New York: Academic Press.
Cermak, L. S. (1979). Amnesic patients' level of processing. In L. S. Cermak & F. I. M. Craik
(Eds.), Levels of processing in human memory (pp. 119-139). Hillsdale, N.J.:
Erlbaum.
Cermak, L. S. (1982). Human memory and amnesia. Hillsdale, N.J.: Erlbaum.
Cermak, L. S. (1984). The episodic-semantic distinction in amnesia. In L. R. Squire & N. But-
ters (Eds.), Neuropsychology of memory (pp. 55-62). New York: Guilford.
Cermak, L. S. (1986). Amnesia as a processing deficit. In G. Goldstein & R. E. Tarter (Eds.),
Advances in clinical neuropsychology (Vol. 3, pp. 265-290). New York: Plenum.
Cermak, L. S., & Butters, N. (1972). The role of interference and encoding in the short-term
memory deficits of Korsakoff patients. Neuropsychologia, 10, 89-95.
Cermak, L. S., Butters, N., & Goodglass, H. (1971). The extent of memory loss in Korsakoff
patients. Neuropsychologia, 9, 307-315.
Cermak, L. S., Reale, L., & Baker, E. (1978). Alcoholic Korsakoff patients' retrieval from
semantic memory. Brain and Language, 5, 215-226.
Cermak, L. S., Talbot, N., Chandler, K., & Wolbarst, L. (1985). The perceptual priming phe-
nomenon in amnesia. Neuropsychologia, 23, 615-622.
Cohen, N. J. (1984). Preserved learning capacity in amnesia: Evidence for multiple memory
systems. In L. R. Squire & N. Butters (Eds.), Neuropsychology of memory (pp. 83-
103). New York: Guilford.
Cohen, N. J., & Squire, L. R. (1980). Preserved learning and retention of pattern analyzing skill
in amnesia: Dissociation of knowing how and knowing that. Science, 210, 207-
209.
Corkin, S., Cohen, N. J., Sullivan, E. V., Clegg, R. A., & Rosen, T. J. (1985). Analyses of global
memory impairments of different etiologies. Annals of the New York Academy of Sci-
ences, 444, 10-40.
Craik, F. I. M. (1986). A functional account of age differences in memory. In F. Klix & H.
Hagendorf (Eds.), Human memory and cognitive capabilities (pp. 409-422). Amster-
dam: North-Holland.
Cutting, J. (1981). Response bias in Korsakoff's syndrome. Cortex, 17, 107-112.
Ericsson, K. A., Chase, W. G., & Faloon, S. (1980). Acquisition of a memory skill. Science, 208,
1181-1182.
Eysenck, M. W. (1979). Depth, elaboration, and distinctiveness. (1979). In L. S. Cermak & F.
1. M. Craik (Eds.), Levels of processing and human memory. Hillsdale, N.J.: Erlbaum.
Freed, D. M., Corkin, S., & Cohen, N. J. (1984). Rate of forgetting in H.M.: A reanalysis. Soci-
ety for Neuroscience Abstracts, 10, 383.
Glenberg, A., Smith, S. M., & Green, C. (1977). Type I rehearsal: Maintenance and more. Jour-
nal of Verbal Learning and Verbal Behavior, 16, 339-352.
 Functional Forms oj Human Memory 131

Glisky, E. L., Schacter, D. L., & Tulving, E. (1986). Computer learning by memory-impaired
patients: Acquisition and retention of complex knowledge. Neuropsychologia, 24,
313-328.
Goldman-Rakic, P. S. (1987). Circuitry of primate prefrontal cortex and the regulation of
behavior by representational memory. In F. Plum (Ed.), Handbook of physiology:
Sec. 1. The nervous system: Vol. 5. Higher functions of the brain (pp. 373-417).
Bethesda, Md.: American Physiological Society.
Graf, P., Squire, L. R., & Mandler, G. (1984). The information that amnesic patients do not
forget. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10,
164-178.
Hirst, W. (1982). The amnesic syndrome: Descriptions and explanations. Psychological Bulle-
tin, 91, 435-460.
Hirst, W., Johnson, M. K., Kim,.J. K., Phelps, E. A., Risse, G.. & Volpe, B. (1986). Recognition
and recall in amnesics. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 12, 445-451.
Hirst, W., Johnson, M. K., Phelps, E. A., & Volpe, B. T. (1988). More on recall and recognition
in amnesia. Journal of Experimental Psychology: Learning, Memory, and Cognition,
14, 758-762.
Hirst, W., Phelps, E. A., Johnson, M. K., & Volpe, B. T. (1988). Amnesia and second language
learning. Brain and Cognition, 8, 105-116.
Huppert, F. A., & Piercy, M. (1976). Recognition memory in amnesic patients: Effects of tem-
poral context and familiarity of material. Cortex, 12, 3-20.
Huppert, F. A., & Piercy, M. (1977). Recognition memory in amnesic patients: A defect of
acquisition? Neuropsychologia, 15, 643-652.
Huppert, F. A., & Piercy, M. (1978). The role of trace strength in recency and frequency judg-
ments by amnesic and control subjects. Quarterly Journal of Experimental Psychol-
ogy, 30, 347-354.
Huppert, F. A., & Piercy, M. (1979a). Dissociation between learning and remembering in
organic amnesia. Nature, 275, 317-318.
Huppert, F. A., & Piercy, M. (1979b). Normal and abnormal forgetting in amnesia: Effect of
locus of lesion. Cortex, 15, 385-390.
Huppert, F. A. & Piercy, M. (1982). In search of the functional locus of amnesic syndromes. In
L. S. Cermak (Ed.), Human memory and amnesia (pp. 123-137). Hillsdale, N.J.:
Erlbaum.
Jacoby. L. L. (1982). Knowing and remembering: Some parallels in the behavior of Korsakoff
patients and normals. In L. S. Cermak (Ed.), Human memory and amnesia (pp. 97-
112). Hillsdale, N.J.: Erlbaum.
Jacoby, L. L., & Craik, F. I. M. (1979). Effects of elaboration of processing at encoding and
retrieval: Trace distinctiveness and recovery of initial context. In L. S. Cermak & F.
I. M. Craik (Eds.), Levels of processing and human memory. Hillsdale, N.J.: Erlbaum.
Jacoby, L. L., & Dallas, M. (1981). On the relationship between autobiographical memory and
perceptual learning. Journal of Experimental Psychology: General, 110, 306-340.
Johnson, M. K. (1983). A multiple-entry, modular memory system. In G. H. Bower (Ed.), The
psychology of learning and motivation: Advances in research and theory (Vol. 17, pp.
81-123). New York: Academic Press.
Johnson, M. K. (1987, February). The origin of memories. Paper presented at the annual meet-
ing of the Lake Ontario Visionary Establishment, Niagara Falls, N.Y.
Johnson, M. K. (1988). Discriminating the origin of information. In T. F. Oltmanns & B. A.
Maher (Eds.), Delusional beliefs: Interdisciplinary perspectives (pp. 34-65). New
York: Wiley.
Johnson, M. K. (in press). Reality monitoring: Evidence from confabulation in organic brain
disease patients. In G. Prigatano & D. L. Schacter (Eds.), Awareness of deficit after
brain injury. New York: Oxford University Press.
132 FORMS OF MEMORY

Johnson, M. K., Doll, T. J., Bransford, J. D., & Lapinski, R. H. (1974). Context effects in sen-
tence memory. Journal of Experimental Psychology, 103, 358-360.
Johnson, M. K., & Hirst, W. (in press). Processing subsystems of memory. In R. G. Lister &
H. J. Weingartner, Perspectives in cognitive neuroscience. New York: Oxford Uni-
versity Press.
Johnson, M. K., Hirst, W., Phelps, E. A., & Volpe, B. T. Context, comprehension and amnesia
(unpublished data).
Johnson, M. K., & Kim, J. K. (1985). Recognition of pictures by alcoholic Korsakoff patients.
Bulletin of the Psychonomic Society, 23, 456-458.
Johnson, M. K., Kim, J. K., & Risse, G. (1985). Do alcoholic KorsakofFs syndrome patients
acquire affective reactions? Journal of Experimental Psychology: Learning, Memory,
and Cognition, 11, 22-36.
Johnson, M. K., & Raye, C. L. (1981). Reality monitoring. Psychological Review, 88, 67-85.
Kinsbourne, M., & Wood, F. (1975). Short-term memory processes and the amnesic syndrome.
In D. Deutsch & J. A. Deutsch (Eds.), Short-term memory (pp. 258-291). New York:
Academic Press.
Kinsbourne, M., & Wood, F. (1982). Theoretical considerations regarding the episodic-seman-
tic memory distinction. In L. S. Cermak (Ed.), Human memory and amnesia (pp.
195-217). Hillsdale, N.J.: Erlbaum.
Kolers, P. A. (1975). Specificity of operations in sentence recognition. Cognitive Psychology, 7,
289-306.
Kovner, R., Mattis, S., & Goldmeier, E. (1983). A technique for promoting robust free recall in
chronic organic amnesia. Journal of Clinical Neuropsychology, 5, 65-71.
Levin, H. S., High, W. M., Meyers, C. A., Von Laufen, A., Hayden, M. E., & Eisenberg, H. M.
(1985). Impairment of remote memory after closed head injury. Journal of Neurol-
ogy, Neurosurgery, and Psychiatry, 48, 556-563.
Lindsay, D. S., & Johnson, M. K. (1987). Reality monitoring and suggestibility: Children's abil-
ity to discriminate among memories from different sources. In S. J. Ceci, M. P.
Toglia, & D. F. Ross (Eds.), Children's eyewitness memory (pp. 92-121). New York:
Springer-Verlag.
Mandler, G. (1967). Organization and memory. In K. W. Spence & J. T. Spence (Eds.), The
psychology of learning and motivation (Vol. 1, pp. 327-372). New York: Academic
Press.
Mandler, G. (1980). Recognizing: The judgment of previous occurrence. Psychological Review,
87, 252-271.
Mayes, A., Meudell, P., & Neary, D. (1980). Do amnesics adopt inefficient encoding strategies
with faces and random shapes? Neuropsychologia, 18, 527-540.
McAndrews, M. P., Glisky, E. L., & Schacter, D. L. (1987). When priming persists: Long-lasting
implicit memory for a single episode in amnesic patients. Neuropsychologia, 25, 497-
506.
McClelland, J. L. (1985). Distributed models of cognitive processes: Applications to learning
and memory. In D. S. Olton, E. Gamzu, & S. Corkin (Eds.), Memory dysfunctions:
An integration of animal and human research from preclinical and clinical perspec-
tives (Annals of the New York Academy of Sciences, Vol. 444). New York: New York
Academy of Sciences.
McDowell, J. (1979). Effects of encoding instructions and retrieval cuing on recall of Korsakoff
patients. Memory and Cognition, 7, 232-239.
Milner, B. (1966). Amnesia following operation on the temporal lobes. In C. W. M. Whitty &
O. L. Zangwill (Eds.), Amnesia (pp. 109-133). London: Butterworths.
Milner, B., Corkin, S., & Teubcr, H. L. (1968). Further analysis of the hippocampal amnesi'
syndrome: 14 year follow-up study of H.M. Neuropsychologia, 6, 215-234.
Milner, B., & Petrides, M. (1984, November). Behavioural effects of frontal-lobe lesions in mar
Transactions in Neuroscience, 403-407.
 Functional Forms of Human Memory 133

Mishkin, M., Malamut, B., & Bachevalier, J. (1984). Memories and habits: Two neural systems.
In G. Lynch, J. L. McGaugh, & N. M. Weinberger (Eds.), Neurobiology of learning
and memory (pp. 65-77). New York: Guilford.
Morton, J. (1985). The problem with amnesia: The problem with human memory. Cognitive
Neuropsychology, 2, 281-290.
Moscovitch, M. (1982). Multiple dissociations of function in amnesia. In L. S. Cermak (Ed.),
Human memory and amnesia (pp. 337-370). Hillsdale, N.J.: Erlbaum.
Moscovitch, M., Winocur, G., & McLachlan, D. (1986). Memory as assessed by recognition
and reading time in normal and memory-impaired people with Alzheimer's disease
and other neurological disorders. Journal of Experimental Psychology: General, 115,
331-347.
Nissen, M. J., & Bullemer, P. (1987). Attentional requirements of learning: Evidence from per-
formance measures. Cognitive Psychology, 19, 1-32.
Parkin, A. J. (1982). Residual learning capability in organic amnesia. Cortex, 18, 417-440.
Parkinson, S. R. (1979). The amnesic Korsakoff syndrome: A study of selective and divided
attention. Neuropsychologia, 17, 67-75.
Phelps, E. A., Hirst, W., Johnson, M. K., & Volpe, B. T. (1988). Amnesia and recall: Building
on prior knowledge. Unpublished manuscript.
Redington, K., Volpe, B. T., & Gazzaniga, M. S. (1984). Failure of preference formation in
amnesia. Neurology, 34, 536-638.
Ribot, T. A. (1977). Diseases of memory: An essay in positive psychology. In D. N. Robinson
(Ed.), Significant contributions to the history of psychology 1750-1920. Washington,
D.C.: University Publications of America. (Original work published 1882)
Schacter, D. L. (1987a). Memory, amnesia, and frontal lobe dysfunction. Psychobiology, 15,
21-36.
Schacter, D. L. (1987b). Implicit memory: History and current status. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 13, 501-518.
Schacter, D. L., Harbluk, J. L., & McLachlan, D. R. (1984). Retrieval without recollection: An
experimental analysis of source amnesia. Journal of Verbal Learning and Verbal
Behavior, 23, 593-611.
Schacter, D. L., & Tulving, E. (1982). Memory, amnesia, and the episodic/semantic distinction.
In R. L. Isaacson & N. E. Spear (Eds.), The expression of knowledge (pp. 33-65). New
York: Plenum.
Schank, R., & Abelson, R. (1977). Scripts, plans, goals and understanding. Hillside, N.J.:
Erlbaum.
Shimamura, A. P., & Squire, L. R. (1988). Long-term memory in amnesia: Cued recall, recog-
nition memory, and confidence ratings. Journal of Experimental Psychology: Learn-
ing, Memory, and Cognition, 14, 763-770.
Spear, N. E. (1984). Behaviours that indicate memory: Levels of expression. Canadian Journal
of Psychology, 38, 348-367.
Squire, L. R. (198 la). Memory and brain. New York: Oxford University Press.
Squire, L. R. (1981b). Two forms of human amnesia: An analysis of forgetting. Journal ofNcu-
roscience, 1, 635-640.
Squire, L. R. (1982). The neuropsychology of human memory. Annual Review ofNeuroscience,
5, 241-273.
Squire, L. R. (1986). Mechanisms of memory. Science, 232, 1612-1619.
Squire, L. R. (1987, November). Neural substrates of memory in humans and non-human pri-
mates. Paper presented at the annual meeting of the Psychonomic Society, Seattle.
Squire, L. R., & Shimamura, A. P. (1986). Characterizing amnesic patients for neurobehavioral
study. Behavioral Neuroscience, 100, 866-877.
Stein, B. S., & Bransford, J. D. (1979). Constraints on effective elaboration: Effects of precision
and subject generation. Journal of Verbal Learning and Verbal Behavior, 18, 769-
777.
134 FORMS OF MEMORY

Suengas, A. G., & Johnson, M. K. (1988). Qualitative effects of rehearsal on memories for per-
ceived and imagined complex events. Journal of Experimental Psychology: General,
117, 377-389.
Talland, G. A. (1965). Deranged memory. New York: Academic Press.
Tulving, E. (1962). Subjective organization in free recall of "unrelated" words. Psychological
Review, 69, 344-354.
Tulving, E. (1983). Elements of episodic memory. New York: Oxford University Press.
Warrington, E. K. (1982). The double dissociation of short- and long-term memory deficits. In
L. S. Cermak (Ed.), Human memory and amnesia (pp. 61-76). Hillsdale, N.J.:
Erlbaum.
Warrington. E. K., & Weiskrantz, L. (1982). Amnesia: A disconnection syndrome? Neuropsy-
chologia, 20, 233-248.
Weinstein, A. (1987). Preserved recognition memory in amnesia. Unpublished doctoral disser-
tation, State University of New York at Stony Brook.
Wickelgren, W. A. (1979). Chunking and consolidation: A theoretical synthesis of semantic
networks, configuring in conditioning, S-R versus cognitive learning, normal forget-
ting, the amnesic syndrome, and the hippocampal arousal system. Psychological
Review, 86, 44-60.
Zajonc, R. B. (1980). Feeling and thinking: Preferences need no inferences, American Psychol-
ogist, 35, 151-175.
COMMENTARIES AND
ALTERNATIVE PERSPECTIVES
This page intentionally left blank
6
Neuromnemonics: Forms and Contents
N O R M A N M. W E I N B E R G E R

"Forms of memory" encompasses a broad and heterogeneous set of problems

and approaches. The preceding chapters provide an indication of the extensive
dimensions of current research.
Carew and his co-authors are concerned with different forms of memory
within the domain of the neurophysiology of habituation, sensitization, and dis-
habituation in a marine gastropod mollusk. Holland also deals with forms of
memory from the viewpoint of conditioning, but he does so within the frame-
work of behavioral analyses alone; the conditioning paradigms used with rats
are considerably more complex than those employed by Carew. Johnson focuses
on three mnemonic processes in normal and amnesic humans, using tasks quite
different from those involved in conditioning as well as stimulus material
unique to human animals, that is, language-based stimuli. Morris emphasizes
the importance of the network within which a mechanism of plasticity is embed-
ded; he argues that although there may be a "cell alphabet" of plasticity, there
is not yet an alphabet of learning.
In this short commentary I will try to point out some agreements and disa-
greements among the authors and emphasize some issues they do not cover.
Specifically, I will deal with some aspects of the contents of memories, in dis-
tinction to forms, types, or processes of neuromnemonics. Future directions
toward achieving closer relations between studies of memory in humans and
other animals will also be considered briefly.

FORMS OF MEMORY

There is not uniform agreement concerning how one delineates forms of mem-
ory. The criteria to determine if there are multiple forms of memory focus on
dissociations of one sort or another. Three chapters seem to consider dissocia-
tions as both necessary and sufficient criteria for the establishment of different
forms, across species and learning situations; in one, judgment is reserved.
Carew and his colleagues argue that habituation, sensitization, and dishabit-

137
138 FORMS OF MEMORY

uation are three different forms of memory because they appear sequentially
during ontogenesis in a gastropod mollusk.
For Johnson, there are three sets of memory subsystems or processes: sensory,
perceptual, and reflective. These are established by behavioral manipulations in
normal and amnesic adults by the use of various tasks that a priori tap these
various processes. Amnesics appear to be impaired most in the use of the reflec-
tive subsystem. Whether one refers to forms or subsystems, the thrust of John-
son's formulation is that of tripartite processing of stored experience.
Although Morris does not explicitly address the criteria for forms, his analysis
of the theoretical position of Kandel implicitly accepts the division of some
forms of learning into sensitization, pseudoconditioning on the one hand and
alpha- and beta-conditioning on the other hand. His emphasis is on relation-
ships between the former and the latter. Within the present context, this quali-
tative distinction is in the spirit of different forms of memory.
Holland provides evidence for independence between three pairs of subsys-
tems: S-S versus S-R, CS-OR versus CS-UR pathways, and simple associative
versus occasion-setting functions. Nonetheless, and unlike the other authors,
Holland is reluctant to conclude that these constitute different forms of memory,
suggesting that reification requires an explication of "functional incompatibil-
ity" within the context of the separate purposes served by each reputed system
(Sherry & Schacter, 1987).
Like Holland, I am uncomfortable in having forms of memory proliferate
with every demonstrable dissociation. The multiplication of types, particularly
in the absence of adequate criteria, may not be particularly illuminating. This is
especially so when reification entails little more than a restatement of opera-
tional definitions. My reservations concern the criteria, particularly the impor-
tance of dissociations. In brief, a dissociation may be a necessary criterion, but
it does not seem to be a sufficient criterion.
For example, dissociations could occur if two behavioral manifestations of a
single underlying process have different thresholds. Regarding this point, I am
reminded of one of the major reasons for extensive use of the nictitating mem-
brane-conditioned response in the rabbit: it is reputed to exhibit no response to
a CS prior to pairing with airpuff to the cornea or shock to the orbit, so that the
conditioned response is a beta rather than an alpha type—a new response rather
than an elaboration of an existing response. These observations are valid at the
behavioral level, but they do not permit conclusions at the neural level. Thus
Cegavske (1974) reported that although an acoustic stimulus, to be used as the
CS, did not produce neuromuscular responses, it did evoke neuronal discharges
in the final common path, the nucleus of the VIth cranial nerve. In other words,
there was an alpha response at the level of the motor neurons but a beta CR at
the level of behavioral detection of conditioning. Moreover, while the behav-
ioral detection of learning at the nictitating membrane exhibits no sensitization
or pseudoconditioning, Cegavske, Patterson, and Thompson (1979) reported
evidence for these processes in the Vlth cranial nucleus. In short, the use of a
dissociation as a necessary criterion for establishing separate forms or processes
 Neuromnemonics: Forms and Contents 139

of learning and memory should be viewed with extreme caution if not consid-
erable skepticism.
From another viewpoint, one may ask about the extent to which dissociations
imply more than a lack of identity. If they do, what are they? If not, then is it
fundamentally different to speak of forms of memory versus forms of limbs,
forms of locomotion, forms of goal-attainment, and so on? In short, we must
take care not to allow the common language terms, or "folk psychology lan-
guage," to assume a role of importance that they do not merit.
At least two approaches might be helpful. First, one might look to other sci-
ences—such as biological taxonomy and systematics—for methodologies for
categorizing forms of memory (e.g., Tulving, 1985). This possibility immedi-
ately raises the question of the extent to which various forms of memory are
related, as opposed to emphasi/ing dissociations that foster (but do not demand)
simpler yes or no answers. This interrelatedness is certainly a very real concern
of all of the authors.
Second, one might look to a logical analysis and a thorough explication of the
various meanings of "forms" of memory. This includes the distinction of Mor-
ris—that a single cellular basis for information storage could be expressed as
different forms of memory, depending on the network within which it is embed-
ded. In this case, one could argue for a single form of memory at a cellular level
but various forms of memory at a network level.
Both approaches are likely to result in a somewhat more precise and helpful
formulation of forms than we now have.

INVOLVEMENT OF SENSORY SYSTEMS

There seems to be a surprisingly high degree of agreement concerning the fact

that sensory systems are intimately involved in the storage of information. Thus
for Carew and his co-authors the primary sensory neurons of Aplysia are actual
sites of learning-induced plasticity. Morris accepts these findings. Holland con-
cludes that associative conditioning can engage the CS "early" in the chain that
ultimately leads to CRs. His conclusion is based on evidence that the CS can
acquire the sensory properties of the US. Johnson's evidence supports the
dynamic involvement of sensory and perceptual systems in learning and mem-
ory, in addition to the reflective processes to which most attention has been
given.
This high degree of agreement is noteworthy in view of a tradition in which
sensory systems have been viewed as of little or no importance in the dynamics
of learning and memory (e.g., Fuster, 1984). One major reason for the reluctance
to accord sensory systems the dynamic, learning-involved properties that have
been repeatedly and exhaustively established since the mid-1940s (see Weinber-
ger & Diamond, 1987, for a review of auditory cortex and learning) seems to
center on a paradox. There seems to be a fundamental conflict between having
sensory system neurons subject to two types of influence: the physical parame-
140 FORMS OF MEMORY

ters of stimuli as transduced by a sheet of sensory epithelium, and the "psycho-

logical parameters" or acquired meaning of stimuli (Weinberger & Diamond,
1988). Neurons subject to both influences would provide ambiguous "reports";
for example, increased discharge might indicate increased stimulus intensity,
increased psychological salience, or both.
We have provided both data and conceptual support for a solution to this
paradox in the auditory system (Ryugo & Weinberger, 1978; Weinberger, 1980,
1982, 1984). Briefly, parallel subcortical paths may deal separately with the
physical and psychological aspects of sound. The two paths (lemniscal and lem-
niscal-adjunct, respectively) coalesce in auditory cortex where in fact neuronal
discharges represent their combined influence in pyramidal neurons. Strictly
speaking, perception is the result.
This issue seems to have received little attention in the neurobiology of learn-
ing and memory. Thus the inherent ambiguity described above must be present
in any organism in which learning alters primary sensory neurons, as in Aplysia
or primary sensory receptors, as in Hermissenda (Alkon, 1986). It is not clear
how or if this ambiguity is resolved in such organisms. A fundamental change
in the role of sensory systems in learning may have evolved for vertebrates,
more complex invertebrates, or both.
As to the engagement of the CS early in the sensory network that processes
the US, Holland has provided evidence that the CS may actually "take on" some
characteristics of the US. One might expect neurons in the CS path to acquire
some of the properties of the US. That this may take place at the level of the
magnocellular medial geniculate nucleus in the auditory system is suggested by
a series of studies by LeDoux and associates. The circuit for conditioned fear to
sound apparently goes from the lemniscal-adjunct auditory thalamus to the
amygdala (LeDoux, Sakaguchi, & Reis, 1983; LeDoux, Sakaguchi, Iwata, &
Reis, 1986; Iwata, LeDoux, Meeley, Arneric, & Reis, 1986). Thus while the story
of brain and learning/memory is still incomplete, it would seem that sensory
system plasticity will be a part of the large picture.

CONTENTS OF MEMORIES

There is another side to memory in addition to its forms: its contents. Inferences
about what is learned—more specifically, what is stored and retrievable—are
based on behavioral tests. Detailed characterization of such contents of memory,
in distinction to the processes involved in storage and retrieval, usually requires
precise tests following the learning experience in question. Failures to detect evi-
dence of learning are a function of the sensitivity of the test, including the type
of question posed by the experimenter. A full account of what was learned is
seldom attempted because of the dauntingly large number of behavioral tests
required. Reasonably, behavioral tests focus on the particular topic of the exper-
imenter's interests, that is, the purpose of the experiment. For example, unless
specifically tested for, the extent of contextual dependence is generally unknown.
 Neuromnemonics: Forms and Contents 141

As an example, Morris cites the oft-ignored report of Wagner (1978) indicating

that habituation can be context dependent.
Neurobiological approaches to memory offer the potential for determining all
of the information stored and the possible postlearning alterations, transfor-
mations, and even loss of the contents of storage. Since all behaviors are expres-
sions of neurobiological structure and function, the neurobiological measures
are potentially more sensitive and direct than are behavioral data. Currently
used behavioral measures have not only the advantage of a resonably well estab-
lished methodology, but provide integrated output from the nervous system.
Neurobiological measures could concentrate on the "trees" while missing the
behavioral "forest."
Nonetheless, it should be admitted that both behavioral and neurobiological
research on memory is largely the study of processes and forms rather than the
study of the contents of memories. A notable attempt to use neurophysiology to
read out content is the corpus of research by E. Roy John (e.g., 1967). This
research attempted to use the frequency of flashes as "tracers" for memory con-
tent (see also John, 1980). Regardless of the ultimate status of this body of work,
it seems clear that few neurobiologists have since trod that path.
To the extent that experiments on the content of memory continue to be dis-
regarded because of technical problems, it is likely that theoretical formulations
and their tests will also ignore issues relating to memories themselves. From this
viewpoint, agreement about forms of memory and about the cellular and net-
work bases of memory storage processes, in the absence of formulations about
the nature of what is actually stored, will still leave us with a highly unsatisfac-
tory account of memory. When neurobiological knowledge of the contents of
memory can be used to accurately predict specific memory-dependent behavior,
not limited to the contractions of a limited set of muscles, then we may achieve
a comprehensive understanding of memory.
The agreement among the four chapters regarding the involvement of sensory
systems in learning and memory suggests, as John argued, the possibility of
"reading out" something about content, particularly within sensory systems.
Through work by my colleagues and me on learning in the auditory system,
something is now known about the "content" of the auditory cortex plasticity.
It is not expressed as a general change in responsiveness to acoustic stimuli.
Rather, the learning-dependent plasticity that rapidly develops in auditory cor-
tex is highly specific to the frequency of the tone used as the conditioned stim-
ulus (Diamond & Weinberger, 1986). The same type of findings have been
obtained for habituation; the decrement in response is specific to the repeated
frequency (Condon & Weinberger, in preparation; Westenberg & Weinberger,
1976). These types of data, in which changes in a stimulus dimension are
induced by learning, may provide a basis for reading out content, to the extent
that the change in representation of the CS, as part of the domain of acoustic
frequency, for example, can be revealed. It seems feasible to extend these studies
to humans because tonotopic organization has been found in human auditory
cortex using noninvasive magnetoencephalography (MEG, Romani et al., 1982).
142 FORMS OF MEMORY

FUTURE DIRECTIONS

That a major gap exists between understanding of human and animal memory
should come as no surprise. Carew and co-authors, Holland, Johnson, and Mor-
ris do not address the problem explicitly, but they illustrate the situation by
emphasizing habituation and classical conditioning tasks with animals versus
the types of tasks Johnson uses with humans. At the same time, some tasks seem
to have a lot in common. Examples are found in the chapters by Holland and
Johnson involving context, stimulus devaluation, and the use of complex tone
sequences.
Regarding the first point, Holland's discussion of "occasion setting" in rats
seems to be closely related to studies of context in humans. From a formal point
of view, both involve control or modulation of relationships between events by
designated cues—conditional learning. In the case of humans, the context is
often more complex than in the rat experiments, such as place versus a flashing
cue light. Nonetheless, the conditional control of associations can be explored
in parallel in both types of subject.
Stimulus devaluation is widely used in animal experiments but not in
humans. However, Johnson presented an intriguing example of stimulus deval-
uation ("John is good; Bill is bad"). She concluded that stimulus devaluation is
an effective means of dissecting memory processes.
Because of the heavy reliance on language in studies of human memory, there
is an insurmountable gap between human and animal research. However, there
are alternatives between presenting animals with isolated stimuli in condition-
ing experiments and presenting humans with linguistic material. I refer here to
music as stimulus. The processing of information and its storage may be quite
different for individual stimuli and for stimulus sequences that are extended in
time. Certainly some aspects of language involve such extended sequences. Fur-
thermore, there are both human (e.g., Warren, 1982) and animal (e.g., D'Amato,
in press) studies of memory for acoustic sequences. Since single neurons in audi-
tory cortex are sensitive to the structure of sequences (Weinberger & McKenna,
1988), it seems of interest to (1) use highly similar experimental designs for
humans and animals and (2) extend neurophysiological approaches to human
auditory cortex.

CONCLUSIONS

This is an exciting time for memory research and for the prospects of achieving
an understanding at the cellular level through behavioral insights into the acqui-
sition and storage of information. As more data accumulate, it becomes increas-
ingly urgent that the conceptual frameworks be developed and that more vig-
orous formulations of key concepts be achieved.
 Neuromnemonics: Forms and Contents 143

REFERENCES

Alkon, D. L. (1986). Persistent calcium-mediated changes of identified membrane currents as

a cause of associative learning. In D. L. Alkon & J. Farley (Eds.), Primary neural
substrates of learning and behavioral change (pp. 291-325). New York: Cambridge
University Press.
Cegavske, C. F. (1974). An approach to the neural substrates of classically conditioned rabbit
nictitating membrane response. Unpublished doctoral dissertation, University of
California, Irvine.
Cegavske, C. F., Patterson, M. M., & Thompson, R. F. (1979). Neuronal unit activity in the
abducens nucleus during classical conditioning of the nictitating membrane response
in the rabbit. Journal of Comparative and Physiological Psychology, 93, 595-609.
Condon, C., & Weinberger, N. M. (in preparation). Habituation in auditory cortex of guinea
pig: Frequency-specific decrements in receptive fields.
D'Amato, M. R. (1988). A search for tonal pattern perception in Cebus monkeys: Why mon-
keys can't hum a tune. Music Perception, 5, 453-480.
Diamond, D. M., & Weinberger, N. M. (1986). Classical conditioning rapidly induces specific
changes in frequency receptive fields of single neurons in secondary and ventral ecto-
sylvian auditory cortical fields. Brain Research, 372, 357-360.
Fuster, J. M. (1984). The cortical substrate of memory. In L. R. Squire & N. Butters (Eds.),
Neuropsychology of memory (pp. 279-286). New York: Guilford.
Iwata, J., LeDoux, J. E., Meeley, M. P. Arneric, S., & Reis, D. J. (1986). Intrinsic neurons in
the amygdaloid field projected by the medial geniculate body mediated emotional
responses conditioned to acoustic stimuli. Brain Research, 383, 195-214.
John, E. R. (1967). Mechanisms of memory. New York: Academic Press.
John, E. R. (1980). A neurophysiological model of purposive behavior. In R. F. Thompson, L.
H. Hicks, & V. B. Shvykov (Eds.), Neural mechanisms of goal-directed behavior and
learning (pp. 93-115). New York: Academic Press.
LeDoux, J. E., Sakaguchi, A., Iwata, J., & Reis, D. J. (1986). Interruption of projections from
the medial geniculate body to an archi-neostriatal field disrupts the classical condi-
tioning of emotional response to acoustic stimuli. Neuroscience, 17, 615-627.
LeDoux, J. E., Sakaguchi, A., & Reis, D. J. (1983). Subcortical efferent projection of the medial
geniculate nucleus mediated emotional responses conditioned to acoustic stimuli.
Journal of Neuroscience, 4, 683-698.
Romani, G. L., Williamson, S. J., Kaufman, L., & Brenner, D. (1982). Characterization of the
human auditory cortex by the neuromagnetic method. Experimental Brain Research,
47, 381-393.
Ryugo, D. K., & Weinberger, N. M. (1978). Differential plasticity of morphologically distinct
neuron populations in the medial geniculate body of the cat during classical condi-
tioning. Behavioral Biology, 22, 275-301.
Sherry, D. F., & Schacter, D. L. (1987). The evolution of multiple memory systems. Psycholog-
ical Review, 94, 439-454.
Tulving, E. (1985). On the classical problem of learning and memory. In L. G. Nilsson & T.
Archer (Eds.), Perspectives on learning and memory (pp. 67-95). Hillsdale, N.J.:
Erlbaum.
Wagner, A. R. (1978). Expectancies and the priming of STM. In S. H. Hulse, H. Fowler, & W.
K. Honig (Eds.), Cognitive processes in animal behavior. Hillsdale, N.J.: Erlbaum.
Warren, R. M. (1982). Auditory perception. A new synthesis. New York: Pergamon Press.
Weinberger, N. M. (1980). Neurophysiological studies of learning in association with the pupil-
lary dilation conditioned reflex. In R. F. Thompson, L. H. Hicks, & V. B. Shvyrkov
(Eds.), Neural mechanisms of goal-directed behavior and learning (pp. 241-261). New
York: Academic Press.
144 FORMS OF MEMORY

Weinberger, N. M. (1982). Sensory plasticity and learning: The magnocellular medial geniculate
nucleus of the auditory system. In C. D. Woody (Ed.), Conditioning: Representation
of involved neural function (pp. 697-710). New York: Plenum.
Weinberger, N. M. (1984). The neurophysiology of learning: A view from the sensory side. In
L. Squire & N. Butters (Eds.), The neuropsychology of memory (pp. 489-503). New
York: Guilford.
Weinberger, N. M., & Diamond, D. M. (1987). Physiological plasticity in auditory cortex:
Rapid induction by learning. Progress in Neurobiology, 29, 1-55.
Weinberger, N. M., & Diamond, D. M. (1988). Dynamic modulation of the auditory system
by associative learning. In G. M. Edelman, W. E. Gall, & W. M. Cowan (Eds.), Audi-
tory function: The neurobiological bases of hearing (pp. 485-512). New York: Wiley.
Weinberger, N. M., & McKenna, T. M. (1988). Sensitivity of single neurons in auditory cortex
to frequency contour: Toward a neurophysiology of music perception. Music Percep-
tion, 5, 355-389.
Westenberg, I. S., & Weinberger, N. M. (1976). Evoked potential decrements in auditory cortex,
II: Critical test for habituation. Electroencephalography and Clinical Neurophysiol-
ogy, 40, 356-369.
7
Time and Memory
ROBERT W. DOTY

While memory preserves times past, a record that will instruct an organism's
future, time also enters the mnemonic process in many other guises. A critical
aspect is how long the mnemonic trace will endure, a time stretching from sec-
onds to a lifetime. Vastly shorter, yet meaningfully measurable to some degree,
is the time it takes to retrieve a memory.
The time between the occurrence of two successive events is also important.
This temporal relation will determine whether, and in what direction, an endur-
ing association will be formed between the events, as in the rules found appli-
cable in forming conditioned reflexes. In terms of expectations from neurophys-
iology (e.g., Auyeung, May, Goh, & Sastry, 1987; Larson & Lynch, 1986), the
direction of this association is paradoxical, for in the formation of a conditioned
reflex it is the response to the first occurring event, the conditional stimulus, that
becomes altered, whereas the response to the subsequent, unconditional stimu-
lus is stable. I have termed this the "temporal paradox" of conditioning (Doty,
1979). This relation prevails for the usual type of Pavlovian conditioning, even
in bees (Menzel, 1983), as well as when assayed with electrical excitation of ran-
domly chosen points in mammalian neocortex (Doty & Giurgea, 1961) or, with
entirely different parameters, for establishing long-term potentiation in the hip-
pocampus with stimulus relations similar to those in conditioning (Levy &
Steward, 1983). The temporal paradox is thus an underlying principle in the
formation of such associations.
A more subtle temporal feature of mnemonic phenomena is the distinction
that memories for events carry a relative "date," so that their sequence as well
as their identity can be retrieved, whereas memories for "rules" or "habits" need
no such tag, nor do they usually bear one. This is by now a familiar dichotomy
(e.g., Mishkin, Malamut, & Bachevalier, 1984; Squire & Cohen, 1984; Thomas,
1984) and, as discussed below, there is strong reason to suppose that it is the
hippocampus which somehow attaches the "date" to memory for unique events.
To varying degrees these themes of time and memory run through what fol-
lows; more important, they provide a number of opportunities for understand-
ing the nature of mnemonic processes.

145
146 FORMS OF MEMORY

SLUGS AND SLUGGISHNESS

As Morris (Chapter 3, this volume) points out, there is a remarkable similarity

in the optimum times for initiation of associative effects in mammals and the
sea slug, Aplysia—something of the order of 500 msec in each case. Similarly,
Carew and colleagues (Chapter 2, this volume) used interstimulus intervals of
1, 5, and 10 sec in a study of habituation in the slug, times wholly familiar from
many comparable studies on mammals. Why, it must be asked, is the slug not
sluggish in this regard? The answer is probably that the determinants are phys-
ical processes rather than chemical reactions in each case and therefore, given
the comparable sizes of the critical elements (neurons, their terminals, and the
intercellular spaces), are relatively independent of the temperature differences
between the mammal and the slug (see, e.g., Hober, 1945).
The slug, too, is rather facile in the time it takes to recover from the pertur-
bation of its synapses achieved by habituation, at least if the series of stimuli is
not repeated. Again, the time course of recovery may reflect what is basically a
diffusional process, complicated by sequestration, transport, or ionically con-
trolled availability of transmitter. In other words, retention of this marginally
mnemonic effect fits the earlier classification (Doty, 1979) as being "ionic"
rather than "macromolecular."
The term macromolecular was introduced to indicate the probability that
mnemonic processes that result in an essentially permanent trace and display
the temporal paradox ultimately will involve more than a shuffling of ionic con-
centrations or permeabilities. A restructuring of synaptic-postsynaptic relations
at a complex chemical level is implied by a macromolecular form of memory.
This was originally visualized as being analogous to an immunochemical bond-
ing, a concept that may still have merit. However, it now appears that although
macromolecular changes occur locally at sites of altered synaptic function in a
mnemonic context (e.g., Alkon, 1986; Hawkins, Clark, & Kandel, 1987; Hu et
al., 1987; Lynch, 1986; Lynch & Baudry, 1984; Montarolo et al, 1986), they are
unlikely to be permanent.
The rather ephemeral nature of these mnemonic effects may well be related to
the metabolic half-life of the relevant macromolecules. Indeed, this severe prob-
lem is difficult to attack and hence often ignored in endeavors to understand
how conditioned reflexes or human memory endures in the face of continual
turnover in the chemical constituents at the presumed site of storage. The max-
imal half-life of neuronal proteins seems to lie in the order of a few weeks, most
being much shorter (Dunlop, 1983; Margolis & Margolis, 1983). It is thus logical
to suppose that this metabolic instability is counteracted by instructions at the
nuclear level, modulating transcription of the genome to provide a continuing
supply of the chemical constituents needed to maintain the mnemonically
altered structure. Thus, although the effect is macromolecular, the essential fea-
ture of enduring memory may be that it is "transnuclear."
This was Halstead's original thesis (e.g., 1951): neuronal memory should util-
 Time and Memory 147

ize the already perfected processes of genetic memory. This concept fostered the
examination of memory in relation to disruption of protein synthesis that has
continued for three decades. The timing in such experiments, however, may go
awry, for it is possible that the presumably chemical signal, which must pass
from terminal or dendrite into the nucleus, is for some time inaccessible to
manipulation during its transport. Kesner, McDonough, and I (1970) subjected
cats to an aversive experience, followed 4 sec later by an electroconvulsive sei-
zure. When tested 24 hr later, they were amnesic for the experience; but if 5 to
15 days later they received a second experience again followed in 4 sec by a
seizure, the aversive experience was remembered. In other words, within 4 sec
after the experience enough of a "trace" has been formed and become inacces-
sible to disruption by seizure activity that two such experiences formed a robust
memory.
That synaptic processes can induce modulation of enzyme synthesis is now
well documented for sympathetic ganglia (Thoenen, Otten, & Schwab, 1979)
and the adrenal medulla (Guidotti & Costa, 1977). Furthermore, Thoenen and
his colleagues showed that presynaptic terminals can imbibe chemical moieties,
such as nerve growth factor, and transport them retrogradely to the nucleus
where they alter protein production. Thus a mechanism can be perceived
whereby both presynaptic and postsynaptic elements could mutually exchange
some chemical constituent, provided that the timing of events in these two ele-
ments is appropriate (the "temporal paradox"?). Attaining the nucleus by micro-
tubular transport, this signal might then permanently alter genetic transcription
to produce whatever chemistry it may be that renders effective and permanent
the connections of one neuron selectively to another.
Should it be true that memory as well as the regulation of neuronal intercon-
nectivity is dependent on imbibition and subsequent transport of some chemical
moiety to the cell nucleus, there to effect a modulation of the genome, then it is
understandable why both vertebrate and invertebrate nervous systems (e.g., Kuf-
fler, 1967; Landolt, 1965) are protected by a glial barrier from inadvertent mac-
romolecular contamination. Unlike the immune system, whose function
requires genomic response to newly present molecular species, function of the
nervous system would, in the present speculation, be severely disrupted by ran-
dom exposure to irrelevant molecules; it thus remains "immunologically privi-
leged," shielded by the blood-brain barrier.
It must be noted that no permanent alteration, of either genomic output or
mnemonic result of transnuclear effects, has yet been observed. The durations,
so far, are measured in days or hours (Alkon, 1986; Guidotti & Costa, 1977; Hu
et al., 1987; Montarolo et al., 1986; Thoenen et al., 1979). This might, of course,
mean that the idea is merely fanciful, and that establishing permanent, mne-
monically relevant synaptic alterations does not involve the hypothesized trans-
nuclear mechanism. On the other hand, the systems amenable to test are not
necessarily those in which a permanent change might be manifested or expected.
There is, indeed, some question as to whether with their shorter life span and
148 FORMS OF MEMORY

changing neuronal configurations (e.g., Carew et al., Chapter 2, this volume)

invertebrates need or would benefit from permanently altered neural responses
consequent to ephemeral environmental events.

SNARSKY'S PROBLEM

Not long after he began analyzing the nature of the "psychic secretions" he had
observed from salivary glands of dogs, Pavlov set his student Snarsky to study-
ing the effect of different materials, what we now call the unconditional stimuli,
upon this secretion. Snarsky used three stimuli: sand, stones, and dry food,
placed in the animal's mouth. He soon found that the dog produced a different
type of salivation merely to the sight of these items (Pawlow, 1904). How was
this to be explained? Pavlov asked, rhetorically, "Must one in order to under-
stand these new phenomena immerse oneself in the mental state of the animal,
endeavor to imagine its sensations, ideas, feelings, and desires similar to our
own?" (Doty, 1987). This was the type of interpretation that Snarsky offered;
but Pavlov, following his mentor Heidenhain, vehemently opted for limiting
himself to what he considered to be the physiology.
Almost a century later Snarsky's approach is still alive and well in psychology.
Holland (Chapter 4, this volume) concedes that "rats probably do not halluci-
nate perfect copies of absent USs" (italics added), yet his endeavors seem largely
directed to the possiblity that they might summon up some form of mental
image, for example, a "representation of food" activated by a tone. Holland, of
course, is not alone in schematizing what may go on in the "black box" of the
rat's mind, leaving, as he says, the plausibility of neurobiological analysis far
behind. This is the stock in trade of cognitive science, discovering the rules with-
out the neurons. The philosopher Searle (1984) astutely noted this penchant of
psychologists for discerning "rules." He questions whether the rules can reveal
the mechanism, for, as he says, "You don't need to suppose that there are any
rules on top of the neurophysiological structures" (p. 51).

HIPPOCAMPUS AND THE STRING OF TIME

One of the common features acquired in the evolution of higher motile organ-
isms is the ability to maintain orientation with respect to places of forage or
safety. This is as true of the sand crab venturing along various radii away from
its hole as of the fox patrolling the likely hunting spots of its territorial domain.
Among the most remarkable and best studied examples of this orientational
memory is that for vision in insects. In his meticulous and fascinating review
Wehner (1981) lists some 140 species and about 200 references on this topic (see
also Cartwright & Collett, 1983; Menzel, 1983).
Wasps and bees are particularly precise in this regard, ranging over a kilo-
meter or more, yet returning within a few centimeters to the nest or hive. Among
 Time and Memory 149

the essential features of this performance are, first, the recording of the sequence
of landmarks encountered on the outward path, and then, second, the ability to
navigate by these guides as they are encountered in reverse order during the
return flight. In other words, there must be a "start recording" marker at the
onset of the flight, differentiating those movements in time from those made
previously; and there must then be a faithful registration of the sequence of the
visual patterns encountered from that moment until the beginning of the return
flight.
Like the sand crab, fossorial mammals must maintain an unerring orientation
as to their position relative to the safety of their den, and larger mammals,
including primitive humans, range over great distances without becoming
"lost." There are two clues that the hippocampus plays a critical role in this
ability. First is the fact that, in the rat, there is a clear candidate for a "now
record" marker, the onset of theta rhythm whenever the rat begins to move itself
through space (e.g., Vanderwolf & Leung, 1983). It is of particular interest that
this rhythm is substantially controlled by pontine neurons (see Vertes, 1986), a
site from which vestibular influences might readily be introduced. The relevance
of vestibular input can be perceived from the experiments of Beritashvili, show-
ing that the human deaf who lack the labyrinthine system are unable to repro-
duce blindly a path over which they have been led, whereas the deaf who lack
only cochlear function are able to do so (Beritachvili, 1963).
The second clue comes from O'Keefe's discovery (see O'Keefe, 1983), now so
rigorously denned by Ranck and his colleagues (e.g., Muller, Kubie, & Ranck,
1987), that many single units in the hippocampus of the rat reflect the animal's
spatial location in their rate of discharge. It is also now clear that the rat needs
its hippocampus to record its past positions in space (Morris, 1983, and Chapter
3, this volume; O'Keefe, 1983; Olton, 1983).
The well-studied case of H.M. (see Corkin, 1984) defined the nature of the
deficit in man following bilateral loss of amygdala, hippocampus, and much of
temporal cortex. The observations of Penfield and Mathieson (1974) and of
Zola-Morgan, Squire, and Amaral (1986), however, suggest that most of the def-
icit arises from the hippocampal loss alone. Yet there is a slight problem here:
first, the comparable syndrome in macaques seems also to require removal of
amygdala as well as hippocampus (Mishkin, 1978; Murray & Mishkin, 1983);
but, second, if H.M. is tested in the same manner as the macaques, he displays
no deficit (Freed, Corkin, & Cohen, 1987).
Despite such discrepancies, there is still general agreement that the hippocam-
pus is intimately involved in the recording and/or the retrieval of human mem-
ory for events, although it is a challenge experimentally to disentangle the
recording from the retrieving. Furthermore, all normal human beings can testify
that the sequence of events is a common and often essential part of their mne-
monic record. Thus it can be hypothesized that in humans and other mammals
the hippocampus subserves a phylogenetically ancient function, similar to that
required by many flying insect species: upon a "start" signal, for example, as
movement through space begins, its function is to tag each sensorial frame for
150 FORMS OF MEMORY

subsequent identification of its position in the sequence. Indeed, this very pro-
cedure, an imaginary walk, was the principal method used by Luria's (1968) pro-
digious mnemonist in tagging items for subsequent retrieval. Compared with
that in rodents, the "start" signal in primates may be relatively continuous, orig-
inating in the massive neocortical input to the hippocampus (Rosene & Van
Hoesen, 1987) rather than or in addition to brain stem input associated with
movement. This might explain the relative paucity (Crowne & Radcliffe, 1975;
Isokawa-Akesson, Wilson, & Babb, 1987) of theta rhythm in primates or pres-
ervation of normal human mentality despite congenital absence of the fornix
(Nathan & Smith, 1950).
It remains to be noted that although the hippocampus seems essential to the
recording of memory for events, it is very unlikely to be the site of storage. This
can be deduced, first, from the fact that hippocampal loss does not eliminate
access to most previously established memories (e.g., Zola-Morgan et al., 1986)
and, second, from the sheer magnitude of the cortical neuronal circuitry avail-
able for retaining a near infinity of memories, compared with the mere 10 mil-
lion pyramidal cells bilaterally in CA1 whose loss so dramatically perturbs their
recording. We have recently been able, somewhat fortuitously, to affirm the
"near infinity of memory" even for macaques. An animal that had had exposure
to a series of images for about 30 sec each was able, 6 or more months later, to
distinguish these from highly comparable images intermingled with the earlier
set (Ringo & Doty, 1985); this is similar to H.M., who can also perform such
recognition of images at a 6-month interval (Freed & Corkin, 1985), even though
missing hippocampus and amygdala.
The problem of mnemonic retrieval is indeed perplexing, be it such recogni-
tion or recall. For instance, mentally normal individuals suddenly afflicted with
transient global amnesia lose contact with a significant, and continuous, portion
of their past and are simultaneously unable to retain a record of present events:
they display both retrograde and anterograde amnesia (see, e.g., Kushner & Hau-
ser, 1985; Patten, 1971). The condition commonly clears within a matter of
hours thus dramatically evidencing the separation between established memo-
ries, which remain intact, and the mechanisms for forming and retrieving them.
Although the etiology of this bizarre confusion remains uncertain, speculation
and suggestive evidence focuses on the hippocampal area.
Such "bidirectional" amnesia may be permanent if there is massive loss of
limbic stuctures (Damasio, Eslinger, Damasio, Van Hoesen, & Cornell, 1985).
More interesting, however, is the situation in which it can be demonstrated that
retrieval per se is perturbed. This can be achieved to varying degrees by electrical
stimulation of the hippocampus or associated structures in rats (e.g., Collier,
Quirk, & Routtenberg, 1987), macaques (e.g., Overman & Doty, 1979), and
humans (e.g., Halgren, Wilson, & Stapleton, 1985), the latter proving specifically
an effect on retrieval per se.
One of the more fascinating and mysterious facts about retrieval of memories
is that following various forms of cerebral trauma the distant past may remain
 Time and Memory 151

available while more recent memories are inaccessible (e.g., Russell, 1959;
Squire & Cohen, 1979). There is thus some temporal gradient in vulnerability,
not of the memories but of their retrieval. This is again seen in humans with
electrical stimulation of the hippocampal area, the distance back from time pres-
ent whose retrieval is perturbed and the duration of the effect both being pro-
portional to the intensity of the stimulation (Bickford, Mulder, Dodge, Svien, &
Rome, 1958). In rare enigmatic cases access to a span of memory dating back
from time present as much as 40 years may be suddenly and permanently lost
(Andrews, Poser, & Kessler, 1982).
The remarkable thing in all these examples is the relative precision of the tem-
poral dividing line between retrievable and irretrievable memories, and the near
continuity of the inaccessible span. Penfield and Mathieson (1974), in their two
autopsied cases of temporal lobectomy, noted that the date of irretrievable mem-
ories extended farther back in time for the patient having the more posteriorly
extending hippocampal excision; they postulated that memories, or their mech-
anism of retrieval, might gradually migrate in time, with serial order preserved,
along the extent of the hippocampus. Given current knowledge, such a mecha-
nism is difficult to imagine, but the numerous instances of date-specific tempo-
rally continuous spans of irretrievability suggest a chronotopic mapping of times
past somewhere within the nervous system. Perhaps the hippocampus is a
microminiature quipu, continually knotting the present and stringing it anatom-
ically into the past.

MACAQUE, MAN, AND CEREBRAL DUALITY

One of the major riddles of memory, seldom posed, is how or whether the two
cerebral hemispheres are united in the mnemonic process: Are memories redun-
dant, stored in each hemisphere, or, as proposed by Doty, Negrao, and Yamaga
(1973), are they unilateral and available to the other hemisphere via the fore-
brain commissures? That one hemisphere via the forebrain commissures can
gain access to memories established exclusively in the other has recently been
affirmed (Kucharski & Hall, 1987). The relevance of these interhemispheric
questions to Johnson's presentation (Chapter 5, this volume) lies in two facts.
First, each hemisphere has somewhat different processing strategies, for exam-
ple, for visual information (see Zaidel, 1987). Thus the "functional forms" of
memory will have to differ even between the hemispheres.
The second issue concerns retrieval and "amnesia." Cronin-Golomb (1986)
demonstrated that even in the absence of the forebrain commissures one hemi-
sphere can gain a certain degree of access to information in the other via tenuous
clues as to the affective tone generated by this information and communicated
bilaterally. The situation is perhaps unique as a pure case of failure of retrieval,
the one hemisphere clearly knowing while the other lacks access to all details.
"Amnesia" is complete, a given hemisphere being unable to designate what is
152 FORMS OF MEMORY

or was viewed by the other, yet the ignorant hemisphere can garner clues from
the innumerably subtle shades of emotional tone generated by items that the
other hemisphere perceives.
Now, is not the amnesic patient preserving some appropriate emotional eval-
uation of a given face, while all details as to the origin of this emotional response
remain unavailable, just what Johnson describes? The point here is not that
amnesia ordinarily arises from failure of interhemispheric communication, but
that mechanisms, as Searle (1984) says, cannot be discerned from rules. Intro-
spection—be it Snarsky (see above) "explaining" forms of salivation in his dogs
or Johnson "noting," "shifting," "refreshing," and the like—describes the prob-
lem but does not usefully clarify it.
There is certainly a need for clarification, since amnesia does have many
forms. As noted, H.M. without amygdala and hippocampus can nevertheless
recognize visual material almost as well as normal macaques (Freed et al., 1987).
Yet even macaques with comparable neurological loss are able somehow not
only to master almost as well as normal animals the acquisition of 20 concurrent
visual discriminations when presented with one trial per day, but they "learn
how to learn" such discriminations as well (Murray, 1987). What, then, of the
vaunted role of the hippocampus (and amygdala?) in memory?
There are at least two intertwined factors here: the role of language and the
role of the hippocampus. When the former factor is essentially unavailable in
tagging visual images, mnemonic performance of macaques overlaps that of
humans (Doty, Ringo, & Lewine, 1988; Ringo, Lewine, & Doty, 1986). In com-
paring the mnemonic performance of patients such as H.M. and macaques with
similar neurological loss, it would seem essential that tests be used in which
human ability is at the same level as that of macaques—without the linguistic
advantage.
It is now abundantly clear that many things are remembered in the absence
of the hippocampus, yet is seems to serve some special role in making memory
traces available for conscious analysis, Johnson's "reflection" if you will. Per-
haps this is achieved by its role in "dating" each trace—preserving its temporal
sequence—so that it may subsequently be identified for retrieval. The amygdala,
in turn, might similarly provide an emotional "tag" to each trace. Some tanta-
lizingly suggestive evidence for the latter is provided by the experiments of
Meyer and his colleagues (see Meyer, 1984), where amnesia following electro-
convulsive shock occurs for motivationally rather than temporally related
"memories."
Again, an understanding of retrieval is essential to understanding memory.
Sternberg (1966) invented an ingenious means of measuring the time it takes to
retrieve a memory, admittedly one of recent vintage. He presented subjects with,
say, one to six target items, and subsequently required them to respond as rap-
idly as possible to distinguish whether a presented item (probe) was or was not
a member of the target set. The results were remarkably clear: there was a nearly
linear relation between number of target items and accrued time to respond.
Each item added roughly 30 msec to the reaction time, and since all other
 Time and Memory 153

times—perception, movement, and so on—should be independent of set size,

this 30 msec should be the time it takes to search and retrieve an additional item
from memory.
Sands and Wright (1982) were able to train a macaque to perform this task,
and they found a similar monotonic relation between reaction time and number
of target items. We repeated this, using the identical material on macaques and
humans (Lewine, unpublished data; Lewine, Doty, Provencal, & Astur, 1987).
While maintaining the same level of accuracy, intact macaques require only 5
to 11 msec per added target, compared with 17 to 31 msec for the human sub-
jects. In other words, macaques can search their memory three times faster than
can man. The most parsimonious explanation would seem to be simple neuro-
physiology: the macaque brain being only one tenth the volume of that in
humans, the relevant conduction paths and times must be significantly shorter.
The possibility does remain, however, that the human mnemonic search
includes a qualitatively different step, perhaps further verification prior to
decision.
This paradigm takes on added interest when the task can be apportioned selec-
tively between the two hemispheres (Lewine & Doty, work in progress). This is
made possible by transection of the optic chiasm. Three macaques have been
observed thus far, two with only the splenium of the corpus callosum and one
with only the anterior commissure remaining of the forebrain commissures. The
relation between reaction time and number of target items is unaltered in these
animals; it is about 6 msec per added item. The error rate is slightly higher (5-
10%) when the target information is given to one hemisphere and the probe
stimuli (target and nontarget items) are presented to the other for response than
when target and probe go to the same hemisphere. In other words, accuracy is
slightly degraded when performance depends on commissural as opposed to
intrahemispheric paths. However, there is no systematic difference in reaction
time for interhemispheric versus intrahemispheric performance.
When target items are apportioned between the hemispheres, the splenium is
wholly effective in unifying the mnemonic performance, so that, for instance, if
three target items are given to one hemisphere and three to the other, both the
error rate and the reaction time reflect the total of six items. However, with the
anterior commissure in such circumstance, while the error rate is the same as if
the six targets have all been given to one hemisphere, the reaction time is appro-
priate for the presence of only three items if an only if the probe presented
matches a target given the responding hemisphere: otherwise here too the time
reflects six items.
From these and other data it appears that unilateral input of a target forms a
bilateral engram, unifying the hemispheres completely in the case of the sple-
nium and nearly so for the anterior commissure. In each case the number of
errors is correlated with the total number of target items, regardless of how target
presentation and probing are distributed between the hemispheres. Nor is there
any indication of the delay that might be expected were it necessary to link a
probe presented to one hemisphere with target information stored in the other.
154 FORMS OF MEMORY

With the anterior commissure, however, it seems that one hemisphere can or
does distinguish targets which it has received from those given to the other. In
such a case, if each hemisphere is given three targets, the reaction time is appro-
priate for three if the probe is a member of the target set given to the hemisphere
viewing the probe, whereas if the probe is not a member of that set, the reaction
time is the same as if all six items were in a single hemisphere.
The absence of an added delay when target and probe are presented interhem-
ispherically rather than intrahemispherically is strong evidence that the target
information has been distributed bilaterally, especially with the splenium (and
contrary to the prediction of Doty et al., 1973). Although it is possible that the
time for interhemispheric transmission might be on the order of only a few mil-
liseconds, the known anatomy and physiology of callosal fibers strongly suggest
that the time should be much longer. Indeed, for humans, if the slower axons
were involved and a conduction path of 200 mm were traversed, a one-way
interhemispheric passage could easily run 200 msec or more (Swadlow,
Geschwind, & Waxman, 1979). Perhaps this is the origin of the inherent slow-
ness in the human search through the inventory of remembered targets, 20 to
30 msec per item versus 6 msec for macaques. Humans, with pronounced hemi-
spheric specialization, must verify the engrams in each hemisphere, whereas the
macaque, lacking such specialization, can rest content with the bilaterality of its
target engrams.

CONCLUSIONS

Time is relevant to memory not only as (1) the duration of the mnemonic
change, which may be permanent and lifelong, but also in (2) the temporal rela-
tion between two stimuli to form an association between them, (3) the relative
precision of the period elided in retrograde amnesia, however produced, and (4)
the time required to retrieve an item from memory. It is proposed that, in the
face of continual metabolic replacement, (1) is achieved by a transport of mate-
rial in and out of the cell nucleus to activate a particular genetic code, and that
this feature is relevant to the phenomena in (2) which, in terms of classical neu-
rophysiology, are temporally paradoxical.
It is also suggested that the hippocampus is the mammalian embodiment of
a phylogenetically ancient mechanism for recording the sequence of events, and
that the sequential "tag" that it provides is of prime importance in the process
of retrieval.
Sternberg's serial probe recognition task was used to study (4) above. It was
found that for recognition of a series of targets each additional item in the mne-
monic inventory added 17 to 31 msec to the reaction time for individual human
subjects, whereas for macaques the corresponding figures were 5 to 11 msec. In
other words, macaques can "search" their mnemonic store about three times
faster than can man. Judged by times accrued in the Sternberg paradigm in var-
ious experiments with "split-brain" macaques, it seems that the splenium of the
 Time and Memory 155

corpus callosum as well as the anterior commissure produce bilateral engrams

of the target items.

Acknowledgments
Preparation of these comments and original work reported here was supported
by a Javits Award, NS20052, from the National Institute of Neurological and
Communicative Disorders and Stroke, National Institutes of Health. I am par-
ticularly indebted to Jeffrey D. Lewine for permission to cite unpublished work
from his forthcoming doctoral dissertation, and to him and James L. Ringo for
consistently stimulating discussions of the phenomena of memory.

REFERENCES

Alkon, D. L. (1986). Changes of membrane currents and calcium-dependent phosphorylation

during associative learning. In D. L. Alkon & C. D. Woody (Eds.), Neural mecha-
nisms of conditioning (pp. 1-18). New York: Plenum.
Andrews, E., Poser, C. M., & Kessler, M. (1982). Retrograde amnesia for forty years. Cortex,
18, 441-458.
Auyeung, A., May, P. B. Y., Goh, J. W., & Sastry, B. R. (1987). Temporal requirements of
associative short-term potentiation in CA, neurons of rat hippocampus. Neurosci-
ence Letters, 79, 117-122.
Beritachvili, I. S. (1963). Les mecanismes nerveux de 1'orientation spatiale chez I'homme. Neu-
ropsychologia, I, 233-249.
Bickford, R. G., Mulder, D. W., Dodge, H. W. Jr., Svien, H. J., & Rome, H. P. (1958). Changes
in memory functions produced by electrical stimulation of the temporal lobe in man.
In H. C. Solomon, S. Cobb, & W. Penfield (Eds.), The brain and human behavior (pp.
227-243). (Research Publications Association for Research in Nervous and Mental
Disease, Vol. 30). Baltimore: Williams & Wilkins.
Cartwright, B. A., & Collett, T. S. (1983). Landmark learning in bees. Experiments and models.
Journal of Comparative Physiology-A, 151, 521-543.
Collier, T. J., Quirk, G. J., & Routtenberg, A. (1987). Separable roles of hippocampal granule
cells in forgetting and pyramidal cells in remembering spatial information. Brain
Research, 409, 316-328.
Corkin, S. (1984). Lasting consequences of bilateral medial temporal lobectomy: Clinical course
and experimental findings in H.M. Seminars in Neurology, 4, 249-259.
Cronin-Golomb, A. (1986). Subcortical transfer of cognitive information in subjects with com-
plete forebrain commissurotomy. Cortex, 22, 499-519.
Crowne, D. P., & Radcliffe, D. D. (1975). Some characteristics and functional relations of the
electrical activity of the primate hippocampus and a hypothesis of hippocampal func-
tion. In R. L. Isaacson & K. H. Pribram (Eds.), The hippocampus: Vol. 2. Neuro-
physiology and behavior (pp. 185-206). New York: Plenum.
Damasio, A. R., Eslinger, P. J., Damasio, H., Van Hoesen, G. W., & Cornell, S. (1985). Mul-
timodal amnesic syndrome following bilateral temporal and basal forebrain damage.
Archives of Neurology, 42, 252-259.
Doty, R. W. (1979). Neurons and memory: Some clues. In M. A. B. Brazier (Ed.), Brain mech-
anisms in memory and learning: From the single neuron to man (pp. 53-63). (Inter-
national Brain Research Organization Monograph Series, Vol. 4). New York: Raven.
Doty, R. W. (1987). W. Horsley Gantt: Physiologist, Pavlovian. In F. J. McGuigan &T. A. Ban
(Eds.), Critical issues in psychology, psychiatry, and physiology. A memorial to W.
Horsley Gantt (pp. 343-356). New York: Gordon & Breach.
156 FORMS OF MEMORY

Doty, R. W., & Giurgea, C. (1961). Conditioned reflexes established by coupling electrical exci-
tation of two cortical areas. In A. Fessard, R. W. Gerard, & J. Konorski (Eds.), Brain
mechanisms and learning (pp. 133-151). Oxford: Blackwell.
Doty, R. W., Negrao, N., & Yamaga, K. (1973). The unilateral engram. Act a Neurobiologiae
Experimental, 33, 711-728.
Doty, R. W., Ringo, J. L., & Lewine, J. D. (1988). Humanlike characteristics of visual mne-
monic system in macaques. In C. D. Woody, D. L. Alkon, & J. L. McGaugh (Eds.),
Cellular mechanisms of conditioning and behavioral plasticity (pp. 313-328). New
York: Plenum.
Dunlop, D. S. (1983). Protein turnover in brain. Synthesis and degradation. In A. Lajtha (Ed.),
Handbook of neurochemistry (2nd ed.): Vol. 5. Metabolic turnover in the nervous sys-
tem (pp. 25-63). New York: Plenum.
Freed, D. M., & Corkin, S. (1985). Rates of forgetting in H.M.: Six-month recognition. Society
for Neuroscience Abstracts, 11, 458.
Freed, D. M., Corkin, S., & Cohen, N. J. (1987). Forgetting in H.M.: A second look. Neuropsy-
chologia, 25, 461-471.
Guidotti, A., & Costa, E. (1977). Trans-synaptic regulation of tyrosine 3-mono-oxygenase bio-
synthesis in rat adrenal medulla. Biochemical Pharmacology, 26, 817-823.
Halgren, E., Wilson, C. L., & Stapleton, J. M. (1985). Human medial temporal-lobe stimulation
disrupts both formation and retrieval of recent memories. Brain and Cognition, 4,
287-295.
Halstead, W. C. (1951). Brain and intelligence. In L. A. Jeffress (Ed.), Cerebral mechanisms in
behavior. The Hixon Symposium (pp. 244-288). New York: Wiley.
Hawkins, R. D., Clark, G. A., & Kandel, E. R. (1987). Cell biological studies of learning in
simple vertebrate and invertebrate systems. In F. Plum (Ed.), Handbook of physiol-
ogy: Sec I. The nervous system: Vol. 5. Higher functions of the brain (pp. 25-83).
Bethesda, Md.: American Physiological Society.
Hober, R. (1945). Physical chemistry of cells and tissues. Philadelphia: Blakiston.
Hu, G.-Y., Hvalby, O., Walaas, S. I., Albert, K. A., Skjeflo, P., Andersen, P., & Greengard, P.
(1987). Protein kinase C injection into hippocampal pyramidal cells elicits features
of long term potentiation. Nature, 328, 426-429.
Isokawa-Akesson, M., Wilson, C. L., & Babb, T. L. (1987). Diversity of periodic pattern of firing
in human hippocampal neurons. Experimental Neurology, 98, 137-151.
Kesner, R. P., McDonough, J. J. Jr., & Doty, R. W. (1970). Diminished amnestic effect of a
second electroconvulsive seizure. Experimental Neurology, 27, 527-533.
Kucharski, D., & Hall, W. G. (1987). New routes to early memories. Science, 238, 786-788.
Kuffler, S. W. (1967). The Ferrier lecture: Neuroglial cells: Physiological properties and a potas-
sium mediated effect of neuronal activity on the glial membrane potential. Proceed-
ings of the Royal Society of London (Biology), 168, 1-21.
Kushner, M. J., &Hauser, W. A. (1985). Transient global amnesia: A case-control study. Annals
of Neurology, 18, 684-691.
Landolt, A. M. (1965). Elektronenmikroskopische Untersuchungen an der Perikaryenschicht
der Corpora pedunculata der Waldameise (Formica lugubris Zett.) mit besonderer
Berucksichtigung der Neuron-Glia-Beziehung. Zeitschrift fur Zellforschung, 66, 701-
736.
Larson, J., & Lynch, G. (1986). Induction of synaptic potentiation in hippocampus by pat-
terned stimulation involves two events. Science, 232, 985-988.
Levy, W. B., & Steward, O. (1983). Temporal contiguity requirements for long-term associative
potentiation/depression in the hippocampus. Neuroscience, 8, 791-797.
Lewine, J. D., Doty, R. W., Provencal, S., & Astur, R. (1987). Monkey beats man in efficiency
of mnemonic retrieval. Society for Neuroscience Abstracts, 13, 206.
Luria, A. R. (1968). The mind ofmnemonist. A little book about a vast memory (L. SolotaroiT,
Trans.). Chicago: Henry Regnery.
 Time and Memory 157

Lynch, G. (1986). Synapses, circuits, and the beginnings of memory. Cambridge, Mass.: MIT
Press.
Lynch, G., & Baudry, M. (1984). The biochemistry of memory: A new and specific hypothesis.
Science, 224, 1057-1063.
Margolis, R. K., & Margolis, R. U. (1983). Glycoproteins and proteoglycans. In A, Lajtha (Ed.),
Handbook of neurochemistry (2nd ed.): Vol. 5. Metabolic turnover in the nervous sys-
tem (pp. 177-204). New York: Plenum.
Menzel, R. (1983). Neurobiology of learning and memory: The honeybee as a model system.
Naturwissenschaften, 70, 504-511.
Meyer, D. R. (1984). The cerebral cortex: Its role in memory storage and remembering. Phys-
iological Psychology, 12, 81-88.
Mishkin, M. (1978). Memory in monkeys severely impaired by combined but not by separate
removal of amygdala and hippocampus. Nature, 273, 297-298.
Mishkin, M., Malamut, B., & Bachevalier, J. (1984). Memories and habits: Two neural systems.
In G. Lynch, J. L. McGaugh, & N. M. Weinberger (Eds.), Neurobiology of learning
and memory (pp. 65-77). New York: Guilford.
Montarolo, P. G., Goelet, P., Castellucci, V. F., Morgan, J., Kandel, E. R., & Schacher, S.
(1986). A critical period for macromolecular synthesis in long-term heterosynaptic
facilitation in Aplysia. Science, 234, 1249-1254.
Morris, R. G. M. (1983). An attempt to dissociate "spatial-mapping" and "working-memory"
theories of hippocampal function. In W. Seifert (Ed.), Neurobiology of the hippocam-
pus (pp. 405-432). London: Academic Press.
Muller, R. U., Kubie, J. L., & Ranck, J. B. Jr. (1987). Spatial firing patterns of hippocampal
complex-spike cells in a fixed environment. Journal of Neuroscience, 7, 1935-1950.
Murray, E. A. (1987). Normal learning set formation in monkeys with combined ablation of
the amygdaloid complex and hippocampal formation. Society for Neuroscience
Abstracts, 13, 206.
Murray, E. A., & Mishkin, M. (1983). Severe tactual memory deficits in monkeys after com-
bined removal of the amygdala and hippocampus. Brain Research, 270, 340-344.
Nathan, P. W., & Smith, M. C. (1950). Normal mentality associated with a maldeveloped
"rhinencephalon." Journal of Neurology, Neurosurgery and Psychiatry, 13, 191-197.
O'Keefe, J. (1983). Spatial memory within and without the hippocampal system. In W. Seifert
(Ed.), Neurobiology of the hippocampus (pp. 375-403). London: Academic Press.
Olton, D. S. (1983). Memory functions and the hippocampus. In W. Seifert (Ed.), Neurobiology
of the hippocampus (pp. 335-373). London: Academic Press.
Overman, W. H. Jr., & Doty, R. W. (1979). Disturbance of delayed match-to-sampie in
macaques by tetanization of anterior commissure versus limbic system or basal gan-
glia. Experimental Brain Research, 37, 511-524.
Patten, B. M. (1971). Transient global amnesia syndrome. Journal of the American Medical
Association, 217, 690-691.
Pawlow, J. P. (1904). Psychische Erregung der Speicheldriisen. Ergebnisse der Physiologic, biol-
ogische Chemie und experimented Pharmakologie, 3, 177-193.
Penfield, W., & Mathieson, G. (1974). Memory; autopsy findings and comments on the role of
hippocampus in experiential recall. Archives of Neurology, 31, 145-154.
Ringo, J. L., & Doty, R. W. (1985). A macaque remembers pictures briefly viewed six months
earlier. Behavioural Brain Research, 18, 289-294.
Ringo, J. L., Lewine, J. D., & Doty, R. W. (1986). Comparable performance by man and
macaque on memory for pictures. Neuropsychologia, 24, 711-717.
Rosene, D. L., & Van Hoesen, G. W. (1987). The hippocampal formation of the primate brain:
A review of some comparative aspects of cytoarchitecture and connections. In E. G.
Jones & A. Peters (Eds.), Cerebral cortex (Vol. 6, pp. 345-456). New York: Plenum.
Russell, W. R. (1959). Brain, memory and learning: A neurologist's view. Oxford: Clarendon
Press.
158 FORMS OF MEMORY

Sands, S. F., & Wright, A. A. (1982). Monkey and human pictorial memory scanning. Science,
216, 1333-1334.
Searle, J. (1984). Minds, brains and science. Cambridge, Mass.: Harvard University Press.
Squire, L. R., & Cohen, N. (1979). Memory and amnesia: Resistance to disruption develops
for years after learning. Behavioral and Neural Biology, 25, 115-125.
Squire, L. R., & Cohen, N. J. (1984). Human memory and amnesia. In G. Lynch, J. L.
McGaugh, & N. M. Weinberger (Eds.), Neurobiology of learning and memory (pp. 3-
64). New York: Plenum.
Sternberg, S. (1966). High-speed scanning in human memory. Science, 153, 652-654.
Swadlow, H. A., Geschwind, N., & Waxman, S. G. (1979). Commissural transmission in
humans. Science, 204, 530-531.
Thoenen, H., Otten, U., & Schwab, M. (1979). Orthograde and retrograde signals for the reg-
ulation of neuronal gene expression: The peripheral sympathetic nervous system as
a model. In F. O. Schmitt & F. G. Worden (Eds.), The neurosciences, fourth study
program (pp. 911-928). Cambridge, Mass.: MIT Press.
Thomas, G. J. (1984). Memory: Time binding in organisms. In L. R. Squire & N. Butters (Eds.),
Neuropsychology of memory (pp. 374-384). New York: Guilford.
Vanderwolf, C. H., & Leung, L.-W. S. (1983). Hippocampal rhythmical slow activity: A brief
history and the effects of entorhinal lesions and phencyclidine. In W. Seifert (Ed.),
Neurobiology of the hippocampus (pp. 275-302). London: Academic Press.
Vertes, R. P. (1986). A new role for FTG neurons? Behavioral and Brain Sciences, 9, 425-426.
Wehner, R. (1981). Spatial vision in arthropods. In H. Autrum (Ed.), Handbook of sensory
physiology: Vol. VII/6C. comparative physiology and evolution of vision in inverte-
brates, C: Invertebrate visual centers and behavior II (pp. 287-616). Berlin: Springer-
Verlag.
Zaidel, D. W. (1987). Hemispheric asymmetry in memory for pictorial semantics in normal
subjects. Neurospsychologia, 25, 487-495.
Zola-Morgan, S., Squire, L. R., & Amaral, D. G. (1986). Human amnesia and the medial tem-
poral region: Enduring memory impairment following a bilateral lesion limited to
field CA, of the hippocampus. Journal of Neuroscience, 6, 2950-2967.
8
Forms of Memory: Issues and Directions
A R T H U R P. S H I M A M U R A

What constitutes evidence for different forms of memory? In behavioral neu-

roscience, memory forms are often distinguished by the kind of test paradigm
that is used—from nonassociative learning paradigms such as habituation and
sensitization to verbal learning paradigms such as word recall and recognition.
A critical issue in understanding the neurobiological basis of memory is to iden-
tify and characterize the neural architecture that supports such behavioral
expressions of memory. In this chapter I evaluate evidence that has been used
to argue for different forms of memory, particularly in the domain of neuropsy-
chological research.

TWO EXTREME VIEWS OF MEMORY REPRESENTATION

We can approach the idea of forms of memory from two extreme views. First,
one could argue that there is only one, undifferentiated neural system for mem-
ory. This view is similar to that proposed by Lashley (1950). That is, the neural
substrate of memory is so distributed and so interactive that one cannot locate
any subcomponents or functional units within the system. Thus any damage
within the system causes some level of memory impairment, and every part of
this system is equally potent in producing impairment. An extreme version of
this view predicts that performance on every test of memory should decrease to
some extent as a result of damage to the brain.
It is easy to reject this extreme view on the basis of human neuropsychological
data. Memory functions do not degrade in a nonspecific manner as a result of
brain injury. That is, not all areas of the brain contribute equally to memory—
there is some differentiation and specialization of memory function. In partic-
ular, damage to the medial temporal or diencephalic region produces a selective
impairment of memory. Moreover, studies of memory in various animal models
(see Chapters 2, 3, and 4, this volume) suggest that it is possible to dissociate the
neural architecture underlying one form of memory from another.
The opposite extreme view is that all memory representations are unique, so

159
160 FORMS OF MEMORY

that divisions of memory into two, three, or more forms are arbitrary and erro-
neous. In some sense, all memory representations should be differentiated from
one another, otherwise they would produce exactly the same behavioral conse-
quence. Yet some divisions in memory may be especially honored by the brain.
That is, some memory forms may share similar yet not identical neural sub-
strates, and others may be quite distinct. In this sense, it would be useful to
develop classes or divisions of memory forms, just as it is useful to classify the
multitude of animal species into a taxonomy. Even though all animal species
are considered to be different to some extent, it has been useful to identify
classes, orders, and families of species.

DIVISIONS OF MEMORY AND BEHAVIORAL DISSOCIATIONS

Many divisions of memory have been proposed, and some of them are listed in
Table 8.1. These classifications are not mutually exclusive, and in fact there is
much overlap between some of them. A common way to establish evidence for
divisions of memory is to identify dissociations between two putatively different
forms of memory. Figure 8.1 illustrates the ideal data set for establishing a single
dissociation. Performance by a control group indicates normal performance on
two tests (tests A and B). Group 1 could represent a group that has a damaged
neural system. This group could be a group of experimentally lesioned animals,
a group of neurologically impaired patients, or a group given a pharmacological
agent. For example, Morris (Chapter 3, this volume) showed that rats are
impaired on a spatial memory test (test A) but not on a visual discrimination
test (test B) when they are injected with D-2-amino-5-phosphonovalerate (AP-
5), a pharmacological agent that blocks yV-methyl-D-aspartate (NMDA) receptor

TABLE 8.1 Divisions of Memory

Fact memory Skill memory
Declarative Procedural
Memory Habit
Explicit Implicit
Knowing that Knowing how
Cognitive mediation Semantic
Conscious recollection Skills
Elaboration Integration
Memory with record Memory without record
Autobiographical memory Perceptual memory
Representational memory Disposilional memory
Vertical association Horizontal association
Locale Taxon
Episodic Semantic
Working Reference
Source: From Squire (1987).
 Forms of Memory: Issues and Directions 161

FIGURE 8.1. Idealized data for

establishing a single dissociation.
Group 1 exhibits impaired perform-
ance on test A but entirely normal
performance on test B. Group 1
represents an impaired group, such
as a group of experimentally le-
sioned animals, a group of neuro-
psychological patients, or a group
given a pharmacological agent.

channels (see Foster & Fagg, 1984; Jahr & Stevens, 1987). This finding suggests
that the behavioral impairment resulting from AP-5 is rather selective.
To demonstrate single dissociations, three important criteria must be satis-
fied. First, one must show entirely normal performance on a test—as shown by
the performance on test B in Figure 8.1. If even a relatively small but significant
impairment is observed in test B, a pure dissociation is not supported, because
it must be allowed that at least some aspect of performance on test B depends
on the neural system that is damaged in group 1. Second, tests A and B must be
sensitive tests. In particular, equivalent performance should not be attributed to
ceiling or floor effects. For example, an extremely easy or extremely hard test B
could result in a spurious finding of no significant difference between groups.
Third, it would be ideal to observe entirely normal performance in test B under
a variety of manipulations, not just one condition, as is shown in Figure 8.1. For
example, normal performance of group 1 might be observed on test B at various
retention intervals along a forgetting curve or for different doses of a drug.
Studies of amnesia have shown that certain memory functions can be disso-
ciated from other cognitive and memory functions. The central feature of
organic amnesia is an impairment of new learning capacity (i.e., anterograde
amnesia) that affects performance on a variety of standard memory tests, such
as tests of free recall, cued recall, recognition, and paired-associate learning (for
review, see Cermak, 1982; Shimamura, in press; Squire, 1987; Warrington &
Weiskrantz, 1982). Severe anterograde amnesia can occur despite relatively
intact intellectual and language functions. Importantly, recent findings suggest
that certain memory forms are expressed in an entirely normal fashion in amne-
sic patients (see Cohen, 1984; Schacter, 1987a; Shimamura, 1986; Squire, 1987).
Such dissociations follow the pattern shown in Figure 8.1 and satisfy all three
criteria mentioned above.
One form of memory that is preserved in amnesia is priming: the unconscious
facilitation or bias of recently presented information (for review, see Shima-
mura, 1986). Priming can be demonstrated in a word association priming task
in which study words (e.g., baby) are cued with semantically related words (e.g.,
child). In one study (Shimamura & Squire, 1984), patients with Korsakoff s syn-
162 FORMS OF MEMORY

drome and alcoholic control subjects were presented a list of words but were not
told to expect a memory test. Following list presentation, subjects were told,
"Now I want to give you a different task. I will say a word and I want you to say
the first word that comes to mind." Patients with Korsakoff s syndrome used
previously presented words in the word association priming test as often as con-
trol subjects, despite severe impairment on a test of explicit recall memory (Fig.
8.2). In fact, both groups exhibited about a twofold increase in their use of study
words, as compared to their baseline tendency to use these words when no prior
presentation is given. When priming was tested after a 2-hr retention interval,
word association performance in both groups was at baseline levels. Examples
of entirely normal priming in amnesia can be found on a variety of other
implicit tests (see Schacter, 1987a; Shimamura, 1986).
Skill learning is another form of memory that can be expressed by amnesic
patients in an entirely normal manner. Pursuit rotor learning (Corkin, 1968) and
mirror reading (Cohen & Squire, 1980) are two examples. What distinguishes
tests of skills and priming from standard memory tests such as tests of recall and

FIGURE 8.2. An example of a single dissociation in amnesia. Compared with alcoholic control
subjects (ALC), patients with Korsakoff s syndrome (KOR) exhibited impaired performance on a
free recall test but entirely normal performance on a word association priming test (data from
Shimamura and Squire, 1984). In the word association test, subjects were presented words that
were related to study words and simply asked to say the first word that came to mind.
 Forms of Memory: Issues and Directions 163

recognition is that skills and priming can be tested without any explicit reference
to the fact that memory is being tested. In tests of skills, patients can exhibit
normal retention across days and even months, with no knowledge of ever hav-
ing practiced the skill before. In priming tests, subjects are asked to say the first
word that comes to mind and study words simply appear to pop to mind.
These findings have suggested that amnesic patients exhibit impairment
whenever they are asked explicitly to declare or recollect previously presented
information. Thus it has been suggested that amnesic patients exhibit a partic-
ular deficit in declarative, conscious, or explicit memory (see Baddeley, 1982;
Schacter, 1987a; Squire, 1982b, 1987). Figure 8.3 illustrates this division of
memory. Memory is divided into declarative and nondeclarative forms. Declar-
ative memory includes episodic (i.e., autobiographical) memory and semantic
(i.e., fact) memory (see Tulving, 1983). In addition, other divisions could be
made within the declarative memory domain. For example, divisions could be
made between recall and recognition memory in the sense that these two forms
may require different kinds of processing, as Johnson (Chapter 5, this volume)
suggested. Declarative memory deficits are found in any test where memory is
explicitly requested. This deficit is found not only on standard tests such as word
recall and recognition, but it also can be seen on tests of fact memory (Shima-
mura & Squire, 1987). In fact memory tests, subjects are presented obscure facts
(e.g., Angel Falls is located in Venezuela) and then are asked explicitly to rec-
ollect the answer (e.g., Where is Angel Falls located?). A deficit occurs even when
the question is presented as a test of general knowledge rather than making ref-
erence to an earlier learning session.
Nondeclarative memory comprises a heterogeneous group of memory func-
tions, all of which can operate without the neural systems damaged in amnesia.
It includes procedural memory or skill learning, simple classical conditioning,
priming, and other forms of memory such as preceptual aftereffects. In the past,

FIGURE 8.3. Forms of declarative and nondeclarative memory. Declarative memory depends
on the medial temporal and diencephalic areas known to be damaged in amnesic patients. Non-
declarative memory comprises a heterogeneous group of memory forms; it can be expressed im-
plicitly, that is, without awareness that memory is being tested.
164 FORMS OF MEMORY

the term "procedural memory" was used to classify all these spared functions.
Yet because so little is known about the similarities and differences among these
spared functions, it is not clear whether the phrase "procedural memory" can
be appropriately applied to all these nondeclarative forms. Thus the more neu-
tral classification of "nondeclarative memory" is used here.
Declarative memory functions are better understood. In particular, they
depend critically on the diencephalic and medial temporal regions. Neuropath-
ological studies of patients with KorsakofFs syndrome (Mair, Warrington, &
Weiskrantz, 1979; Victor, Adams, & Collins, 1971) have implicated the dien-
cephalic region as important for the establishment of declarative memory. Ever
since the now-classic studies of patient H. M. (see Corkin, 1984; Scoville & Mil-
ner, 1957), we have known that damage to the medial temporal region—includ-
ing the hippocampal gyrus, amygdala, and hippocampus proper—is important
for new learning ability. Recently the importance of the hippocampus was dem-
onstrated by amnesic case R.B., who in a postmortem study by Zola-Morgan,
Squire, and Amaral (1986) was found to have bilateral lesions restricted to the
CA1 subfield of the hippocampus.
Figure 8.4 illustrates why the medial temporal region may be so important for
memory. It depicts the fact that unimodal and polymodal information from
many areas of neocortex projects to entorhinal cortex either directly, as indi-
cated by projections from neocortical areas such as orbitofrontal cortex, cingu-
late cortex, and superior temporal gyrus, or indirectly from other neocortical
areas (e.g., Brodmann areas 9 and 23) that influence the entorhinal cortex
through their projections to the parahippocampal gyrus or perirhinal cortex. The
entorhinal cortex projects to the hippocampus via the perforant path. In addi-
tion, there are reciprocal projections back to cortical areas from the hippocam-
pus via the subiculum and entorhinal cortex. Thus the medial temporal region—
particularly the hippocampus—is privy to extensive cortical activity from a
variety of areas. One possibility is that the hippocampus acts to integrate, asso-
ciate, or index cortical events (for further discussion see Rolls, Chapter 9, this
volume; Squire, Shimamura, & Amaral, 1989; Teyler & DiScenna, 1986).
It is presumed that the establishment of declarative memory depends criti-
cally on the medial temporal region. That is, when the medial temporal region
is damaged, performance on memory tests that tap conscious or explicit mem-
ory should be impaired. Evidence suggests that within this domain, the severity
of amnesia depends on the extent of damage. For example, compared with case
R.B., H.M.'s medial temporal damage is more extensive and his amnesia is
more severe. Similarly, damage to areas outside the medial temporal and dien-
cephalic regions, such as the frontal cortex, may cause additional memory
impairment. Additional damage could account for disproportionate deficits seen
on two different declarative memory tests such as recall and recognition. That
is, patients with more widespread damage may exhibit cognitive and memory
impairment not seen in other amnesic patients with more circumscribed
damage.
FIGURE 8.4. Many cortical areas project to entorhinal cortex either directly as shown by the
projections at the bottom (e.g., orbitofrontal cortex, cingulate cortex) or indirectly by way of the
parahippocampal gyms or perirhinal cortex (e.g., Brodmann areas shown above). The entorhinal
cortex projects to the hippocampus and also sends reciprocal projections back to cortical areas.
(From Squire, Shimamura, and Amaral, 1989)
166 FORMS OF MEMORY

PARTIAL DISSOCIATIONS: TRUE OR SPURIOUS?

Disproportionate deficits on two declarative memory tests could resemble the

data set illustrated in Figure 8.5. I will refer to such a pattern of results as a
partial dissociation. In this example, group 1 exhibits impairment on tests A and
B relative to control subjects, yet performance on test A is disproportionately
low. This pattern of results is not uncommon (e.g., recall vs. recognition). In
many comparisons between two tests of memory, amnesic patients appear to be
more impaired on one test than another. A simple explanation is that one of the
memory tests is a more sensitive measure than the other. Other scaling artifacts
can also produce spurious dissociations (see Loftus, 1978). Yet another possi-
bility is that amnesic patients in fact exhibit a qualitatively different (i.e., a dis-
proportionate) deficit on one memory test compared to another. That is, they
may indeed have more difficulty with the function measured by one test than
with the function measured by the other.
To distinguish between these possibilities, one needs to know what would
happen to control subjects if their memory were weakened so that performance
on one of the memory tests was matched to the performance of amnesic
patients. Memory could be weakened in control subjects by testing after a longer
retention interval or by presenting information at faster exposure times (see
Meudell & Mayes, 1982). Two possible outcomes that may occur with an addi-
tional group of "weakened" control subjects are shown in Figure 8.6. One pos-
sibility is that as memory weakens, performance on test A declines faster than
performance on test B so that weakened control subjects actually perform like
the amnesic patients on both tests. Figure 8.6B illustrates this outcome, which
could occur as a result of differences in test sensitivity or differences in scaling
effects between tests. For example, it was once thought that recognition memory
performance was disproportionately impaired relative to cued recall perfor-
mance (Warrington & Weiskrantz, 1974; see Mayes & Meudell, 1981). Yet when
weakened control subjects were used, it was found that this apparent dissocia-
tion was spurious (Huppert & Piercy, 1978; Mayes & Meudell, 1981; Squire,
Wetzel, & Slater, 1978). These findings suggest that some "disproportionate"

FIGURE 8.5. Partial dissocia-

tions occur when performance by
an impaired group (group 1) is dis-
proportionately impaired on one
test (test A) relative to another (test
B). Note, however, that a complete
dissociation does not exist because
entirely normal performance is not
observed on test B.
 Forms of Memory: Issues and Directions 167

FIGURE 8.6. To determine if partial dissociations are true or spurious, an additional (weakened)
control group is used to assess performance in control subjects who are matched to the impaired
group on one test (test B). Performance can be weakened by using a longer retention interval. (A)
A true dissociation occurs because even when performance is weakened in control subjects, group
1 still exhibits a disproportionate impairment on test A. (B) A spurious dissociation occurs
when weakened control subjects exhibit a decline in performance on tests in the same manner
as group 1.

effects may simply be a consequence of weak memory. That is, some differences
between groups could be explained as a result of quantitative differences rather
than qualitative differences between tests (for another example, see Shimamura
& Squire, 1988).
Some phenomena in amnesia cannot be explained solely on the basis of weak
memory. Figure 8.6A illustrates a situation in which the data suggest a true sin-
168 FORMS OF MEMORY

gle dissociation. Amnesic patients (i.e., group 1) and weakened control subjects
exhibited the same performance level on test B, yet amnesic patients exhibited
a disproportionate deficit on test A. Such findings suggest a true (qualitative)
dissociation between tests A and B. The deficit in free recall compared with rec-
ognition memory is one example (e.g., Hirst et al., 1986). Other examples
include the disproportionate impairment in metamemory (the ability to know
about one's own memory abilities) in some amnesic patients and the presence
of source amnesia (a memory loss for where and when information was learned)
in some patients (see Schacter, Harbluk, & McLachlan, 1984; Shimamura &
Squire, 1986, 1987). Initially it was difficult to determine whether these phenom-
ena were true or spurious dissociations. Upon further studies with weakened
control subjects it was possible to conclude that a true dissociation had been
observed.
The extent and severity of neurological damage in amnesic patients can vary
greatly. One possibility is that partial dissociations reflect damage that occurs
outside the medial temporal and diencephalic areas. One candidate structure
that could contribute to memory impairment is the frontal lobes. Indeed,
patients with KorsakofFs syndrome have cortical atrophy in frontal areas, as
indicated by quantitative analyses of computed tomography (CT) scans (Shi-
mamura, Jernigan, & Squire, 1988). Moreover, patients with KorsakofFs syn-
drome exhibit cognitive and memory deficits that are not seen in other amnesic
patients (see Moscovitch, 1982; Oscar-Berman, 1980; Squire, 1982a; Talland,
1965). For example, Korsakoff patients exhibit particular deficits in the ability
to remember the temporal order of events (Huppert & Piercy, 1978; Squire,
1982a) and in metamemory (Shimamura & Squire, 1986). Other amnesic
patients do not exhibit these memory deficits. These deficits may be related to
deficits in planning and in spatial/temporal memory which are characteristic of
patients with frontal lobe damage (for review, see Milner & Petrides, 1984;
Schacter, 1987b).
Perhaps the best neuropsychological method for addressing these issues would
be to compare amnesic patients directly with patients with frontal lobe damage.
By using two neurologically impaired groups, it may be possible to locate double
dissociations as shown in Figure 8.7. For instance, group 1 could represent a
group of patients with circumscribed medial temporal damage who show mem-
ory impairment on test A but not on test B. Conversely, group 2 could represent
a group of patients with frontal damage who show the reverse pattern. Findings
of double dissociations between groups would greatly strengthen our under-
standing of the role of various neural systems on memory processing. In fact, it
may be that patients with KorsakofFs syndrome would constitute a third group
in which performance is impaired on both tests A and B, because of their com-
bined diencephalic and frontal damage.
Recent findings suggest that patients with frontal lobe damage exhibit impair-
ment on some tests of memory (Janowsky, Shimamura, Kritchevsky, & Squire,
1989). A group of seven patients with relatively circumscribed frontal lobe
lesions was studied. Two patients had left frontal damage, three had right frontal
 Forms of Memory: Issues and Directions 169

FIGURE 8.7. Double dissociations provide strong behavioral evidence for the independence of
performance on two tests. One impaired group exhibits normal performance on test B but impaired
performance on test A, whereas another impaired group exhibits the reverse pattern.

damage, and two had bilateral damage. We found reliable memory dysfunction
in these patients, though they do not come close to exhibiting the amnesic dis-
order seen in patients with diencephalic or medial temporal damage. For exam-
ple, preliminary findings suggest that some aspects of metamemory are impaired
in these frontal lobe patients (Janowsky, Shimamura, & Squire, 1989).
Patients with frontal damage also appear to have deficits on tests of free recall,
despite relatively preserved performance on tests of recognition memory. Inter-
estingly, relative to anoxic or ischemic patients, patients with KorsakofFs syn-
drome also appear to have a disproportionate deficit on free recall tests relative
to performance on recognition memory tests. These findings provide evidence
for the contribution of frontal lobes to performance on some memory tests. Free
recall deficits may be linked to the planning and spatial/temporal memory def-
icits that are known to occur as a result of frontal lobe damage. Such deficits
could affect retrieval processes or strategies which may be more important on
tests of free recall.

FORMS OF MEMORY

Figure 8.8 summarizes in simple terms some of the points made here about dif-
ferent forms of memory. The figure shows what might be called three forms of
memory. Each form depends on the interconnections of a number of brain struc-
tures or systems, which are identified by circles. Each is different, yet each shares
some structures with others. Suppose structure B represents the hippocampus.
Then by definition, memory forms 1 and 2 are two forms of declarative mem-
ory. For example, memory forms 1 and 2 could represent the processes that con-
tribute to recall and recognition memory, respectively. Memory form 3 could
170 FORMS OF MEMORY

FIGURE 8.8. Three forms of memory.

represent neural structures that contribute to priming. If structure B is damaged,

then memory forms 1 and 2 should be affected (i.e., a declarative memory
impairment). In contrast, performance on tests that tap memory form 3 should
be entirely normal. In other words, single dissociations should be observed
between memory form 3 and the other two memory forms (1 and 2).
Suppose further that structure C represents the frontal lobes such that memory
form 1 depends critically on both the hippocampus and frontal lobes (structures
B and C). Memory form 2 depends on the hippocampus but not on the frontal
lobes. If structures B and C are damaged, then form 1 should be disproportion-
ately affected on tests that involve both structures B and C. Yet both forms 1
and 2 should be affected on tests that involve structure B. Consequently, a par-
tial dissociation would occur between memory forms 1 and 2, because both
forms would be impaired on declarative tests, but on some tests form 1 would
be more impaired than form 2. This scheme could help to explain dissociations
between various forms of declarative memory, such as the dissociation between
recall and recognition memory.
 Forms of Memory: Issues and Directions 171

CONCLUSIONS

Evidence of different forms of memory as indicated by neuropsychological stud-

ies of amnesia can provide important clues about the organization of memory.
Yet a complete understanding of memory cannot be achieved by such studies
alone. It is extremely important to characterize the neural systems more explic-
itly using animal models—from invertebrate preparations to nonhuman pri-
mate models. Such models are important for providing information about basic
principles of memory. Also, models in the nonhuman primate (see Mishkin,
1982; Squire & Zola-Morgan, 1983) may provide better information at the neu-
ral systems level because with such models one can gain more precise control of
the structures that when damaged cause amnesia. Finally, the field seems ripe
for computational models of memory. In particular, the advent of connection-
istic or parallel distributed processing approaches (see Byrne & Berry, 1989;
Rumelhart & McClelland, 1986) is a great boon for theories of memory because
they can bring together information from neuroscience, neuropsychology, and
cognitive science.

Acknowledgments
This work was supported by the Medical Research Service of the Veterans
Administration, by National Institute of Mental Health Grant MH24600, and
by the Office of Naval Research. I thank Larry Squire and Jeri Janowsky for
helpful comments on earlier drafts.

REFERENCES

Baddeley, A. D. (1982). Domains of recollection. Psychological Review, 89, 708-729.

Byrne, J. H., & Berry, W. O. (Eds.). (1989). Neural models of plasticity. Orlando, Fla.: Academic
Press.
Cermak, L. S. (1982). Human memory and amnesia. Hillsdale, N.J.: Erlbaum.
Cohen, N. J. (1984). Preserved learning capacity in amnesia: Evidence for multiple memory
systems. In L. R. Squire & N. Butters (Eds.), Neuropsychology of memory (pp. 83-
103). New York: Guilford.
Cohen, N. J., & Squire, L. R. (1980). Preserved learning and retention of pattern analyzing skill
in amnesia: Association of knowing how and knowing that. Science, 210, 207-209.
Corkin, S. (1968). Acquisition of motor skill after bilateral medial temporal lobe excision. Neu
ropsychologia, 6, 225-265.
Corkin, S. (1984). Lasting consequences of bilateral medial temporal lobectomy: Clinical course
and experimental findings in H.M. Seminars in Neurology, 4, 249-259.
Foster, A. C., & Fagg, G. E. (1984). Acidic amino acid binding sites in mammalian neuronal
membranes: Their characteristics and relationship to synapttc receptors. Brain
Research Review, 7, 103-164.
Hirst, W., Johnson, M. K., Kim, J. K., Phelps, E. A., Risse, G., & Volpe, B. T. (1986). Recog-
nition and recall in amnesics. Journal of Experimental Psychology: Learning, Mem-
ory, and Cognition, 12, 445-451,
172 FORMS OF MEMORY

Huppert, F. A., & Piercy, M. (1978). The role of trace strength in recency and frequency judge-
ments by amnesic and control subjects. Quarterly Journal of Experimental Psychol-
ogy, 30, 346-354.
Jahr, C. E., & Stevens, C. F. (1987). Glutamate activates multiple single channel conductances
in hippocampal neurons. Nature, 325, 522-525.
Janowsky, J. S., Shimamura, A. P., Kritchevsky, M., & Squire, L. R. (1989). Cognitive impair-
ment following frontal lobe damage and its relevance to human amnesia. Behavioral
Neuroscience, 103, 548-560.
Janowsky, J. S., Shimamura, A. P., & Squire, L. R. (1989). Memory and metamemory: Com-
parisons between patients with frontal lobe lesions and amnesic patients. Psychobi-
ology, 17, 3-11.
Lashley, K. S. (1950). In search of the engram. Symposia of the Society of Experimental Biology,
4, 454-482.
Loftus, G. (1978). On the interpretations of interactions. Memory and Cognition, 6, 312-319.
Mair, W. G. P., Warrington, E. K., & Weiskrantz, L. (1979). Memory disorder in Korsakoff
psychosis: A neuropathological and neuropsychological investigation of two cases.
Brain, 102, 749-783.
Mayes, A. R., & Meudell, P. R. (1981). How similar is the effect of cueing in amnesics and
normal subjects following forgetting? Cortex, 171, 113-124.
Meudell, P. R., & Mayes, A. R. (1982). Normal and abnormal forgetting: Some comments on
the human amnesic syndrome. In L. A. Ellis (Ed.), Normality and pathology in cog-
nition function (pp. 203-237). London: Academic Press.
Milner, B., & Petrides. M. (1984). Behavioural effects of frontal-lobe lesions in man. Trends in
Neuroscience, 7, 403-407.
Mishkin, M. (1982). A memory system in the monkey. Philosophical Transactions of the Royal
Society (London), 298, 85-95.
Moscovitch, M. (1982). Multiple dissociations of function in amnesia. In L. Cermak (Ed.),
Human memory and amnesia (pp. 337-370). Hillsdale, N.J.: Erlbaum.
Oscar-Berman, M. (1980). The neuropsychological consequences of long-term chronic alcohol-
ism. American Scientist, 68, 410-419.
Rumelhart, D. E., & McClelland, J. L. (1986). Parallel distributed processing. Cambridge,
Mass.: MIT Press.
Schacter, D. L. (1987a). Implicit memory: History and current status. Journal of Experimental
Psychology: Learning, Memory and Cognition, 13, 501-518.
Schacter, D. L. (1987b). Memory, amnesia, and frontal lobe dysfunction. Psychobiology, 15,
21-36.
Schacter, D. L., Harbluk, J. L., & McLachlan, D. R. (1984). Retrieval without recollection: An
experimental analysis of source amnesia. Journal of Verbal Learning and Verbal
Behavior, 23, 593-611.
Scovillc, W. B., & Milner, B. (1957). Loss of recent memory after bilateral hippocampal lesions.
Journal of Neurology, Neurosurgery, and Psychiatry, 20, 11-21.
Shimamura, A. P. (1986). Priming effects in amnesia: Evidence for a dissociable memory func-
tion. Quarterly Journal of Experimental Psychology, 38A, 619-644.
Shimamura, A. P. (in press). Disorders of memory: The cognitive science perspective. In F.
Boiler & J. Grafman (Eds.), Handbook of clinical neuropsychology. Amsterdam:
Elsevier.
Shimamura, A. P., Jernigan, T. L., & Squire, L. R. (1988). Radiological (CT) findings in patients
with KorsakofTs syndrome and their relationship to memory impairment. Journal
of Neuroscience, 8, 4400-4410.
Shimamura, A. P., & Squire, L. R. (1984). Paired-associate learning and priming effects in
amnesia: A neuropsychological study. Journal of Experimental Psychology: General,
113, 556-570.
 Forms of Memory: Issues and Directions 173

Shimamura, A. P., & Squire, L. R. (1986). Memory and metamemory: A study of the feeling-
of-knowing phenomenon in amnesic patients. Journal of Experimental Psychology:
Learning, Memory and Cognition, 12, 452-460.
Shimamura, A. P., & Squire, L. R. (1987). A neuropsychological study of fact memory and
source amnesia. Journal of Experimental Psychology: Learning, Memory and Cog-
nition, 13, 464-473.
Shimamura, A. P., & Squire, L. R. (1988). Long-term memory in amnesia: Cued recall, recog-
nition memory, and confidence ratings. Journal of Experimental Psychology: Learn-
ing, Memory and Cognition, 14, 763-770.
Squire, L. R. (1982a). Comparisions between forms of amnesia: Some deficits are unique to
Korsakoffs syndrome. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 8, 560-571.
Squire, L. R. (1982b). The neuropsychology of human memory. Annual Review of Neurosci-
ence, 5, 241-273.
Squire, L. R. (1987). Memory and brain. New York: Oxford University Press.
Squire, L. R., Shimamura, A. P., & Amaral, D. (1989). Memory and the hippocampus. In J. H.
Byrne & W. O. Berry (Eds.), Neural models of plasticity (pp. 208-239). Orlando, Fla.:
Academic Press.
Squire, L. R., Wetzel, C. D., & Slater, P.C. (1978). Anterograde amnesia following ECT: An
analysis of the beneficial effect of partial information. Neuropsvchologia, 116, 339-
347.
Squire, L. R., & Zola-Morgan, S. (1983). The neurology of memory: The case for correspon-
dence between the findings for human and nonhuman primates. In J. A. Deutsch
(Ed.), The physiological basis of memory (2nd ed.). New York: Academic Press.
Talland, G. A. (1965). Deranged memory. New York: Academic Press.
Teyler, T. J., & DiScenna, P. (1986). The hippocampal memory indexing theory. Behavioral
Neuroscience, 100, 147-154.
Tulving, E. (1983). Elements of episodic memory. Oxford: Clarendon Press.
Victor, M., Adams, R. D., & Collins, G. H. (1971). The Wernicke-Korsakoff syndrome. Phila-
delphia: Davis.
Warrington, E. K., & Weiskrantz, L. (1974). The effect of prior learning on subsequent retention
in amnesic patients. Neuropsychologia, 12, 419-428.
Warrington, E. K., & Weiskrantz, L. (1982). Amnesia: A disconnection syndrome? Neuropsy-
chologia, 20, 233-248.
Zola-Morgan, S., Squire, L. R., & Amaral, D. G. (1986). Human amnesia and the medial tem-
poral region: Enduring memory impairment following a bilateral lesion limited to
field CA1 of the hippocampus. Journal of Neuroscience, 6, 2950-2967.
This page intentionally left blank
II
Regulation of Cortical Function
in Memory
This page intentionally left blank
Introduction
M A R K R. R O S E N Z W E I G

A major theme that runs through this section is differential changes and differ-
ential activities in the nervous system in relation to learning and memory, and
often in relation to specific kinds of learning and memory. These changes can
occur during learning and formation of memory and during development and
self-organization of the nervous system; morever, differential changes in the
brain caused by diseases lead to predictable changes in specific abilities to learn
and remember.
Let us look at a few examples of these differential effects from the following
chapters. Rolls (Chapter 9) takes up specific kinds of learning known to involve
the primate hippocampus. He reports that about one tenth of the neurons in the
hippocampus come to respond differentially to the combination of a particular
object with a particular location in space; about one seventh of the neurons in
the hippocampus come to respond differentially to the rewarded combination of
a particular object with a particular motor response. Moreover, it appears that
although many hippocampal neurons originally respond to a new combination,
the few that show sustained responses to it inhibit others that show only tran-
sient modification, so that relatively few neurons are allocated to a particular
association. In turn, these hippocampal neurons may allocate memory storage
to specific regions of cerebral cortex for efficient use of cortical capacity. Subse-
quent use of this stored material may produce progressively more specific rep-
resentation. Thus all along the way, differential responsiveness is stressed.
Van Hoesen (Chapter 11) considers whether Alzheimer's disease (AD) dissects
memory into distinct behavioral compartments and whether such compart-
ments are related to damage in distinct locations in the brain. He points out that
neither the course of the memory changes in AD nor the sequence of changes in
the brain have been studied fully by rigorous methods; Van Hoesen nevertheless
produces evidence for regularity in time courses and for relationships between
the behavioral and the neuroanatomical changes. Here are a few examples. The
earliest changes in AD may be deficiencies in anterograde episodic memories;
this may be related to the fact that the most persistent and perhaps earliest neu-
roanatomical changes occur in the hippocampus, particularly in its subicular

177
178 REGULATION OF CORTICAL FUNCTION IN MEMORY

subdivision, and in the anterior parahippocampal gyrus including the entorhinal

cortex. Impairments of retrograde episodic memories also occur relatively early
in AD, and they often cause the victim or the family to seek medical advice.
These losses appear to be related to pathology in posterior parahippocampal,
lateral temporal, and anterior insular cortices. Still later, generic and semantic
memory decline when the multimodal association cortices are involved. Even
when a region of cortex is affected in AD, pathology is not uniformly distributed
but is confined to specific cell types and to specific cortical layers. Specific sub-
cortical formations and specific subregions within some of them are also
involved in AD. In fact, although this section of the book deals with "Regulation
of Cortical Function in Memory," the authors frequently treat the cortex in the
context of its connections to other parts of the brain.
Singer (Chapter 10) considers some of the mechanisms of development of self-
organization of the brain, especially the organization of binocularity in the
visual system, and investigates the extent to which they may be precursors of
mechansims that mediate learning in the adult. Singer points out that neural
activity, including the results of information processing, becomes part of the
environment relevant for self-organization, and the spiral of reciprocal interac-
tions leads to ever more differentiated structures. At least early in life, both for-
mation and withdrawal of synaptic connections are involved, and, especially in
adult life, changes in the efficacy of synaptic connections are involved. Record-
ing from individual cortical neurons, Singer and his colleagues studied changes
during monocular visual deprivation and during reversed monocular occlusion.
They conclude that two different processes are involved: rapid inactivation of
deprived connections and slow increase in effectiveness of previously inacti-
vated connections. (Chapter 15 by Cooper, Bear, Ebner, and Scofield suggests
that the changes in visual dominance occur chiefly at excitatory synapses on
excitatory cells, a type of synapse that is formed exclusively on dendritic spines,
a feature that distinguishes it from other types of cortical synapse—again a
highly specific site.) Singer reports that changes in ocular dominance occur only
when the animal is alert and using its visual system, so postsynaptic modulatory
inputs to the plastic cells are also important. Permissive gating signals may
involve both cholingergic and noradrenergic modulatory systems and also Ca2+
fluxes governed by NMDA receptors, so a variety of neurochemical processes is
involved.
A less overt theme in this section is the overcoming of certain older ideas
about the nervous system, learning, and memory as new ideas are presented and
demonstrated—as well as the revival of other old ideas. One dogma that had
wide currency in the fifties was that the adult nervous system is quite stable and
could not be altered in any measurable way by such benign activities as learning
and formation of memory. In light of some of the reports and discussions in this
section, it may be difficult for many readers to realize how fixed the adult ner-
vous system seemed then. But I remember clearly how exciting it was for
Edward Bennett, David Krech, and me when our data forced us to conclude that
giving formal training or informal experience to experimental animals caused
 Introduction 179

measurable changes in the biochemistry of their brains and that these changes
were not caused by associated variables (Krech, Rosenzweig, & Bennett, 1960).
Soon thereafter, along with Marian Diamond, we were startled to find that small
but significant changes in gross anatomy of the cerebral cortex were induced by
differential experience in enriched or impoverished environments (Rosenzweig,
Krech, Bennett, & Diamond, 1962).
A further belief that gave way before empirical evidence was the idea that the
brain, or at least the cerebral cortex, responded in a rather general, uniform way
to training or differential experience. Our early experiments demonstrated that
the cerebral cortex responded differently to experience than did noncortical tis-
sue, and within the cerebral cortex different regions responded with reliably
larger or smaller changes (Rosenzweig et al., 1962). Going to a finer level of res-
olution, we and others soon found that neurons show a variety of morphological
responses to differential experience, using measures such as dendritic branching
(e.g., Holloway, 1966; Volkmar & Greenough, 1972), numbers of synapses (e.g.,
Diamond, Lindner, Johnson, Bennett, & Rosenzweig, 1975; West & Greenough,
1972) and numbers of dendritic spines (Globus, Rosenzweig, Bennett, & Dia-
mond, 1973). These morphological changes were accompanied by an increase in
protein in the cortical regions involved (Bennett, Diamond, Krech, & Rosen-
zweig, 1964). Complementary experiments demonstrated that administering
agents that inhibit protein synthesis in the brain shortly before either appetitive
or aversive training prevents formation of long-term memory, although it does
not prevent acquisition of short-term or intermediate-term storage (e.g., Bar-
ondes & Cohen, 1968; Flood, Bennett, Orme, & Rosenzweig, 1975; Flood, Ben-
nett, Rosenzweig, & Orme, 1972, 1973). As pointed out at the First Irvine Con-
ference, the synthesis of proteins required for formation of long-term memory
usually occurs in the minutes following training, but "this synthesis can be
deferred during a period of inhibition of protein synthesis ... the stronger the
training, the longer the synthesis can be deferred but still take place" (Rosen-
zweig & Bennett, 1984, p. 268).
Can differential experience cause changes in the numbers of cortical synapses?
Morris (Chapter 3) raised this question earlier in this book, and it is also perti-
nent here. Although Singer doubts that the visual cortex can show changes in
synaptic number in the adult, several laboratories have found changes in syn-
aptic numbers in occipital cortex of the rat as a result of differential experience
(e.g., Bhide & Bede, 1984; Diamond et al., 1975; Globus et al., 1973; Sirevaag &
Greenough, 1986), and Colonnier and Beaulieu (1985) found similar results in
the cat. Wenzel et al. (1977) also found increases in synaptic numbers in the
hippocampus of the rat as a result of formal training. Stewart, Rose, Kind, Gab-
bott, and Bourne (1984) found that peck-avoidance training in chicks caused a
significant increase in the number of synapses per unit volume of neuropil in
the left cerebral hemisphere. These findings in vertebrates are now being
repeated in invertebrates. Recently investigators studying the mechansims of
long-term sensitization in Aplysia showed that the number of synapses on sen-
sory neurons is significantly increased after training, and the number then
180 REGULATION OF CORTICAL FUNCTION IN MEMORY

declines over a few weeks in parallel with decreasing retention (Bailey & Chen,
1987). Nevertheless, opinion has fluctuated on this question ever since Tanzi
(1893) proposed that the plastic changes in learning would be found at the junc-
tions between neurons. I remember Eccles exclaiming at a conference, "Learning
can produce only bigger and better synapses, not new ones." (He wrote much
the same thing, but not quite so pithily [Eccles, 1973, p. 184].) On the contrary,
Eccles's teacher Sherrington was an early proponent of the hypothesis that learn-
ing could result in formation of new synaptic connections. In the chapter in
which he proposed the name "synapse," Sherrington offered this portrayal of
educable neurons:
Shut off from all opportunities of reproducing itself and adding to its number by
mitosis or otherwise, the nerve cell directs its pent-up energy towards amplifying its
connections with its fellows, in response to the events which stir up. Hence, it is
capable of an education unknown to other tissues. (Foster & Sherrington, 1897, 1117)
I believe that Eccles was wrong on two counts when he asserted that learning
involves only increasing the size of existing synapses. First, as noted, results
from several laboratories show that training or enriched experience does cause
the formation of new synapses. Second, it appears that enriched experience may
cause some synapses to become smaller while others become larger, as we
pointed out some time ago (Rosenzweig, Diamond, Bennett, & Mollgaard,
1972). The latter point is reminiscent of the findings on the inhibitory as well as
excitatory changes during learning reported by Rolls and of the concept of prun-
ing connections through experience as mentioned by Singer.
To claim that formation of new synapses (and elimination of existing synap-
ses) is employed in some kinds of memory storage is not to hold that such
changes occur in all memory storage. Some forms of memory (especially short-
term) are very unlikely to require changes in synaptic number. Other kinds may
very well employ this mechanism, as shown by the evidence cited earlier. Kety
(1976) put the case for plurality of memory mechansims well:
So profound and powerful an adaptation as learning or memory is not apt to rest
upon a single modality. Rather, I suspect that advantage is taken of every opportunity
provided by evolution. There were forms of memory before organisms developed
nervous systems, and after that remarkable leap forward it is likely that every new
pathway and neural complexity, every new neurotransmitter, hormone, or metabolic
processes that played upon the nervous sytem and subserved a learning process was
preserved and incorporated, (pp. 321-322)
Another old idea, now refuted, was that the two cerebral hemispheres of ani-
mals play essentially similar and equal roles in learning and memory formation.
It is not so long ago that many investigators supposed that only the human being
employs the two hemispheres in differential fashion, but now rodents and chicks
show clear hemispheric differences in ways that we are still far from understand-
ing (e.g., Horn, 1985; Patterson, Alvarado, Warner, Bennett, & Rosenzweig,
1986; Rose & Csillag, 1985; Stewart et al., 1984). In particular, Patterson et al.
(1986) found that to form memories of peck-aversion learning in chicks, the left
 Introduction 181

hemisphere is the one that counts in regard to one brain structure (the inter-
mediate medial hyperstriatum ventrale) but the right hemisphere is the impor-
tant one for another structure (the lateral neostriatum).
An old idea that keeps coming back is Tanzi's general concept that the same
mechanisms may be involved in the development of the nervous system and in
learning and memory storage. A strong statement of this hypothesis was made
by Galambos (1961) at an international conference on brain mechanisms and
learning in 1959:
It could be argued ... that no important gap separates the explanations for how the
nervous sytem comes to be organized during embryological development in the first
place; for how it operates to produce the innate responses characteristic of each spe-
cies in the second place; and for how it comes to be reorganized, finally, as a result
of experiences during life. If this idea should be correct, the solution of any one of
these problems would mean that the answers to the others would drop like a ripe
plum, so to speak, into our outstretched hands, (pp. 238-239)

Although this hypothesis has intrigued many investigators, it is only in recent

years that substantial empirical support has been obtained for it, as exemplified
by Singer's chapter in this book and many other current papers and symposia
(e.g., Black & Greenough, 1986; Purves, Smith, Bailey, & Greenough, 1987).
It is worth noting that the kinds of learning and memory considered in this
section, as well as in Part I, represent a considerable variety. This may seem
quite natural these days, even though the discussions in this book show that we
are far from agreeing about the categories of learning and memory or the defi-
nitions of various proposed categories. But in the 1940s, when behavioral
"schools" of learning theory held sway, it was generally believed that all learning
is basically similar and that one set of laws should be able to encompass all
instances of learning. It was startling then for Edward Tolman to write an article
entitled "There Is More Than One Kind of Learning" (1949), and he probably
did not convince many readers. Considerably later, Neal Miller (1967) wrote
about "Grade-A Certified Learning" (associative learning in mammals) as if that
was the kind to be studied. Miller admitted that there are other phenomena such
as habituation or even immune reactions "that may have some elements in com-
mon with learning and that may provide the key to an understanding of its phys-
ical basis.... In our present state of ignorance we cannot afford to discard any
approaches" (p. 643). That almost exclusive focus on associative learning is cer-
tainly not the view of the authors in this volume. Carew, Marcus, Nolen, Rank-
ing, and Stopfer (Chapter 2), for example, reported important results on the
developmental emergence of habituation, dishabituation, and sensitization. The
authors in Part II specify a number of different kinds of learning and memory,
including certain types involving spatial location that require the hippocampus,
anterograde and retrograde impairments in recall, episodic and semantic mem-
ories, working and long-term memories. In fact, when Singer draws up similar-
ities between developmental self-organization and "adult learning," one won-
ders what kind or kinds of adult learning he has in mind.
182 REGULATION OF CORTICAL FUNCTION IN MEMORY

So the following chapters consider various kinds of learning and memory and
a variety of possible neurophysiological, neuroanatomical, and neurochemical
correlates and mechansims. All this implies to me that much progress is being
made in getting beyond oversimple generalities; investigators are beginning to
reach the specifics of the processes involved in the many kinds of learning and
memory that human beings and other animals demonstrate and experience. To
celebrate these differential approaches and findings, let me conclude the intro-
duction to the following chapters by expanding a famous aphorism: "Vivent les
differences!"

REFERENCES

Bailey, C. H., & Chen M. (1987). The course of the structural changes at identified sensory
neuron synapses during long-term sensitization in Aplysia. Society for Neuroscience
Abstracts, 13 (Part 1), 617.
Barondes, S. H., & Cohen, H. D. (1968). Memory impairment after subcutaneous injection of
acetoxycycloheximide. Science, 160, 556-557.
Bennett, E. L., Diamond, M. C., Krech, D., & Rosenzweig, M. R. (1964). Chemical and ana-
tomical plasticity of brain. Science, 146, 610-619.
Bhide, P. G., & Bede, K. S. (1984). The effects of a lengthy period of environmental diversity
on well-fed and previously undernourished rats. II. Synapse-to-neuron ratios. Jour-
nal of Comparative Neurology, 227, 305-310.
Black, J. E., & Greenough, W. T. (1986). Developmental approaches to the memory process.
In J. L. Martinez, Jr., 7 R. P. Kesner (Eds.), Learning and memory: A biological view
(pp. 55-81). Orlando, Fla.: Academic Press.
Colonnier, M., & Beaulieu, C. (1985). The differential effect of impoverished and enriched envi-
ronments on the number of "round asymmetrical" and "flat symmetrical" synapses
in the visual cortex of cat. Society for Neuroscience Abstracts, 11, 226.
Diamond, M. C, Lindner, B., Johnson, R., Bennett, E. L., & Rosenzweig, M. R. (1975). Dif-
ferences in occipital cortex synapses from environmentally enriched, impoverished,
and standard colony rats. Journal of Neuroscience Research, I, 109-119.
Eccles, J. C. (1973). The understanding of the brain. New York: McGraw-Hill.
Flood, J. F., Bennett, E. L., Orme, A. E., & Rosenzweig, M. R. (1975). Effects of protein syn-
thesis inhibition on memory for active avoidance training. Physiology and Behavior,
14, 177-184.
Flood, J. F., Bennett, E. L., Rosenzweig, M. R., & Orme, A. E. (1972). Influence of training
strength on amnesia induced by pretraining injections of cycloheximide. Physiology
and Behavior, 9, 589-600.
Flood, J. F., Bennett, E. L., Rosenzweig, M. R., & Orme, A. E. (1973). The influence of duration
of protein synthesis inhibition on memory, Physiology and Behavior, 10, 555-562.
Foster, M., & Sherrington, C. S. (1897). A textbook of physiology: Part 3. The central nervous
system. New York: Macmillan.
Galambos, R. (1961). Changing concepts of the learning mechanism. In A. Fessard, R. W.
Gerard, & J. Konorski (Eds.), Brain mechanisms and learning (pp. 231-241). Spring-
field, 111.: Charles C. Thomas.
Globus, A., Rosenzweig, M. R., Bennett, E. L., & Diamond, M. C. (1973). Effects of differential
experience on dendritic spine counts in rat cerebral cortex. Journal of Comparative
and Physiological Psychology, 82, 175-181.
Holloway, R. L. (1966). Dendritic branching: Some preliminary results of training and testing
in rat visual cortex. Brain Research, 2, 393-396.
 Introduction 183

Horn, G. (1985). Memory, imprinting and the brain. Oxford: Clarendon Press.
Kety, S. (1976). Biological concomitants of affective states and their possible role in memory
processes. In M. R. Rosenzweig and E. L. Bennett (Eds.), Neural mechanisms of
learning and memory (pp. 321-326). Cambridge, Mass.: MIT Press.
Krech, D., Rosenzweig, M. R., & Bennett, E. L. (1960). Effects of environmental complexity
and training on brain chemistry. Journal of Comparative and Physiological Psychol-
ogy, 53, 509-519.
Miller, N. E. (1967). Certain facts of learning relevant to the search for its physical basis. In G.
C. Quarton, T. Melnechuk, & F. O. Schmitt (Eds.), The neurosciences (pp. 643-653).
New York: Rockefeller University Press.
Patterson, T. A., Alvarado, M. C., Warner, I. T., Bennett, E. L., & Rosenzweig, M. R. (1986).
Memory stages and brain asymmetry in chick learning. Behavioral Neuroscience,
100, 856-865.
Purves, D., Smith, S., Bailey, C. H., & Greenough, W. T. (1987). Symposium: Dynamic changes
of synaptic structure under normal and experimental conditions. Society for Neuro-
science Abstracts, 13 (Part 2), 1001.
Rose, S. P. R., & Csillag, A. (1985). Passive avoidance training results in lasting changes in
deoxyglucose metabolism in left hemisphere regions of chick brain. Behavioral and
Neural Biology, 44, 315-324.
Rosenzweig, M. R. & Bennett, E. L. (1984). Basic processes and modulatory influences in the
stages of memory formation. In G. Lynch, J. L. McGaugh, & N. M. Weinberger
(Eds.), Neurobiology of learning and memory (pp. 263-288). New York: Guilford.
Rosenzweig, M. R., Diamond, M. C., Bennett, E. L. & Mollgaard, K. (1972). Negative as well
as positive synaptic changes may store memory. Psychological Review, 79, 93-96.
Rosenzweig, M. R., Krech, D., Bennett, E. L., & Diamond, M. C. (1962). Effects of environ-
mental complexity and training on brain chemistry and anatomy: A replication and
extension. Journal of Comparative and Physiological Psychology, 55, 429-437.
Sirevaag, A. M., & Greenough, W. T. (1986). Multivariate analyses of morphological measure-
ments can distinguish among rats reared in complex, social or individual environ-
ments. Society for Neuroscience Abstracts, 12, 1284.
Stewart, M. G., Rose, S. P. R., Kind, T. S., Gabbott, P. L. A., & Bourne, R. (1984). Hemispheric
asymmetry of synapses in chick medial hyperstriatum ventrale following passive
avoidance training: A stereological investigation. Developmental Brain Research, 12,
261-269.
Tanzi, E. (1893). I fatti e le induzioni nell'odierna istologia del sistema nervoso. Revista Sper-
imentale di Freniatria e di Medicina Lega/e, 19, 419-472.
Tolman, E. C. (1949). There is more than one kind of learning. Psychological Review, 56, 144-
155.
Volkmar, F. R. & Greenough, W. T. (1972). Rearing complexity affects branching of dendrites
in the visual cortex of the rat. Science, 176, 1447-1449.
Wenzel, J., Kammerer, E., Joschko, R., Joschko, M., Kaufmann, W., Kirsche, W., & Matthies,
H. (1977). Der Einfluss eines Lernexperimentes auf die Synaptenzahl in Hippocam-
pus der Ratte. Elekronenmikroskpische und morphometrische Untersuchungen.
Zeitschrift fur mikroskopische-anatomische Forschung, 91, 57-73.
West, R. W. & Greenough, W. T. (1972). Effects of environmental complexity on cortical syn-
apses of rats: Preliminary results. Behavioral Biology, 7, 279-283.
9
Functions of Neuronal Networks
In the Hippocampus
and of Backprojections
in the Cerebral Cortex in Memory
E D M U N D T. R O L L S

FUNCTIONS OF THE PRIMATE HIPPOCAMPUS IN MEMORY

It is known that damage to certain regions of the temporal lobe in humans pro-
duces anterograde amnesia, manifested as a major deficit following the damage
in learning to recognize new stimuli (Milner, 1972; Scoville & Milner, 1957;
Squire, 1986, 1987). The anterograde amnesia has been attributed to damage to
the hippocampus, which is within the temporal lobe, and to its associated path-
ways such as the fornix (Gaffan, 1974, 1977; Milner, 1972; Scoville & Milner,
1957; Zola-Morgan, Squire, & Amaral, 1986), but this has been questioned, and
it has been suggested that damage to both the hippocampus and the amygdala
is crucial in producing anterograde amnesia, in that combined but not separate
damage to the hippocampus and amygdala produced severe difficulty with
visual and tactual recognition tasks in the monkey (Mishkin, 1978, 1982; Mur-
ray & Mishkin, 1984). In investigations of the particular aspects of memory for
which the hippocampus may be essential, it has been shown that monkeys with
damage to the hippocampo-fornical system have a learning deficit on memory
tasks that requires them to make associations between a stimulus, for example,
a picture, and a spatial motor response such as touching one part of a screen
(Gaffan, 1985; Rupniak & Gaffan, 1987); they are also impaired on memory
tasks that require combinations of stimulus attributes with their locations in
space to be processed together, such as memory not only for which object was
shown but where it was shown (Gaffan & Saunders, 1985). Further, humans with
right temporal lobe damage are also impaired in conditional spatial response
and object-place memory tasks (Petrides, 1985; Smith & Milner, 1981).
To analyze the functions being performed by the hippocampus in memory,

184
 Functions of Hippocampal and Cortical Networks 185

the activity of 1510 single hippocampal neurons was recorded in rhesus mon-
keys learning and performing these memory tasks known to be impaired by
damage to the hippocampus or fornix (Cahusac, Miyashita, & Rolls, 1989;
Miyashita, Rolls, Cahusac, & Niki, 1989; Rolls et al, 1989).
In an object-place memory task in which the monkey had to remember not
only which object had been seen in the previous 7 to 15 trials, but also the posi-
tion in which it had appeared on a video monitor, neurons were found which
responded differentially depending on where on the monitor screen the objects
were shown (Rolls et al., 1989). These neurons composed 9.3% of the population
recorded. It is notable that these neurons responded to particular postions in
space (whereas "place" cells in the rat respond when the rat is in a particular
place; O'Keefe, 1984). In addition, 2.4% of neurons responded more to a stim-
ulus the first time it was shown in a particular position than the second time.
These neurons thus responded to a combination of information about the stim-
ulus being shown and about position in space, for only by responding to a com-
bination of this information could the neurons respond when a stimulus was
shown for the first time in a certain position in space.
In tasks in which the monkeys had to acquire associations between visual
stimuli and spatial responses, 14.2% of the neurons responded to particular com-
binations of stimuli and responses (Miyashita et al., 1989). For example, in a
task in which the monkey had to perform one response (touching a screen three
times) when one visual stimulus was shown, but had to perform a withholding
response for 3 sec to obtain reward when a different stimulus was shown (Gaffan,
1985), 9.2% of the neurons responded to one of the stimuli if it was linked to
one of the responses in this task. The same neurons typically did not respond if
the same stimuli or the same responses were used in different tasks, or if other
stimuli were associated with the same responses in this task. Thus these neurons
responded to a combination of a particular stimulus with a particular spatial
motor response (Miyashita et al., 1989).
It was possible to study the activity of 70 neurons in the hippocampus or para-
hippocampal gyrus while the monkeys learned new associations between visual
stimuli and spatial responses. In some cases it was possible to show that the
activity of these neurons became modified during this learning (Cahusac, Fei-
genbaum, Rolls, Miyashita, & Niki, 1986). Interestingly, 22% of the neurons
acquired a sustained differential response during the learning, and 45% of the
neurons differentiated between the stimuli only at or just before the monkey
learned the task and stopped differentiating after 5 to 10 more trials. This is
consistent with the possibility (discussed later) that the neurons which show
large sustained differential responses inhibit the other neurons which show tran-
sient modification, so that as a result of competition not all neurons are allocated
to one stimulus—spatial response association.
These results show that hippocampal neurons in the primate have responses
related to certain types of memory. One type of memory involves complex con-
junctions of environmental information, for example, when information about
position in space (perhaps reflecting information from the parietal cortex) must
A
FIGURE 9.!. (A) Representation of connections within the hippocampus. Inputs reach the hippocampus through the perforant path, which makes synapses
with the dendrites of the dentate granule ceils and also with the apical dendrites of the CAS pyramidal cells. The dentate granule cells project via the mossy
fibers to the CA3 pyramidal cells. The well-developed recurrent collateral system of the CAS cells is indicated. The CAS pyramidal cells project via the
Schaffer collaterals to the CA1 pyramidal cells, which in turn have connections to the subiculum. (B) Schematic representation of the connections of the
hippocampus, showing also that the cerebral cortex (neocortex) is connected to the hippocampus via the parahippocampal gyrus and entorhinal cortex, and
that the hippocampus projects back to the neocortex via the subiculum, entorhinal cortex, and parahippocampal gyrus.
188 REGULATION OF CORTICAL FUNCTION IN MEMORY

be memorized in conjunction with what that object is (perhaps reflecting infor-

mation from the temporal lobe visual areas) so that where a particular object
was seen in space can be remembered. The hippocampus is ideally placed ana-
tomically for detecting such conjuntions in that it receives highly processed
information from association areas such as the parietal cortex (conveying infor-
mation about position in space), the inferior temporal visual cortex (conveying
a visual specification of an object), and the superior temporal cortex (conveying
an auditory specification of a stimulus) (Van Hoesen, 1982; see Fig. 9.5). The
positions of stimuli in space may be represented by the firing of hippocampal
neurons as described above so that conjunctions of, for example, objects and
their position can be formed. It is also suggested that one neurophysiological
mechanism by which "place" cells in the rat (see O'Keefe, 1984) may be formed
is by conjunction learning of sets of simultaneously occurring stimuli in different
parts of space, each set of which defines a place.

COMPUTATIONAL THEORY OF THE HIPPOCAMPUS

A possible theoretical basis for these results, and in particular how the hippo-
campus may perform the conjunction or combination learning just described, is
now considered. Figure 9.1 A is schematic of the connections of the hippocam-
pus. One feature is that there is a sequence of stages, in each of which there is a
major set of input axons that connect by a form of matrix with the output neu-
rons of that stage. The type of computation that could be performed by one of
these stages is considered first.
The perforant path connections with the dentate granule cells may be taken
as an example. A version of this is represented as a simplified matrix in Figure
9.2. Although the perforant path makes one set of synapses with the output neu-
rons in the form of a matrix, the matrix is clearly very different from an asso-
ciation matrix memory in that in the hippocampal system there is no uncondi-
tioned stimulus that forces the output neurons to fire (see Rolls, 1987). Nor is
there a climbing fiber for each output cell that acts as a teacher as in the cere-
bellum (see Ito, 1984). In the hippocampal circuit there is apparently no teacher;
that is, this appears to be an example of an unsupervised learning system. One
mode of operation for such a network is described next, and properties of the
hippocampus which suggest that it may operate in this way are discussed later.
Consider a matrix memory of the form shown in Figure 9.2 in which the
strengths of the synapses between horizontal axons and the vertical dendriles
are initially random (postulate 1). Because of these random initial synaptic
weights, different input patterns on the horizontal axons will tend to activate
different output neurons (in this case, granule cells). The tendency for each pat-
tern to select or activate different neurons can then be enhanced by providing
mutual inhibition between the output neurons, to prevent too many neurons
responding to that stimulus (postulate 2). This competitive interaction can be
viewed as enhancing the differences in the firing rates of the output cells (cf. the
 Functions of Hippocampal and Cortical Networks 189

FIGURE 9.2. A matrix for competitive learning in which the input stimuli are presented along
the rows of the input axons (ar), which make modifiable synapses (s rc ) with the dendrites of the
output neurons, which form the columns (dc) of the matrix.

contrast enhancement described by Grossberg, 1982). Synaptic modification

then occurs according to the rules of long-term potentiation in the hippocampus;
that is, synapses between active afferent axons and strongly activated postsyn-
aptic neurons increase in strength (see Kelso, Ganong, & Brown, 1986; Levy,
1985; McNaughton, 1984; Wigstrom, Gustaffson, Huang, & Abraham, 1986)
(postulate 3). The effect of this modification is that the next time the same stim-
ulus is repeated, the neuron responds more (because of the strengthened syn-
apses), more inhibition of other neurons occurs, and then further modification
produces even greater selectivity. The response of the system as a classifier thus
climbs over repeated iterations. One effect of this observed in simulations is that
a few neurons then obtain such strong synaptic weights that almost any stimulus
which has any input to that neuron will succeed in activating it. The solution to
this is to limit the total synaptic weight that each output (postsynaptic) neuron
can receive (postulate 4). In simulations this is performed by normalizing the
sum of the synaptic weights on each neuron to a constant (e.g., 1.0) (see Rumel-
hart & Zipser, 1986; von der Malsburg, 1973). This has the effect of distributing
the output neurons evenly between the different input patterns received by the
network. In the brain this may be achieved by using a modified Hebb rule which
produces some decrease in synaptic strength if there is high presynaptic but low
postsynaptic activation (see Bear, Cooper, & Ebner, 1987).
A simulation of the operation of such a competitive neuronal network is
reproduced as Figure 9.3. It is shown that the network effectively selects different
output neurons to respond to different combinations of active input horizontal
lines. It thus performs a type of classification in which different complex input
patterns are encoded economically onto a few output lines. It should be noted
 Functions ofHippocampal and Cortical Networks 191

that this classification finds natural clusters in the input events; orthogonalizes
the classes so that overlap in input events can become coded onto output neu-
rons with less overlap and in that many active input lines may be coded onto
few active output lines; and does not allocate neurons to events that never occur
(see Grossberg, 1982, 1987; Marr, 1970, 1971; Rumelhart & Zipser, 1986). It
may be noted that there is no special correspondence between the input pattern
and the output lines selected. Thus it is not useful for any associative mapping
between an input and an output event and is thus different from associative
matrix memories (Rolls, 1987). Instead, this type of matrix finds associations or
correlations between input events (which are expressed as sets of simultaneously
active horizontal input lines or axons), allocates output neurons to reflect the
complex event, and stores the required association between the input lines onto
the few output neurons activated by each complex input event.
There is evidence that in the hippocampus the synapses between inactive
axons and active output neurons become weaker (see Bear et al., 1987; Levy,
1985; McNaughton, 1984). In the learning system described this would facilitate
accurate and rapid classification, in that weakening synapses onto a postsynaptic
neuron from axons that are not active when it is strongly activated would reduce
the probability that it will respond to a stimulus which must be placed into a
different class. It is also of interest that postulate 4 above (that the total synaptic
strength onto a postsynaptic neuron is somewhat fixed) is not physiologically
unreasonable (Levy & Desmond, 1985).
Another feature of hippocampal circuitry is the mossy fiber system, which
connects the granule cells of the dentate gyrus to the CA3 pyramidal cells of the

FIGURE 9.3. Simulation of learning in a competitive matrix memory. The architecture is as

shown in Figure 9.2, except that there are 64 horizontal axons and 64 vertical dendrites, which
form the rows and columns respectively of the 64 x 64 matrix of synpases. The strength of each
synapse, which was initially random, is indicated by the darkness of each pixel. The activity of
each of the 64 input axons is represented in the 64-element vector at the left of the diagram by the
darkness of each pixel. The output firing of the vertical neurons is represented in the same way by
the output vectors at the bottom of the diagram. The upper output vector is the result of multiplying
the input stimulus through the matrix of synaptic weights. The vector resulting from the application
of competition between the output neurons (which produces contrast enhancement between the
elements or neurons of the vector) is shown below by the vector labeled "after filter." The state
of the matrix is shown after 203 cycles in each of which stimuli with 8 of 64 active axons were
presented, and the matrix was allowed to learn as described in the text. The stimuli, presented in
random sequence, consisted of a set of vectors which overlapped i n O , 1, 2, 3, 4, 5, or 6 positions
with the next vector in the set. The columns of the matrix were sorted after the learning to bring
similar columns together, so that the types of neuron formed and the pattern of synapses formed
on their dendrites can be seen easily. The dendrites with random patterns of synapses have not
been allocated to any of the input stimuli. It is shown that application of one of the input stimuli
(overlap four A) or vectors which overlapped in four of eight positions with another stimulus
(overlap four B) produced one pattern of firing of the output neurons, and that application of input
stimulus overlap four B produced a different pattern of firing of the output neurons. Thus the stimuli
were correctly categorized by the matrix as being different.
 Functions of Hippocampal and Cortical Networks 193

hippocampus. Each mossy fiber forms approximately 10 "mosses," in which

there are dendrites of perhaps 5 different CA3 pyramidal cells. Thus each dentate
granule cell may contact approximately 50 CA3 pyramidal cells (in the mouse;
see Braitenberg & Schuz, 1984). In the rat, each mossy fiber forms approximately
14 mosses or contacts with CA3 cells, there are 1 X 106 dentate granule cells and
thus 14 X 106 mosses onto 0.18 X 106 CA3 cells (D. G. Amaral, personal com-
munication, 1988), and thus each dentate granule cell may contact approxi-
mately 78 CA3 pyramidal cells. This means that (in the rat) the probability that
a CA3 cell is contacted by a given dentate granule cell is 78 synapses per 106
granule cells, or .000078. These mossy fiber synapses are very large, presumably
because with such a relatively small number on each CA3 cell dendrite (and a
much smaller number active at any one time), each synapse must be relatively
strong. One effect that can be achieved by this low probability of contact of a
particular dentate granule cell with a particular pyramidal cell is pattern sepa-
ration. Consider a pattern of firing present over a set of dentate granule cells.
The probability that any two CA3 pyramidal cells receive synapses from a sim-
ilar subset of the dentate granule cells is very low (because of the low probability
of contact of any one dentate granule cell with a pyramidal cell), so that each
CA3 pyramidal cell is influenced by a very different subset of the active dentate
granule cells. Thus each pyramidal cell effectively samples a very small subset
of the active granule cells, and it is therefore likely that each CA3 pyramidal cell
will respond differently to the others, so that in this way pattern separation is
achieved. (The effect is similar to codon formation described in other contexts
by Marr, 1970.) With modifiability of the mossy fiber synapses, CA3 neurons
learn to respond to just those subsets of activity that do occur in dentate granule
cells. Moreover, because of the low probability of contact, and because of the
competition between the CA3 neurons, the patterns that occur are evenly dis-
tributed over different ensembles of CA3 neurons. This pattern separation effect
can be seen in Figure 9.4. (It may be noted that even if competition does not

FIGURE 9.4. Hippocampal stimulation. Conventions as in Figure 9.3. The dentate gyrus is
shown as a competition matrix at the left, receiving input stimuli from the perforant path. The
vertical dendrites of the CA3 pyramidal cells extend throughout the three submatrices shown in
the middle. The middle submatrix, which operates as a competition matrix, receives the output of
the dentate granule cells via the mossy fibers with potentially powerful synapses and a low contact
probability. Pattern separation can be seen to operate in that input vectors are converted into output
vectors with many elements activated by the inputs about which the submatrix has learned, and in
that different output vectors are produced for even quite similar input vectors. The upper CA3
submatrix operates as a competition matrix with a direct perforant path input. The lower CA3
submatrix operates as an autoassociation matrix (formed by the recurrent collaterals). The output
of the CA3 cells (summed vertically up and down the dendrite) is then used as the input (via the
Schaffer collaterals) to the CA1 cells, which operate as a competition matrix. The states of the
matrices after 2000 presentations of the same set of stimuli used for Figure 9.3. are shown. One
point demonstrated is that two very similar stimuli, overlap six A and overlap six B, produce
output vectors at CA1 which are relatively orthogonal to each other.
194 REGULATION OF CORTICAL FUNCTION IN MEMORY

operate in this system to increase orthogonality, the low probability of connec-

tions just described would nevertheless mean that the hippocampus could oper-
ate to produce relatively orthogonal representations.)
It is notable that in addition to the mossy fiber inputs, the CA3 pyramidal
cells receive inputs directly from perforant path fibers (see Figs. 9.1 and 9.2).
This is not a sparse projection, in that each pyramidal cell may receive on the
order of 2300 such synapses. (This is calculated using the evidence that of 15
mm of total dendritic length with 10,000 spines, approximately 3.5 mm [range
2.5-4.5 mm] is in the lacunosum moleculare and thus receives inputs from the
perforant path; D. G. Amaral, personal communication, 1988). What would be
the effect of this input together with the very sparse but strong synapses from
the mossy fibers? One effect is that the mossy fiber input would cause the pattern
of synapses (considered as a vector) on each pyramidal cell to point in a different
direction in a multidimensional space. However, the precise direction in that
multidimensional space could not be well specified by the relatively small num-
ber of mossy fiber synapses onto each CA3 pyramidal cell. But a particular cell,
once pointed to that part of space by the mossy fibers, would show cooperative
Hebbian learning between its activation by the mossy input and the direct per-
forant path input, allowing the direct input to come by learning to specify the
exact direction of that cell in multidimensional space much more effectively
than by the coarse mossy fiber input alone. This effect can be seen in Figure 9.4.
The relative weighting in this simulation was that the mossy fiber input had an
effect on each neuron which was five times greater than that of the direct per-
forant path input. Thus it is suggested that the combination of the sparse mossy
fiber input and the direct perforant path input is to achieve pattern separation,
and at the same time to allow the response of the neuron to be determined not
just by the sparse mossy fiber input, but much more finely by making use of the
direct perforant path input.
An additional feature of the hippocampus, developed in the CA3 pyramidal
cells in particular, is the presence of strong recurrent collaterals, which return
from the output of the matrix to cross over the neurons of the matrix, as shown
in Figures 9.1 and 9.2. This anatomy immediately suggests that this is an autoas-
sociation (or autocorrelation) matrix. The autoassociation arises because the
output of the matrix, expressed as the firing rate of the CA3 pyramidal cells, is
fed back along the horizontally running axons so that the pattern of activity in
this part of the matrix (the CA3 pyramidal cells) is associated with itself (see,
e.g., Kohonen, Oja, & Lehtio, 1981; Rolls, 1987). It can be noted here that for
this suggestion to be the case, the synapses of the recurrent collaterals would
have to be modifiable, and the modification rule would require alteration of syn-
aptic strength when both the presynaptic fiber and the postsynaptic dendrite
were strongly activated. Further, the probability of contact of the neurons in the
autoassociation matrix must not be very low if it is to operate usefully (see Marr,
1971). Given that the region of the CA3 cell dendrites on which the recurrent
collaterals synapse is long (approximately 11.5 mm), and that the total dendritic
length is approximately 15 mm and has approximately 10,000 spines (D. G.
 Functions of Hippocampal and Cortical Networks 195

Amaral, personal communication, 1988), approximately 7700 synapses per CA3

pyramidal cell could be devoted to recurrent collaterals, which with 180,000
CA3 neurons in the rat makes the probability of contact between the CA3 neu-
rons .043. This is high enough for the system to operate usefully as an autoas-
sociation memory (see Marr, 1971). It is remarkable that the contact probability
is so high, and also that the CA3 recurrent collateral axons travel so widely in
all directions that they can potentially come close to almost all other CA3 neu-
rons (D. G. Amaral, personal communication, 1988).
The importance of the autoassociation performed by this part of the matrix is
that it forms a recognition memory, with all the advantageous emergent prop-
erties of a matrix memory, such as completion, generalization, and graceful deg-
radation (see Kohonen, 1984; Kohonen et al, 1977, 1981; Rolls, 1987). One par-
ticularly relevant property here is completion: if part of a stimulus (or event)
occurs, then the autoassociation part of the matrix completes that event. Com-
pletion may operate particularly effectively here, because it operates after the
granule cell stage, which will reduce the proportion of neurons firing to represent
an input event to a low number partly because of the low probability of contact
of the granule cells with the CA3 pyramidal cells. Under these conditions the
simple autocorrelation effect can reconstruct the whole of one pattern without
interference, which would arise if too high a proportion of the input neurons
was active. A scheme of this type, although not expressed this way in the autoas-
sociation matrix formulation, was proposed by Marr in 1971. Another effect of
the autoassociation matrix is that patterns of activity which are not similar to
those already learned by this type of recogniton memory are lost, so that noisy
patterns can be cleaned up by the autoassociation matrix. It is further notable
that these completion and cleaning up processes may benefit from several iter-
ations (repeated cycles) of the autoassociation feedback effect. It has been sug-
gested by McNaughton (personal communication, 1987) that one function of
hippocampal theta activity may be to allow this autoassociation effect produced
by the recurrent collaterals to cycle for several iterations (in a period of approx-
imately 50 msec) and then to stop, so that the system can operate again with
maximal sensitivity to new inputs received on the mossy and perforant path
systems by the CA3 cells.
The CA1 pyramidal cells that receive from the CA3 cells are considered to
form a further stage of competitive learning, which has the effect of further clas-
sification of signals received, perhaps enabled by the pattern of connections
within the hippocampus to form these classifications over inputs received from
any part of the association neocortex. The firing of the CA1 cells would thus
achieve a much more economical and orthogonal classification of signals than
that present in the perforant path input to the hippocampus. These signals are
then returned to the association neocortex through the subiculum, entorhinal
cortex, and parahippocampal gyrus, as indicated in Figures 9.1 and 9.2. It is
suggested below that one role which these economical (in terms of the number
of activated fibers) and relatively orthogonal signals play in neocortical function
is to guide information storage or consolidation in the neocortex.
196 REGULATION OF CORTICAL FUNCTION IN MEMORY

Multilayer networks (such as the hippocampus) can potentially solve classes

of problem that cannot be solved in principle by single-layer nets (Rumelhart &
Zipser, 1986), because subcategories formed in an early stage of processing can
enable a later stage to find solutions or categories that are not linearly separable
in the input information space.
Having considered the computational theory of how the hippocampal cir-
cuitry may function, we can now turn to a systems level analysis, in which the
inputs and outputs of the hippocampus are considered and the function per-
formed by the hippocampus in relation to overall brain function can be
formulated.

SYSTEMS LEVEL THEORY OF HIPPOCAMPAL FUNCTION

The anatomical connections of the primate hippocampus with the rest of the
brain are considered first, to provide a basis for considering how the computa-
tional ability of the hippocampus could be used by the rest of the brain. The
hippocampus receives inputs by two main routes, the entorhinal cortex and the
fimbria/fornix. The entorhinal cortex (area 28) provides the hippocampus with
extensive inputs from the neocortex (i.e., cerebral cortex; see Fig. 9.5). Thus all
temporal neocortical areas project to area 35, the perirhinal cortex, or to area
TF-TH, in the parahippocampal gyrus, in turn projecting to the entorhinal cor-
tex (Amaral, 1987; Van Hoesen, 1982; Van Hoesen & Pandya, 1975a). The pari-
etal cortex (area 7) projects to area TF-TH, and thus can potentially influence
the hippocampus. The orbitofrontal cortex, areas 12 and 13, projects directly to
the entorhinal cortex (Van Hoesen, Pandya, & Butters, 1975). In addition, the
entorhinal cortex receives inputs from the amygdala. The entorhinal cortex itself
projects via the perforant path to reach primarily the dentate granule cells of the
hippocampus proper. By these routes the hippocampus receives information
after it has been highly processed through the temporal, parietal, and frontal
cortices. It must thus be of great importance for hippocampal function in the
primate that by its main input, the perforant path, it receives information from
the highest parts of the neocortex. There are also inputs to the hippocampus via
the fimbria/fornix from the cholinergic cells of the medial septum and the
adjoining limb of the diagonal band of Broca. The hippocampus also receives a
noradrenergic input from the locus coeruleus and a 5-hydroxytryptamine (5-HT)
input from the median raphe nucleus.
A major output of the hippocampus arises from the hippocampal pyramidal
cells and projects back via the subiculum to the entorhinal cortex, which in turn
has connections back to area TF-TH, which, in turn, projects back to the neo-
cortical areas from which it receives inputs (Van Hoesen, 1982); see Figs. 9.1
and 9.5). Thus the hippocampus can potentially influence the neocortical
regions from which it receives inputs. This is the pathway that seems to be
involved in guiding memory storage in the neocortex. A second efferent projec-
tion of the hippocampal system reaches the subiculum from the CA1 pyramidal
 Functions of Hippocampal and Cortical Networks 197

FIGURE 9.5. Connections of the primate hippocampus with the neocortex. A medial view of
the macaque brain is shown below, and a lateral view is shown inverted above. The hippocampus
receives its inputs via the parahippocampal gyrus, areas TF and TH, and the entorhinal cortex,
area 28. The return projections to the neocortex (shown on the right) pass through the same areas.
Cortical areas 19, 20, and 21 are visual association areas, 22 is auditory association cortex, 7 is
parietal association cortex, and 9, 46, 12, and 13 are frontal cortical association areas. (From Van
Hoesen, 1982)

cells and travels via the fimbria and (postcommissural) fornix to the anterior
thalamus and to the mammillary bodies, which, in turn, project to the anterior
thalamus. The anterior thalamus, in turn, projects into the cingulate cortex,
which itself has connections to the supplementary motor cortex, providing a
potential route for the hippocampus to influence motor output (Van Hoesen,
1982). Functions of the hippocampus in, for example, conditional spatial
response learning apparently utilize this output path to the motor system.
The connections of the hippocampus with other parts of the brain and the
internal connections and synaptic modifiability previously described, suggest
that the hippocampus should be able to detect, and classify onto a few specifi-
cally responding neurons, specific conjunctions of complex (cortically processed)
events, such as the information that a particular object (presumably reflecting
temporal lobe visual processing) has appeared in a particular position in space
(probably reflecting parietal input). Another example might be that a particular
stimulus should be associated with a particular spatial motor response. It has
been shown that this is the type of specific information which comes to activate
198 REGULATION OF CORTICAL FUNCTION IN MEMORY

different hippocampal neurons while monkeys are performing object-place

memory and conditional spatial response learning tasks. Indeed, the neurophys-
iological findings discussed earlier provide evidence supporting the model of
hippocampal function just described. The model is also supported by the evi-
dence that during the learning of conditional spatial responses some hippocam-
pal neurons start, but then stop showing differential responses to the different
stimuli, consistent with competitive interactions between hippocampal neurons
during learning, so that only some hippocampal neurons become allocated to
any one learned event or contingency.
The foregoing analyses have shown that the hippocampus receives from high-
order areas of association cortex; is able by virtue of the large number of syn-
apses on its dendrites to detect conjunctions of events even when these are
widely separated in information space, with their origin from different cortical
areas; allocates neurons to code efficiently for these conjunctions probably using
a competitive learning mechanism; and has connections back to the neocortical
areas from which it receives, as well as to subcortical structures via the fimbria/
fornix system. What could be achieved by this system? It appears that the long-
term storage of information is not in the hippocampus, at least in humans,
where damage to the hippocampus does not necessarily result in major retro-
grade amnesia (Squire, 1986, 1987). On the other hand, the hippocampus does
appear to be necessary for the storage of certain types of information (charac-
terized by the description declarative, or knowing what, as contrasted with pro-
cedural, or knowing how). How could the hippocampus then be involved in the
storage of information?
These and the other findings suggest that the hippocampus is specialized to
detect the best way in which to store information and then, by the return paths
to the neocortex, directs memory storage there. Clearly the hippocampus, with
its large number of synapses on each neuron and its potentiation type of learn-
ing, is able to detect when there is coherence (i.e., conjunctive activation of arbi-
trary sets of its input fibers) and is able, as indicated both theoretically and by
recordings made in the behaving monkey, to allocate neurons to economically
(i.e., with relatively few neurons active) code for each complex input event. Such
neurons could then represent an efficient way in which to store information, in
that redundancy would effectively have been removed from the input signal. In
a sense, the optimal way in which to build high-level feature analyzers could be
determined by the hippocampus. It should be noted that this theory is not incon-
sistent with the possibility that the hippocampus provides a working memory,
in that in the present theory the hippocampus sets up a representation using
Hebbian learning, which is useful in determining how information can best be
stored in the neocortex. (The representation found by the hippocampus could
provide a useful working memory [see Olton, 1984]; indeed, in the object-place
memory task described earlier, the object and place combinations formed onto
single hippocampal neurons would provide a useful working memory. An
understanding of the operations performed by the hippocampus at the neuronal
 Functions of Hippocampal and Cortical Networks 199

network level may enable one to visualize how the hippocampus could contrib-
ute to several functions that are not necessarily inconsistent.)
Where the long-term storage occurs and how it may be directed by the hip-
pocampus mut be considered next. The hippocampus is reciprocally connected
via the subiculum and entorhinal cortex with the parahippocampal gyrus, which
in turn is reciprocally connected with many high-order areas of association neo-
cortex (see Fig. 9.5). It is therefore possible that the actual storage takes place in
the parahippocampal gyrus and that this might be particularly appropriate for
multimodal memories. However, having detected that, for example, a visual
stimulus is regularly associated with an event in another modality such as a
movement, it might be useful to direct the unimodal representation of that
visual image, so that it is stored efficiently and can provide a useful input to the
multimodal conjunction store. Thus return pathways (via, e.g., the parahippo-
campal gyrus) to unimodal cortex (e.g., inferior temporal visual cortex, area TE)
might be used to direct unimodal storage by contributing to detection of the
most economical way in which to store representations of stimuli.
The question of how the hippocampal output is used by the neocortex (i.e.,
cerebral cortex) will be considered next. Given that the hippocampal output
returns to the neocortex, a theory of backprojections in the neocortex will be
needed. By way of introduction, it may be noted that the particular hippocampal
neurons that happen to represent a complex input event are not determined by
any teacher or forcing (unconditioned) stimulus. The neocortex must be able to
utilize the signal cleverly. One possibility is that any neocortical neuron with a
number of afferents active at the same time that hippocampal return fibers in its
vicinity are active modifies its responses so that it comes to respond better to
those afferents the next time they occur. This learning by the cortex would
involve a Hebb-like learning mechanism. One function served by what are thus
in effect backprojections from the hippocampus is some guidance for or super-
vision of neocortical learning. It is a problem of unsupervised learning systems
that they can efficiently detect local conjunctions, which are not necessarily
those of most use to the whole system. It is exactly this problem that the hip-
pocampus apparently helps to solve by detecting useful conjunctions globally
(i.e., over the whole of information space) and then directing storage locally at
earlier stages of processing so that filters are built locally which provide repre-
sentations of input stimuli useful for later processing.

THEORETICAL SIGNIFICANCE OF BACKPROJECTIONS

IN THE NEOCORTEX

The forward and backward projections that will be considered are shown in Fig-
ure 9.6 (for further anatomical information see Peters & Jones, 1984). In pri-
mary sensory cortical areas, the main extrinsic forward input is from the thala-
200 REGULATION OF CORTICAL FUNCTION IN MEMORY

FIGURE 9.6. Schematic diagram of forward and backward projections in the neocortex. The
superficial pyramidal cells (triangles) in layers 2 and 3 project forward to terminate in layer 4 of
the next cortical area. The deep pyramidal cells in the next area project back to layer 1 of the
preceding cortical area, in which there are apical dendrites of pyramidal cells. The hippocampus
and amygdala also are the source of backprojections which end in layer 1. Spiny stellate cells are
represented by small circles in layer 4. See text for further details.

mus and ends in layer 4, where synapses are formed onto spiny stellate cells.
These in turn project heavily onto pyramidal cells in layers 3 and 2, which in
turn send projections forward to terminate strongly in layer 4 of the next cortical
layer (on small pyramidal cells in layer 4 or on the basal dendrites of the layer
2 and 3 [superficial] pyramidal cells). Although the forward afferents end
strongly in layer 4, the forward afferents have some synapses also onto the basal
dendrites of the layer 2 pyramidal cells, as well as onto layer 6 pyramidal cells
and inhibitory interneurons. Inputs reach the layer 5 (deep) pyramidal cells from
the pyramidal cells in layers 2 and 3 (Martin, 1984), and the deep pyramidal
cells send backprojections to end in layer 1 of the preceding cortical area (see
Figure 9.6), where there are apical dendrites of pyramidal cells. There are few
current theories about the functions subserved by the cortico-cortical backpro-
jections, even though there are almost as many backprojecting as forward-pro-
jecting axons. It is important to note that in addition to the axons and their
terminals in layer 1 from the succeeding cortical stage, there are also axons and
terminals in layer 1 in many stages of the cortical hierarchy from the amygdala
and (via the subiculum, entorhinal cortex, and papahippocampal cortex) from
the hippocampal formation (see Fig. 9.6; Amaral, 1986, 1987; Amaral & Price,
1984; Turner, 1981; Van Hoesen, 1981). The amygdala and hippocampus ar
stages of information processing at which the different sensory modalities (such
as vision, hearing, touch, and taste for the amygdala) are brought together, so
that correlations between inputs in different modalities can be detected in these
regions but not at prior cortical processing stages in each modality, since these
 Functions of Hippocampal and Conical Networks 201

cortical processing stages are unimodal. As a result of bringing together the two
modalities, significant correspondences between the two modalities can be
detected. For example, a particular visual stimulus may be associated with the
taste of food; another visual stimulus may be associated with painful touch.
Thus at these stages of processing, but not before, the significance of, say, visual
and auditory stimuli can be detected and signaled, and sending this information
back to the neocortex thus can provide a signal that indicates to the cortex that
information should be stored, but—even more than this—it provides an orthog-
onal signal that could help the neocortex to store the information efficiently.
The way in which backprojections could assist learning in the cortex can be
considered using the architecture shown in Figure 9.7. The (forward) input stim-
ulus occurs as a vector applied to (layer 3) cortical pyramidal cells through mod-
ifiable synapses in the standard way for a competitive net. (If it is a primary
cortical area, the input stimulus is at least partly relayed through spiny stellate
cells, which may help to normalize and orthogonalize the input patterns in a
preliminary way before the patterns are applied to the layer 3 pyramidal cells. If
it is a nonprimary cortical area, the cortico-cortical forward axons may end more
strongly on the basal dendrites of neurons in the superficial cortical layers.) The
lower set of synapses on the pyramidal cells would then start by competitive
learning to set up representations on the lower parts of these neurons, which
would represent correlations in the input information space and could be said
to correspond to features in the input information space, where a feature is
defined simply as the representation of a correlation in the input information
space.

FIGURE 9.7. The architecture used to simulate the properties of backprojections. The forward
(input stimulus) and backprojected axons make Hebb-modifiable synapses onto the same set of
vertical dendrites, which represent cortical pyramidal cells.
202 REGULATION OF CORTICAL FUNCTION IN MEMORY

Consider the application of one of the (forward) input stimulus vectors with
the conjunctive application of a pattern vector via the backprojection axons
with terminals in layer 1. Given that all the synapses in the matrix start with
random weights, some of the pyramidal cells will by chance be strongly influ-
enced both by the (forward) input stimulus and by the backprojecting vector.
These strongly activated neurons will then compete with each other as in a stan-
dard competitive net, to produce contrast enhancement of their firing patterns.
(The relatively short-range [50 /um] excitatory operations produced by bipolar
and double bouquet cells, together with more widespread [300 to 500 jum] recur-
rent lateral inhibition produced by the smooth nonpyramidal cells and perhaps
the basket cells, may be part of the mechanism of this competitive interaction.)
Next, Hebbian learning takes place as in a competitive net, with the addition
that not only are the synapses between forward-projecting axons and active post-
synaptic neurons modified, but the synapses in layer 1 between the backward-
projecting axons and the (same) active postsynaptic neurons are modified.
This functional architecture has the following properties. First, the backpro-
jections, which are assumed to carry signals that are relatively information-rich
(Rolls, 1987) and orthogonal to each other as a result of the conjunctions formed
by the hippocampus, amygdala, or next cortical stage, help the neurons to learn
to respond differently to (and thus separate) input stimuli (on the forward pro-
jection lines) even when the input stimuli are very similar. This is illustrated in
the simulation seen in Figure 9.8, where input stimuli that overlap even in six
positions out of eight can be easily learned as separate if presented conjunctively
with different orthogonal backprojecting "tutors." (For a similar idea on the
guidance of a competitive learning system see Rumelhart & Zipser, 1986.)
Another way in which the learning of pyramidal cells can be influenced is that
if two relatively different input stimuli are being received, but the same back-
projecting signal occurs with these two somewhat different forward inputs, then
the pyramidal cells, guided by the tutor, will build the representation of the for-
ward input stimuli on the same pyramidal cells. If the representations of the
forward inputs built by modification of the synapses on the pyramidal cells are
considered as vectors in a multidimensional space, then the representations will
be forced toward each other as a result of the operation of the common back-
projecting tutor.
In the cerebral cortex, the backprojecting tutors can be of two types. One orig-
inates from the amygdala and hippocampus, and, by benefiting from cross-

FIGURE 9.8. Simulation of neocortical backprojection learning matrix. Conventions as in Figure

9.3. (A) During learning, both a forward input (chosen from the same set used in Figure 9.3) and
a backprojected vector which was orthogonal to the other backprojected vectors are presented
simultaneously. After 206 cycles with input stimulus-backprojected tutor pairs chosen in random
sequence, the synapses had modified as shown. In (B,C) two similar input stimuli (overlap six A
and six B) produce different outputs. The learning has been guided by the backprojected tutors
presented during learning (D). If only the tutor originally paired with input stimulus overlap six A
is presented, then recall of the output vector normally recognized by input stimulus overlap six A
occurs.
(A)

(B)
(C)

(D)
 Functions of Hippocampal and Cortical Networks 205

modal comparison, can provide an orthogonal backprojected vector. This back-

projection, moreover, may be activated only if the multimodal areas detect that
the visual stimulus is significant, for example, if it is associated with a pleasant
taste. This provides one way in which a competitive learning system can be
guided as to what it should learn, so that it does not attempt to lay down rep-
resentations of all incoming sensory information, but only those shown to be
significant as a result of comparison made with inputs originating in a different
modality. Another way for this important function to be achieved is by activa-
tion of neurons that "strobe" the cortex when new or significant stimuli are
shown. The cholinergic system originating in the basal forebrain (which itself
receives information from the amygdala) and the noradrenergic input to layer 1
of the cortex from the locus coeruleus may also contribute to this function (see
Bear & Singer, 1986; Rolls, 1987). However, in that there are relatively few neu-
rons in these basal forebrain and noradrenergic systems, it is suggested that these
projections provide only a simple strobe (see Rolls, 1987). In contrast, there are
sufficient axons in the backprojecting systems from the hippocampus and amyg-
dala to carry pattern-specific information, so that the hypothesis presented here
is that these backprojections guide the consolidation of information by influenc-
ing how information is stored on pyramidal cells in the way described above.
The situations in which the amygdala and hippocampal backprojections guide
learning (as well as recall and attention, as discussed later) are probably different
in the following way. The amygdala is particularly involved when associations
are made between a neutral stimulus (such as a visual stimulus) and a reinforc-
ing stimulus (e.g., from the taste or touch inputs it receives; see Rolls, 1985,
1987); in contrast, the hippocampus is particularly involved when conjunctions
must be detected between spatial inputs (received by the major inputs of the
hippocampus which originate from the parietal cortex) and its other inputs (e.g.,
the inferior temporal visual cortex), in, for example, conditional spatial response
and object-place learning tasks.
The second type of backprojection is from the next cortical area in the hier-
archy. The next cortical area operates in the same manner, and because it is a
competitive system, it is able to further categorize or orthogonalize the stimuli
it receives. This next cortical stage then projects back these more orthogonal
representations as tutors to the preceding stage, to effectively build at the pre-
ceding stage better filters for the diagnosis of the categories being found at the
next stage.
A second property of this architecture is that it provides for recall. Recall of
a previously stored representation by another stimulus or event occurs in the
following way. If only the tutor is presented, then the neurons originally acti-
vated by the forward-projecting input stimuli are activated. This occurs because
the synapses from the backprojecting axons onto the pyramidal cells have been
modified only where there was conjunctive forward-projected and backprojected
activity during learning. Figure 9.8 is a simulation of this. Consider the situation
when in the visual system the sight of food is forward projected onto pyramidal
cells, and conjunctively there is a backprojected signal elicited by the taste of the
food and originating from a multimodal area such as the amygdala. Neurons
206 REGULATION OF CORTICAL FUNCTION IN MEMORY

that have conjunctive inputs from these two stimuli set up representations of
both, so that later if a backprojected signal produced by the taste occurs, then
the visual neurons originally activated by the sight of that food will be activated,
and thus recall of the visual representation is achieved. In this way many of the
low-level details of the original visual stimulus can be recalled. Evidence that
during recall relatively early cortical processing stages are activated comes from
cortical blood flow studies in humans, in which it has been found, for example,
that early visual cortical association areas are activated during recall of visual
(but not auditory) information (Roland & Friberg, 1985; Roland, Vaernet, &
Lassen, 1980). Recall could operate as a result of signals backprojected directly
from the amygdala or parahippocampal gyrus or indirectly through series of
back-connected cortical stages.
A third property of this architecture is that attention could operate from
higher to lower levels to selectively facilitate only certain pyramidal cells by
using the backprojections. Indeed, the backprojections described could produce
many of the "top-down" influences common in perception. A fourth property is
that semantic priming could operate by using the backprojecting neurons to pro-
vide a small activation of just those neurons in earlier stages which are appro-
priate for responding to that semantic category of input stimulus.
A fifth property of such a return mechanism, which, on detecting a conjunc-
tion (perhaps across modalities), improves unimodal representations of the
input stimuli, would be a form of positive feedback that would gradually
improve storage through the reciprocal interactions, as the feedback effect pro-
duced a better representation at the preceding level to be fed forward, with this
occurring repeatedly. This might provide a neurophysiological and computa-
tional basis for any gradient of retrograde amnesia that occurs for the period just
before disruption of temporal lobe function (Squire, 1986, 1987). A sixth prop-
erty of the backprojections is that they would assist the stability of the preceding
competitive networks by providing a relatively constant guiding signal as a
result of associations made at a higher stage, for example, to an unconditioned
taste or somatosensory input.
This theory of the functions of backprojections in the cerebral cortex requires
a large number of backprojecting axons, as pattern-specific information (used to
guide learning by providing a set of mutually orthogonal guidance signals, or to
produce recall) must be provided by the backprojections. It also solves the dead-
dressing problem, for the hippocampus does not need to know exactly where in
the cortex information should be stored. Instead, the backprojection signal
spreads widely in layer 1, and the storage site is simply on those neurons that
happen to receive strong (and precise) forward activity as well as backprojected
activity, so that Hebbian learning by conjunction occurs there. This scheme is
consistent with neocortical anatomy, in that it requires the same pyramidal cell
to receive both forward-projected and (more diffuse) backprojected activity,
which the arrangement of pyramidal cells with apical dendrites extending all the
way up into layer 1 achieves (see Peters & Jones, 1984). Indeed, in contrast to
the relatively localized terminal distributions of forward cortico-cortical and
 Functions of Hippocampal and Cortical Networks 207

thalamo-cortical afferents, the cortico-cortical backward projections that end in

layer 1 have a much wider horizontal distribution, of up to several millimeters
(Amaral, 1986). The suggestion is thus that this enables the backward-projecting
neurons to search over a large number of pyramidal cells in the preceding cor-
tical area for activity conjunctive with their own. Note that the theory utilizes a
Hebbian learning scheme, which provides for learning to occur when conjointly
there is forward and backprojected input to a pyramidal cell resulting in suffi-
cient postsynaptic activation to provide for modification of synapses that hap-
pen to be active. This provides the opportunity to make it clear that the theo-
retical ideas introduced here make clear predictions that can be empirically
tested. For example, the theory of backprojections just proposed predicts that
the backprojections in the cerebral cortex have modifiable synapses on pyrami-
dal cells in the previous cortical area. If this were found not to be the case in
empirical tests, then the theory would be rejected.
The ideas introduced here also have many theoretical implications. One is
that if the backprojections are used for recall, as seems likely (as discussed
above; see also Roland & Friberg, 1985), then this would place severe con-
straints on their use for functions such as error backpropagation. Error back-
propagation is an interesting and powerful algorithm in parallel distributed pro-
cessing networks for setting the weights in hidden units (i.e., nodes in layers
which are not input and output layers) to allow networks to learn useful map-
pings between input and output layers (Rumelhart, Hinton, & Williams, 1986).
However, the backprojections in the architecture in which this algorithm is
implemented have precise functions in conveying error from the output layer
back to the earlier, hidden layers. It would be difficult to use the weights (syn-
aptic strengths) from the backprojecting neurons to neurons in earlier layers
both to convey the error correctly and to have the appropriate strengths for
recall.

CONCLUSIONS

In this chapter, experimental evidence on and theoretical approaches to the

function of the hippocampus and of backprojections in the cerebral cortex were
considered. Theories of how the hippocampus functions and of the functions of
backprojections in the neocortex were proposed. The theories are at the level of
neuronal networks, and are based on evidence on the fine architecture of the
networks, on the rules of synaptic modifiability incorporated, and on the sys-
tems level connections. It is suggested that this approach will be useful in linking
anatomical evidence on structure to physiological evidence on modifiability,
understanding the global properties of the networks, and thus understanding the
role of the networks in brain function and behavior.

Acknowledgments
The author has worked on some of the experiments and neuronal network mod-
eling described here with P. Cahusac, D. Cohen, J. Feigenbaum, R. Kesner, Y.
208 REGULATION OF CORTICAL FUNCTION IN MEMORY

Miyashita, and H. Niki, and their collaboration is sincerely acknowledged. Dis-

cussions with David G. Amaral of the Salk Institute, La Jolla, were also much
appreciated. This research was supported by the Medical Research Council.

REFERENCES

Amaral, D. G. (1986). Amygdalohippocampal and amygdalocortical projections in the primate

brain. In R. Schwarz and Y. Ben-Ari, (Eds.), Excitatory amino acids and epilepsy (pp.
3-17) New York: Plenum.
Amaral, D. G. (1987). Memory: Anatomical organization of candidate brain regions. In F.
Plum (Ed.), Handbook of physiology: Sec. 1. The nervous system (pp. 211-294).
Bethesda, Md.: American Physiological Society.
Amaral, D., & Price, J. L. (1984). Amygdalo-cortical projections in the monkey (Macaco fas-
cicularis). Journal of Comparative Neurology, 230, 465-496.
Bear, M. F., Cooper, L. N., & Ebner, F. B. (1987). A physiological basis for a theory of synapse
modification. Science, 237, 42-48.
Bear, M. F., & Singer, W. (1986). Modulation of visual cortical plasticity by acetylcholine and
noradrenaline. Nature, 320, 172-176.
Braitenberg, V., & Schuz, A. (1982). Some anatomical comments on the hippocampus. In W.
Seifert (Ed.), Neurobiology of the Hippocampus (pp. 21-37), London: Academic
Press.
Cahusac, P. M. B., Feigenbaum, J., Rolls, E. T., Miyashita, Y., & Niki, H. (1986). Modifications
of neuronal activity during learning in the primate hippocampus. Neuroscience Let-
ters, S24, S29.
Cahusac, P. M. B., Miyashita, Y., & Rolls, E. T. (1989). Responses of hippocampal neurons in
the monkey related to delayed spatial response and object-place memory tasks.
Behavioural Brain Research 33, 229-240.
Corkin, S. (1984). Lasting consequences of bilateral medial temporal lobectomy: Clinical course
and experimental findings in H. M. Seminars in Neurology, 4, 249-259.
Gaifan, D. (1974). Recognition impaired and association intact in the memory of monkeys after
transection of the fornix. Journal of Comparative and Physiological Psychology, 86,
1100-1109.
Gaffan, D. (1977). Monkeys' recognition memory for complex pictures and the effects of fornix
transection. Quarterly Journal of Experimental Psychology, 29, 505-514.
Gaffan, D. (1985). Hippocampus: Memory, habit and voluntary movement. Philosophical
Transactions of the Royal Society B, 308, 87-99.
Gaffan, D., & Saunders, R. C. (1985). Running recognition of configural stimuli by fornix tran-
sected monkeys. Quarterly Journal of Experimental Psychology, 37B, 61-71.
Grossberg, S. (1982). Studies of mind and brain. New York: Reidel.
Grossberg, S. (1987). Competitive learning: From interactive activation to adaptive resonance.
Cognitive Science, 11, 23-63.
Hinton, G. E., & Anderson, J. A. (1981). Parallel models of associative memory. Hillsdale, N.J.:
Erlbaum.
Ito, M. (1984). The cerebellum and neural control. New York: Raven.
Jones, E. G., & Peters, A. (Eds.). (1984). The cerebral cortex: Vol. 2. Functional properties of
cortical cells. New York: Plenum.
Kelso, S. R., Ganong, A. H., & Brown, T. H. (1986). Hebbian synapses in the hippocampus.
Proceedings of the National Academy of Sciences, 83, 5326-5330.
Kohonen, T. (1984). Self-organization and associative memory. Berlin: Springer-Verlag.
Kohonen, T., Lehtio, P., Rovamo, J., Hyvarinen, J., Bry, K., & Vainio, L. (1977). A principle
of neural associative memory. Neuroscience, 2, 1065-1076.
 Functions of Hippocampal and Cortical Networks 209

Kohonen, T., Oja, E., & Lehtio, P. (1981). Storage arid processing of information in distributed
associative memory systems. In G. E. Hinton & J. A. Anderson (Eds.), Parallel mod-
els of associative memory (pp. 105-143). Hillsdale, N.J.: Erlbaum.
Levy, W. B. (1985). Associative changes in the synapse: LTP in the hippocampus. In W. B.
Levy, J. A. Anderson, & S. Lehmkuhle (Eds.), Synoptic modification, neuron selec-
tivity, and nervous system organization (pp. 5-33). Hillsdale, N.J.: Erlbaum.
Levy, W. B., & Desmond, N. L. (1985). The rules of elemental synaptic plasticity. In W. B.
Levy, J. A. Anderson, & S. Lehmkuhle (Eds.), Synaptic modification, neuron selec-
tivity, and nervous system organization (pp. 105-121). Hillsdale, N.J.: Erlbaum.
Marr, D. (1970). A theory for cerebral cortex. Proceedings of the Royal Society B, 176, 161-
234.
Marr, D. (1971). Simple memory: A theory for archicortex. Philosophical Transactions of the
Royal Society B, 262, 23-81.
McNaughton, B. L. (1984). Activity dependent modulation of hippocampal synaptic efficacy:
Some implications for memory processes. In W. Seifert (Ed.), Neurobiology of the
hippocampus (pp. 233-252). London: Academic Press.
Martin, K. A. C. (1984). Neuronal circuits in cat striate cortex. In E. G. Jones & A. Peters (Eds.),
The cerebral cortex: Vol 2. Functional properties of cortical cells (pp. 241-285). New
York: Plenum.
Milner, B. (1982). Disorders of learning and memory after temporal lobe lesions in man. Clin-
ical Neurosurgery, 19, 421-446.
Mishkin, M. (1978). Memory severely impaired by combined but not separate removal of
amygdala and hippocampus. Nature, 273, 297-298.
Mishkin, M. (1982). A memory system in the monkey. Philosophical Transactions of the Royal
Society B, 298, 85-95.
Miyashita, Y., Rolls, E. T., Cahusac, P. M. B., & Niki, H. (1989). Activity of hippocampal
formation neurons in the monkey related to a stimulus-response association task.
Journal of Neurophysiology, 61, 669-678.
Murray, E. A., & Mishkin, M. (1984). Severe tactual as well as visual memory deficits follow
combined removal of the amygdala and hippocampus in monkeys. Journal ofN'eu-
roscience, 4, 2565-2580.
Murray, E. A., & Mishkin, M. (1985). Amygdalectomy impairs crossmodal association in mon-
keys. Science, 228, 604-606.
O'Keefe, J. (1984). Spatial memory within and without the hippocampal system. In W. Seifert
(Ed.), Neurobiology of the Hippocampus (pp. 375-403). London: Academic Press.
Olton, D. S. (1984). Memory functions and the hippocampus. In W. Seifert (Ed.), Neurobiology
of the hippocampus, (pp. 335-373). London: Academic Press.
Palm, G. (1982). Neural assemblies. Berlin: Springer-Verlag.
Peters, A., & Jones, E. G. (1984). The cerebral cortex: Vol. 1. Cellular components of the cerebral
cortex. New York: Plenum.
Petrides, M. (1985). Deficits on conditional associative-learning tasks after frontal- and tem-
poral-lobe lesions in man. Neuropsychologia, 23, 601-614.
Roland, P. E., & Friberg, L. (1985). Localization of cortical areas activated by thinking. Journal
of Neurophysiology, 53, 1219-1243.
Roland, P. E., Vaernet, K., & Lassen, N. A. (1980). Cortical activations in man during verbal
report from visual memory. Neuroscience Letters (Suppl. 5), S478.
Rolls, E. T. (1985). Connections, functions and dysfunctions of limbic structures, the prefrontal
cortex, and hypothalamus. In M. Swash & C. Kennard (Eds.), The scientific basis of
clinical neurology (pp. 201-213). London: Churchill Livingstone.
Rolls, E. T. (1987). Information representation, processing and storage in the brain: Analysis
at the single neuron level. In J.-P. Changeux & M. Konishi (Eds.), The neural and
molecular bases of learning (pp. 503-540). Chichester: Wiley.
Rolls, E. T. (1988). Visual information processing in the primate temporal lobe. In M. Imbert
(Ed.), Models of visual perception: From natural to artificial. Oxford: Oxford Uni-
versity Press.
210 REGULATION OF CORTICAL FUNCTION IN MEMORY

Rolls, E. T., Miyashita, Y., Cahusac, P. M. B., Kesner, R. P., Niki, H., Feigenbaum, J., & Bach,
L. (1989). Hippocampal neurons in the monkey with activity related to the place in
which a stimulus is shown. Journal of Neuroscience, 9, 1835-1845.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning internal representations
by error propagation. In D. E. Rumelhart & J. L. McClelland (Eds.), Parallel distrib-
uted processing: Vol. L Foundations (pp. 318-362). Cambridge, Mass.: MIT Press.
Rumelhart, D. E., & Zipser, D. (1986). Feature discovery by competitive learning. In D. E.
Rumelhart & J. L. McClelland (Eds.), Parallel distributed processing: Vol. 1. Foun-
dations (pp. 151-193). Cambridge, Mass.: MIT Press.
Rupniak, N. M. J., & Gaffan, D. (1987). Monkey hippocampus and learning about spatially
directed movements. Journal of Neuroscience, 7, 2331-2337.
Scoville, W. B., & Milner, B. (1957). Loss of recent memory after bilateral hippocampal lesions.
Journal of Neurology Neurosurgery & Psychiatry, 20, 11-21.
Smith, M. L., & Milner, B. (1981). The role of the right hippocampus in the recall of spatial
location. Neuropsychologia, 19, 781-793.
Squire, L. (1986). Mechanisms of memory. Science, 232, 1612-1619.
Squire, L. (1987). Complex connections involved in memory functions. In J. Byrne & W. O.
Berry (Eds.), Neural models of plasticity: Theoretical and empirical approaches.
Orlando, Fla.: Academic Press.
Turner, B. H. (1981). The cortical sequence and terminal distribution of sensory related affer-
ents to the amygdaloid complex of the rat and monkey. In Y. Ben-Ari (Ed.), The
amygdaloid complex. Amsterdam: Elsevier.
Turner, B. H., Mishkin, M., & Knapp, M. (1980). Organization of the amygdalopetal modality-
specific cortical association areas in the monkey. Journal of Comparative Neurology,
191, 515-543.
Van Hoesen, G. W. (1981). The differential distribution, diversity and sprouting of cortical
projections to the amygdala in the rhesus monkey. In Y. Ben-Ari (Ed.), The amyg-
daloid complex (pp. 79-90). Amsterdam: Elsevier.
Van Hoesen, G. W. (1982). The parahippocampal gyrus. New observations regarding its cor-
tical connections in the monkey. Trends in Neurosciences, 5, 345-350.
Van Hoesen, G. W., & Pandya, D. N. (1975a). Some connections of the entorhinal (area 28)
and perirhinal (area 35) cortices in the monkey. I. Temporal lobe afferents. Brain
Research, 95, 1-24.
Van Hoesen, G. W., & Pandya, D. N. (1975b). Some connections of the entorhinal (area 28)
and perirhinal (area 35) cortices in the monkey. III. Efferent connections. Brain
Research, 95, 39-59.
Van Hoesen, G. W., Pandya, D. M., & Butters, N. (1975). Some connections of the entorhinal
(area 28) and perirhinal (area 35) cortices in the monkey. II. Frontal lobe afferents.
Brain Research, 95, 25-38.
Von der Malsburg, C. (1973). Self-organization of orientation-sensitive columns in the striate
cortex. Kybernetik, 14, 85-100.
Wigstrom, H., Gustaffson, B., Huang, Y.-Y., & Abraham, W. C. (1986). Hippocampal long-
term potentiation is induced by pairing single afferent volleys with intracellularly
injected depolarizing currents. Acta Physiologica Scandinavica, 126, 317-319.
Zola-Morgan, S., Squire, L. R., & Amaral, D. G. (1986). Human amnesia and the medial tem-
poral region: Enduring memory impairment following a bilateral lesion limited to
field CA1 of the hippocampus. Journal of Neuroscience, 6, 2950-2957.
10
Ontogenetic Self-Organization
and Learning
WOLF SINGER

From a reductionist point of view, learning consists of a stimulus- and hence

activity-dependent neuronal process that leads to a lasting change of stimulus-
response relations. The formation of a memory trace may thus be considered a
modification of the neuronal program that specifies input-output functions.
This program is contained in the architecture of interneuronal connectivity and
in the differential weighing of the transfer functions of these connections. Thus
any activity-dependent process that modifies, in a sufficiently stable and long-
lasting way, the excitatory or inhibitory interactions between pairs of neurons
could serve as a mechanism of learning, and any long-lasting alteration of inter-
cellular communication can be considered an engram.
Activity-dependent long-term changes of neuronal interactions have been
observed in a number of brain structures of adult animals. They are a ubiquitous
phenomenon during early ontogeny, where they play a crucial role in the self-
organization of the brain. It is the goal of this short review to illustrate some of
the mechanisms of developmental self-organization and to investigate the extent
to which they can be considered precursors of mechanisms that mediate learning
in the adult.

PRINCIPLES OF SELF-ORGANIZATION

It has often been emphasized that genes cannot contain, in any naive sense, the
full information necessary to describe the brain. Cerebral cortex alone contains
on the order of 1014 synapses. Forgetting considerations of genome size, one can
hardly imagine how ontogeny could select the correct wiring diagram out of all
alternatives if these were of equal likelihood. Besides, judging from the vari-
ability of the vertebrate brain structure, the precision of the ontogenetic process
is not sufficient to specify individual connections. This implies that ontogeny
has to make use of self-organization: of general rules for pattern generation and
of simple principles of error correction.

211
212 REGULATION OF CORTICAL FUNCTION IN MEMORY

The development of organs and organisms is based on a permanent interac-

tive exchange between the information stored in the genome and the informa-
tion available in the cellular "environment" of the genome.
The cellular environment determines which of the genetic commands should
be expressed at particular developmental stages. As the developmental process
proceeds and organs differentiate, the microenvironment of the genome changes
and hence modifies continuously the further expression of the genes. During
early phases of brain development gene expression and posttranslational modi-
fications of gene products are controlled as in any other organ by the biochem-
ical signals produced in the cellular microenvironment. Later, however, brain
development starts to differ radically from the development of other organs
because electrical activity is added to the biochemical messengers as a further
signaling system in the self-organizing dialogue between the genes and their
respective environment. It has to be assumed that these electrical signals are
capable of influencing gene expression and posttranslational modifications.
This has a number of extremely important implications. Electrical signals are
transported by neuronal processes over great distances and with high topological
selectivity. This enlarges dramatically the range and the complexity of the envi-
ronment that is available to the self-organization process. From a certain devel-
opmental stage on the brain possesses functioning sense organs that convert sig-
nals from within the organism and even from extracorporal space into electrical
messages. Thus the environment relevant for brain self-organization ultimately
includes all domains with which the evolving brain is capable of interacting and
from which it can receive messages. Another important aspect is that the very
same electrical signals which convey sensory messages are used by the brain as
information carriers for computational processes. Hence the powerful capacities
of nerve nets to perform complex logic operations on large-parameter sets
become available to the self-organization process. As the complexity of the brain
increases, the computational power and the complexity of its interactions with
its environment increase as well. As a consequence, the set of parameters deter-
mining further brain development becomes more complex and capable of sup-
porting self-organization toward even more differentiated structures. Because of
this spiral of reciprocal interactions between the genome and its increasingly
complex environment, a rather small set of genetic rules suffices to promote the
development of such highly differentiated structures as the human brain. As will
become clear later, these very processes that support experience-dependent self-
organization during development are closely related to the more general pro-
cesses that serve to adapt the mature brain to its environment and are com-
monly known as learning.

THE VISUAL SYSTEM AS A MODEL

Most of our knowledge of activity-dependent specification of neuronal connec-

tions comes from the visual system. "Seeing" has to be learned during a critical
 Ontogenetic Self-Organization and Learning 213

period of early postnatal development. The most dramatic evidence for this
comes from patients who suffered from congenital opacities of the eyes during
early childhood and therefore were unable to perceive contours. With the devel-
opment of lens transplants the optical media of these patients' eyes could be
restored, but unexpectedly these patients were unable to recover visual functions
when operated on as juveniles or adults. Experiments in visually deprived ani-
mals revealed that these functional deficits are primarily due to abnormalities
in visual centers of the cerebral cortex. The reason is that certain cortical func-
tions can be developed only if visual experience is available.
The following example illustrates this. Higher mammals and humans who
have frontally positioned eyes with overlapping visual fields can compute the
distance of objects in space from the differences between the images in the two
eyes. This ability has two obvious advantages. First, the distance of an object in
space can be assessed with great precision even if object and observer are sta-
tionary and do not produce any motion parallaxes. Second, the separation of
figures from ground, a primordial prerequisite for pattern recognition, is greatly
facilitated by evaluating spatial distance. The basis for this function are neurons
in the visual cortex which possess two receptive fields, one in each eye, that have
the same features and are precisely superimposed in visual space. Thus during
development the 1 million afferents arriving from each eye have to be arranged
so that only those pairs of afferents converge onto cortical cells which originate
from precisely corresponding retinal loci. The problem is that there is no way to
predict with any great precision which retinal loci will actually be corresponding
in the mature visual system. Retinal correspondence depends on parameters
such as the size of the eyeballs, the position of the eyeballs in the orbit, and the
interocular distance. Clearly, these parameters are strongly influenced by epige-
netic factors. It follows that genetic instructions alone, even if they are quanti-
tatively sufficient, cannot in principle suffice to determine with the required pre-
cision the pattern of interocular connections. An elegant possibility exists,
however, to identify fibers as coming from corresponding retinal loci by evalu-
ating their electrical activity. When a target is fixated with both eyes, corre-
sponding retinal loci are stimulated by the same contours. Therefore, neuronal
responses in afferents from corresponding retinal loci are more correlated than
those in afferents from noncorresponding loci. What is required then is a devel-
opmental mechanism capable of consolidating selectively those retinal afferents
which convey correlated activation patterns.

ACTIVITY-DEPENDENT CONSOLIDATION AND DISRUPTION

OF FUNCTIONAL BINOCULARITY

In the following paragraphs I shall briefly review evidence suggesting that such
mechanisms do indeed exist in the mammalian brain. I shall further demon-
strate that these processes are not solely dependent on local interactions but are
in addition controlled by more globally organized modulatory systems, which
214 REGULATION OF CORTICAL FUNCTION IN MEMORY

are implicated in the control of central states such as arousal, attention, and
motivation.
Before visual experience becomes effective in influencing the ocularity of cor-
tical neurons, most cells in the visual cortex of cats and monkeys respond to
stimulation of both eyes. This condition is maintained with normal experience
but also with dark rearing. With monocular deprivation, however, the large
majority of cortical cells loses the ability to respond to the deprived eye (Blake-
more, Garey, & Vital-Durand, 1978; Wiesel & Hubel, 1965a). The cells that con-
tinue to respond to the deprived eye are located preferentially within layer IV
in the center of the shrunken termination fields of the deprived eye (Shatz &
Stryker, 1978). This suggests that only those cells which have been monocularly
driven by this eye right from the beginning continue to respond to the deprived
eye. With reverse suture, if it occurs early enough, the previously closed eye may
recover control over most cortical neurons while the previously open eye in turn
becomes ineffective (Blakemore et al, 1978; Wiesel & Hubel, 1965b).
When induced early in development, at times when axonal segregation is not
yet completed, these functional changes go in parallel with the retraction and
expansion of thalamic afferents within layer IV and reflect to some extent the
formation and withdrawal of synaptic connections between thalamic afferents
and cortical target cells. However, these modifications of ocular dominance
most certainly also include changes in the efficacy of intracortical synaptic con-
nections, this latter mechanism becoming relatively more prominent at later
stages of development when axonal segregation is already in an advanced stage.
This has to be inferred from the evidence that marked changes in the ocularity
of cortical neurons can occur without or with only minor modifications of the
termination patterns of thalamic afferents in layer IV (for review see Tieman,
1985). One example is the late change of ocular dominance in dark-reared kit-
tens. In normally reared kittens monocular deprivation is effective only during
a critical period that lasts about 3 months (Olson & Freeman, 1980). In dark-
reared kittens monocular deprivation induces changes of ocular dominance
even after the end of this period (Cynader, 1983). These late changes of ocularity
are, however, not associated with changes of columnar patterns in layer IV
(Mower, Caplan, Christen, & Duffy, 1985; Mower & Christen, 1985). Another
example for such a dissociation is the modification occurring with strabismus.
Here the morphological pattern of ocular dominance columns in layer IV is
nearly unchanged (LeVay, Stryker, & Shatz, 1978), while the functional ocularity
of cortical neurons is drastically altered. Most cells are monocular and driven
either by the right or the left eye (Hubel & Wiesel, 1965). Finally, there is evi-
dence from single-cell recording and deoxyglucose mapping in monocularly
deprived cats that activity is still present within the deprived eye's columns but
is confined to layer IV, indicating transmission failure at the level of intracortical
connections (Bonds, Silverman, Sclar, & Tootell, 1980; Shatz & Stryker, 1978).
It is these latter changes in excitatory transmission rather than the macroscopic
rearrangements of columnar patterns that will be dealt with in the following
section.
 Ontogenetic Self-Organization and Learning 215

THE EFFECTS OF MONOCULAR OCCLUSION ON INDIVIDUAL

CORTICAL NEURONS

Thus far the changes resulting from monocular deprivation have never been fol-
lowed in individual neurons. All interpretations and conclusions are based on
the comparison of independently analyzed samples of neurons which in most
cases have been recorded from different animals that were raised under different
conditions. We have therefore developed a method for chronic recording which
allows us to determine repeatedly the receptive field properties of cortical neu-
rons in alert and only lightly restrained kittens. This made it possible to study
the effects of monocular deprivation and reverse occlusion in individual cortical
cells (Mioche & Singer, 1989).
The first effect of monocular deprivation was the disappearance of binocular
summation, that is, the superiority of binocular over monocular responses.
Overt changes of ocular dominance have been observed as soon as 6 hr after the
beginning of monocular deprivation and a complete loss of excitatory responses
to deprived eye stimulation was seen as early as 12 hr after occlusion. These
changes in ocularity resulted mainly from a gradual decrease of the excitatory
response to deprived eye stimulation and were only occasionally associated with
a moderate increase of responses to the normal eye. During this shift neurons
preserved their orientation and direction selectivity. Interestingly, stimulation
of the deprived eye continued to evoke inhibitory responses even after excit-
atory responses had vanished completely. The interval required for the mani-
festation of these alterations appeared to be independent of the duration of pre-
ceding visual exposure, suggesting that the duration of visual stimulation
required for the induction of a change is shorter than the time needed for sub-
sequent expression of the change. Neurons that had already been monocular
before deprivation were found resistant to monocular deprivation and showed
changes neither in responsiveness nor in ocularity.
The recovery of responses to the previously deprived eye in the case of reverse
occlusion had a considerably slower time course: at least 24 hr of reverse occlu-
sion was required before the responses to the deprived eye reappeared and
growth of these responses continued for up to 74 hr, the longest period investi-
gated. The first change after reversal was a reduction of the response to the newly
deprived eye. In numerous neurons responses to the newly open eye reappeared
only after the cell had ceased to be excitable from the previously open eye. Thus
these neurons passed through a phase during which they were entirely unre-
sponsive to visual stimulation. In the remaining cells responses to the newly
open eye reappeared before responses to the now deprived eye had disappeared.
During this reversal of ocular dominance all cells preserved their original ori-
entation preference, but there were indications for a reduction of orientation
tuning.
The results of the monocular deprivation experiments suggest several conclu-
sions. First, the downregulation of the deprived afferents cannot be considered
the result of a homeostatic response to the upregulation of inputs from the view-
216 REGULATION OF CORTICAL FUNCTION IN MEMORY

ing eye. Our results rather favor a mechanism that leads to primary inactivation
of deprived inputs. This may either be due to weakening of excitatory or to
strengthening of inhibitory inputs. The observation that inhibitory responses
persisted once the excitatory response had disappeared is compatible with the
latter interpretation. Also compatible is the finding that blockade of GABAergic
inhibition reverses the early effects of monocular deprivation (Sillito, Kemp, &
Blakemore, 1983). However, the alternative—inhibitory connections remain
unmodified while excitatory transmission is weakened—has to be considered as
equally likely.
The second conclusion is that the time required for the manifestation of
changes in ocular dominance does not depend solely on the amount of visual
stimulation but seems to be determined by the time constants of a process that
is independent of experience. This has a bearing on the controversial discussion
about the existence of a consolidation period for the expression of ocular dom-
inance changes. Our data favor the notion of the involvement of two different
processes: an early, vision-dependent induction phase and a subsequent vision-
independent manifestation period. The recent finding of Rauschecker and Hahn
(1987) that anesthesia, if induced within 1 hr after monocular exposure, pre-
vents the manifestation of ocular dominance changes also points in this
direction.
Finally, the results of the study of monocularly deprived kittens showed that
the breakdown of binocular summation precedes the first signs of an ocular
dominance shift. This suggests that binocular cooperativity is a particularly vul-
nerable network property. The earlier reports that short periods of monocular
deprivation led to a reduction of binocular neurons rather than to a shift of ocu-
lar dominance toward the open eye may be relevant in this context (Freeman &
Olson, 1982).
The results of the reversal experiments support the conclusion that the reduc-
tion of transmission in deprived afferents is not a direct consequence of
increased efficiency of afferents from the open eye since the former process
seemed to precede the latter. Differential gain changes in deprived and nonde-
prived afferents do not appear to result from a normalizing process aimed at
keeping the overall excitatory input to a neuron constant. Rather it appears as
if there are two processes which have different time constants and interact with
each other through some still unidentified links: one leads to a rapid inactivation
of deprived afferents while the other promotes a slow increase in the efficacy of
the previously inactivated connections. The latter process appears to take over
only after competing excitatory inputs are weakened.

THE ROLE OF POSTSYNAPTIC ACTIVATION

It has become clear that changes in ocular dominance do not depend solely on
the level of activity in the afferent connections (Cynader & Mitchell, 1977;
 Modification Rules Input selection by coherency matching

FIGURE 10.1. The rules that describe activity-dependent modifications of synaptic connections
in the developing visual cortex. As the rules in the left column indicate, connections consolidate
if the probability is high that presynaptic and postsynaptic elements are active in temporal contiguity
(rule I), while connections destabilize if the probability is high that the presynaptic terminal is
inactive at the same time as the postsynaptic target is activated (rule II). In the absence of post-
synaptic responses there appears to be no selective change in input connection irrespective of the
state of the activation of the presynaptic input (rule III). When applied to conditions where two
inputs converge on the same target (right column) these local rules lead to selective stabilization
of converging inputs that convey correlated activity (condition I) while they lead to competition
between converging inputs if these convey noncorrelated activity (condition II). In this case one
input will consolidate at the expense of the other.
218 REGULATION OF CORTICAL FUNCTION IN MEMORY

Greuel, Luhmann, & Singer, 1988; Singer, Rauschecker, & Werth, 1977). It
rather appears that the critical variable is the state of activation of the postsyn-
aptic neuron and in particular the degree of temporal correlation between pre-
synaptic and postsynaptic activation (Rauschecker & Singer, 1979, 1981). The
use-dependent modifications of excitatory transmission actually seem to follow
rules that closely resemble those postulated by Hebb (1949) for adaptive neu-
ronal connections (Rauschecker & Singer, 1981). As summarized in Figure 10.1,
the basic assumption is that the direction of the change—increase or decrease of
efficacy—depends on the correlation between presynaptic and postsynaptic acti-
vation. The efficacy of excitatory transmission appears to increase when presyn-
aptic afferents and the postsynaptic cell are active in temporal contiguity and to
decrease when the postsynaptic target is active while the presynaptic terminal is
silent. These rules, when applied to circuits where two (or more) afferent path-
ways converge onto a common postsynaptic target cell, have the effect of selec-
tively stabilizing and hence associating pathways that convey correlated activity.
Similarly, these selection criteria lead to competition between converging path-
ways if these convey uncorrelated activity. In that case one subset of afferents is
always inactive while the other is driving the postsynaptic cell. Hence one subset
will increase its gain at the expense of the other. Eventually those afferents that
have the highest probability of being active in temporal contiguity with the post-
synaptic target cell will win.

THE DURATION OF THE INTEGRATION INTERVAL

FOR HEBBIAN MATCHING

The modification rules predict that pathways consolidate only when they are
active in contingency with the postsynaptic cell. For converging pathways, such
as those arriving from the two eyes, this implies that they have to be active
simultaneously if they are to remain connected to a common target cell. As soon
as they get out of phase they will compete with each other and one pathway will
become repressed. We recently attempted to determine the cutoff point beyond
which asynchrony between the activity patterns from the two eyes leads to dis-
ruption of binocularity. For that purpose high-speed solid-state shutters were
mounted in goggles that were fitted to the head of freely moving kittens and
allowed for rapidly alternating monocular occlusion (Altmann, Luhmann,
Greuel, & Singer, 1987). The maximal interval of asynchrony still compatible
with the maintenance of binocular connections was found to be on the order of
200 to 400 msec. This agrees with the psychophysical evidence that the neuronal
representation of a stimulus persists for 300 to 400 msec (Altmann, Eckhorn, &
Singer, 1986); signals can be integrated over such long time spans. The impli-
cation is that the signal which is generated by one eye persists, presumably in
the cortical cells, for a few hundred milliseconds. This excludes individual
action potentials as a basis for contingency matching. The neuronal processes
 Ontogenetic Self-Organization and Learning 219

subserving persistence have not yet been identified. Long-lasting postsynaptic

conductance changes, persistence of second messenger signals, and reverbera-
tion in reciprocally coupled neuron ensembles are likely candidates.

THE ROLE OF CENTRAL GATING MECHANISMS

It is obvious that such activity-dependent selection can solve our specification

problem and stabilize selectively those afferents from the two eyes which origi-
nate from corresponding retinal loci. It is further clear that such experience-
dependent pruning of neuronal circuits is a very powerful and versatile mecha-
nism to establish associations and hence has all the potential to serve as a basis
for learning. Before discussing this fascinating aspect, however, further con-
straints on the experience-dependent developmental process have to be consid-
ered. In our special case selection of afferent connections from the two eyes can
be successful only if it occurs while the animal is actually fixating a nonambi-
guous target with both eyes. Pruning must not take place when the two eyes are
moving in an uncoordinated way. In this latter case, the images processed by
the two eyes are different and hence all retinal signals, even those originating
from corresponding retinal loci, are uncorrelated. All afferents from the two eyes
would therefore compete with one another and the consequence would be com-
plete disruption of binocular connections. The same would be the case if the
spontaneously produced bursts of activity that occur, for example, in the genic-
ulate afferents during certain sleep stages, were capable of inducing changes in
circuitry. Moreover, to assure a sufficient degree of overlap of the images in the
two eyes, some evaluation of the best match between the coarsely prespecified
retinotopic representations ought to be made prior to selection. This requires
preprocessing and control of eye movements and hence an aroused and attentive
brain. The selection process must therefore be gated by nonretinal control sys-
tems capable of determining the instances at which retinal activity may induce
changes in circuitry. In agreement with these postulates it has been found that
nonretinal afferents to striate cortex play a crucial role in gating ocular domi-
nance plasticity.
Even when contour vision is unrestricted and retinal signals readily elicit
responses in the neurons of the visual cortex, vision-dependent modifications of
excitatory transmission fail to occur in a variety of conditions. Thus neurons of
the kitten's striate cortex may remain binocular despite monocular deprivation
when the proprioceptive input from extraocular muscles is disrupted (Buisseret
& Singer, 1983), when the open eye is surgically rotated within the orbit (Singer,
Tretter, & Yinon, 1982), when large-angle squint is induced in both eyes (Singer,
von Grunau, & Rauschecker, 1979), or when strabismus is induced by bilateral
cyclotorsion (Crewther, Crewther, Peck, & Pettigrew, 1980). In these cases con-
tour vision per se is unimpaired but the abnormal eye position and motility lead
to massive disturbances of the kittens' visuomotor coordination. Initially the
inappropriate retinal signals cause abnormal visuomotor reactions and during
220 REGULATION OF CORTICAL FUNCTION IN MEMORY

this period are effective in influencing cortical ocular dominance. Subsequently,

however, the kittens rely less and less on visual cues and develop a near com-
plete neglect of the visual modality. In this phase, retinal signals no longer mod-
ify ocular dominance, and they also fail to support the development of orien-
tation-selective receptive fields.
These results suggest that retinal signals influence the development of cortical
functions only when the animal uses them for the control of behavior. Indeed,
retinal signals do not lead to changes of cortical functions when the kittens are
paralyzed and/or anesthetized while exposed to visual patterns (Buisseret, Gary-
Bobo, & Imbert, 1978; Freeman & Bonds, 1979; Singer, 1979; Singer & Raus-
checker, 1982). Vision-dependent modifications also failed to occur in kittens in
which a sensory hemineglect was induced by unilateral surgical lesions of the
intralaminar nuclear complex of the thalamus (Singer, 1982). Following mon-
ocular deprivation, neurons in the visual cortex of the normal hemisphere had
become monocular, as is usual with monocular deprivation. However, neurons
in the visual cortex of the hemisphere with the lesion had remained binocular.
Thus although both hemispheres had received identical signals from the open
eye, these signals induced the expected modifications in the normal hemisphere
only, remaining rather ineffective in the hemisphere which—because of the
diencephalic lesion—"attended" less to retinal stimulation. Recent experiments
revealed that these effects of the thalamic lesion cannot be attributed to the pro-
jections originating from the intralaminar nuclei alone (Bear, Kleinschmidt, &
Singer, 1988). Cortical plasticity was found unimpaired when the intralaminar
lesions were made by injections of the cytotoxin NMDA, which selectively
destroys neurons but leaves fibers of passage intact. This suggests that the sur-
gical lesions were effective because they comprised both thalamic neurons and
fiber systems traversing the intralaminar nuclei.

THE CHEMICAL NATURE OF PERMISSIVE GATING SIGNALS

Kasamatsu and Pettigrew (1979) showed that neurons of the kitten striate cortex
remain binocular despite monocular deprivation when cortical norepinephrine
(NE) is depleted by local infusion of the neurotoxin 6-hydroxydopamine (6-
OHDA). Since they could demonstrate that microperfusion of the depleted
cortical tissue with NE reinstalls ocular dominance plasticity (Kasamatsu, Pet-
tigrew, & Ary, 1979), these authors proposed that normal NE levels are a nec-
essary prerequisite for ocular dominance plasticity.
Subsequent independent investigations indicated that ocular dominance
changes can be induced despite NE depletion. In these investigations 6-OHDA
was either injected prior to monocular deprivation or the noradrenergic input
to cortex was blocked by means other than local 6-OHDA application (Adrien
et al., 1985; Bear & Daniels, 1983; Bear et al., 1983; Daw, Robertson, Rader,
Videen, & Cosica, 1984; Daw, Videen, Parkinson, & Rader, 1985). Apparent
controversy may have been resolved by the demonstration that ocular domi-
 Ontogenetic Self-Organization and Learning 221

nance plasticity is influenced both by noradrenergic and cholinergic mechanisms

(Bear & Singer, 1986). Ocular dominance plasticity is abolished only if both the
noradrenergic pathway from locus coeruleus and the cholinergic projection from
the basal forebrain are lesioned. Disruption of either system alone is not suffi-
cient to arrest plasticity. These results suggest that both neuromodulators ace-
tylcholine and norepinephrine have a permissive function in cortical plasticity,
either system being capable of replacing the other.

INVOLVEMENT OF NMDA RECEPTOR ACTIVATION AND Ca 21

FLUXES IN USE-DEPENDENT PLASTICITY

The lesion experiments reported above indicated that use-dependent modifica-

tions of synaptic transmission require a certain amount of cooperativity between
retinal input and additional internally generated signals. This suggested that the
process mediating cortical plasticity has a threshold which is reached only when
retinal signals are coincident with additional facilitatory input. As a possible
substrate for such a threshold process the high-threshold voltage-dependent Ca2+
channels were considered. Their voltage threshold is higher than that of the Na+
channels which mediate the regenerative action potential, and they appear to be
located in dendritic rather than in somatic membranes (Llinas, 1979). More-
over, there is evidence from developing Purkinje cells and from binding studies
in the kitten visual cortex (Bode-Greuel & Singer, 1988) that these channels are
particularly numerous in the membranes of developing neurons (Llinas & Sug-
imori, 1979). An attempt was made, therefore, to determine with ion-selective
electrodes whether stimuli known to modify cortical response properties also led
to an activation of Ca2+ channels. The results revealed that stimulation condi-
tions leading to changes in neuronal response properties also caused activation
of Ca2+ fluxes from extracellular to intracellular compartments. Measurable
Ca2+ fluxes occurred only when patterned retinal activity was coincident with
electrical activation of the brain stem reticular core or of the intralaminar nuclei
(Geiger & Singer, 1986). These Ca2+ fluxes were much more pronounced in 4-
to 5-week-old kittens than in adult cats. This is compatible with the hypothesis
that the respective Ca2+ conductances are related to mechanisms involved in
developmental plasticity (see also below).
The hypothesis that Ca2+ currents are involved in activity-dependent long-
term changes of neuronal excitability receives further support from the recent
finding that the activation of the NMDA receptor, a subtype of the receptors for
excitatory amino acids, is also crucial for the induction of ocular dominance
changes (Kleinschmidt, Bear, & Singer, 1987). Infusion of 2-amino-5~phospho-
novaleric acid (APV), a selective antagonist of the NMDA receptor, completely
blocks any experience-dependent modifications of striate cortex functions with-
out abolishing neuronal responses to light. The NMDA receptor is activated by
excitatory amino acids and has a selective affinity for NMDA. It is coupled to
an ionophore that is permeable for Ca2+ ions. A special property of this NMDA-
222 REGULATION OF CORTICAL FUNCTION IN MEMORY

FIGURE 10.2. Functional properties of the

NMDA receptor. The ionophore associated with
the NMDA receptor becomes permeable for cal-
cium ions only if the receptor is occupied by its
endogenous ligand and if the membrane potential
of the postsynaptic neuron is sufficiently depolar-
ized to remove the magnesium block of the NMDA
channel. In condition I, only input A is active and
the membrane potential is not sufficiently positive
to remove the magnesium block. In condition II,
inputs A and B are active simultaneously; mem-
brane depolarization is now sufficient for the re-
moval of the magnesium block, and the NMDA
ionophores become permeable for calcium-ions.
The result is a strong depolarization of the mem-
brane and influx of calcium. By virtue of its voltage
dependence the NMDA mechanism is capable of
evaluating the temporal contiguity of activity in
converging pathways and promoting calcium entry
if there is sufficient cooperativity.

sensitive channel is its voltage dependence (Fig. 10.2). It is activated only when
the receptor is occupied by its endogenous ligand and when the membrane is
sufficiently depolarized. Thus this mechanism is ideally suited to "evaluate" the
contingency of presynaptic and postsynaptic activation. It contributes to the
cells' depolarization only if there is sufficient cooperativity between converging
afferents. Hence the NMDA mechanism could be the substrate for the threshold
that needs to be reached in order to obtain long-term modifications of neuronal
excitability.
Experiments in slices of the visual cortex seem to confirm this interpretation.
They show that the conventional single spike responses to orthodromic stimu-
lation of the optic radiation do not depend on the activation of the NMDA
mechanism. However, NMDA channels seem to be involved in triggering the
high-frequency bursts that may also be evoked by orthodromic stimulation,
especially when stimuli are applied at higher frequency or when postsynaptic
inhibition is reduced by application of bicuculline. These burst discharges are
likely to depend on the activation of Ca2+ conductances and arc blocked by
 Ontogenetic Self-Organization and Learning 223

application of the NMD A receptor blocker APV. In a substantial fraction of cor-

tical cells prolonged application of burst-generating stimuli induces a long-term
enhancement of excitatory transmission that is similar to long-term potentiation
in the hippocampus. When the NMDA receptors are blocked by APV these same
stimuli no longer generate burst discharges and also no longer induce changes in
synaptic transmission (Artola & Singer, 1987). Thus the activation of the
NMDA-dependent excitatory postsynaptic potential (EPSP) and the resulting
burst discharge appear to be necessary conditions for the induction of long-term
modifications. Since both phenomena are most likely associated with postsyn-
aptic Ca2+ fluxes through NMDA-dependent ionophores and/or high-threshold
Ca2+ channels, these results support the hypothesis that activity-dependent
modifications require a surge of intracellular Ca2+ ions. Since APV also blocked
experience-dependent modifications of cortical response properties in vivo, it is
likely that these adaptive processes are based on mechanisms similar to those
observed in vitro.
Supporting evidence for this possibility comes from experiments in which
attempts were made to induce lasting changes of neuronal response properties
by combining retinal stimulation with direct pharmacological modulation of the
excitability of cortical neurons (Greuel et al., 1988). In anesthetized, paralyzed
kittens cortical cells were activated from the retina with light while at the same
time they were exposed to ionophoretically applied pulses of the excitatory
amino acids NMDA and glutamate and the neuromodulators acetylcholine
(ACh) and NE. In more than half of the neurons 30 min of conditioning with
combined light and pharmacological stimulation were sufficient to induce a
selective enhancement of the responses to the light stimulus used for condition-
ing and to produce a concomitant depression of responses to nonconditioned
stimuli. Thus with appropriate light stimulation, drastic changes in ocular dom-
inance and/or orientation and direction preference could be obtained that out-
lasted the end of the conditioning sequence by more than 30 min. These results
support the notion that induction of use-dependent long-term modifications of
response properties requires cooperativity between retinal input and further,
internally generated, facilitatory influences. The permissive effects of NMDA,
particularly, are suggestive of an involvement of Ca2+ currents in these
processes.

CONSTRAINTS FOR THE INDUCTION

OF EXPERIENCE-DEPENDENT MODIFICATIONS

The data reviewed thus far could be used to define some of the constraints of
use-dependent modifications of neuronal connectivity. These are summarized
schematically in Figure 10.3. Local correlations between the activity patterns in
modifiable afferents and their target cells determine the sign of the modification.
Connections increase their gain and/or stabilize if their activity is correlated
with that of their target cell. If presynaptic and postsynaptic activities are not
Factors influencing Neuronal Plasticity
 Ontogenetic Self-Organization and Learning 225

sufficiently contiguous, the respective connection weakens and if this weakening

occurs during a critical period of development, it leads to physical removal of
the synaptic connection. The integration interval during which presynaptic and
postsynaptic activation must coincide in order to lead to stabilization of a path-
way has a duration of maximally 200 to 300 msec. However, the decision as to
whether such local patterns of activity lead to modifications depends on addi-
tional hierarchically organized gating systems. Thus it can be inferred from the
requirement to reach the activation threshold of the NMDA mechanism that a
critical level of postsynaptic depolarization has to be reached. This probability
increases to the extent that excitatory and inhibitory inputs to the same dendrite
are activated and silenced, respectively. Since all inhibitory inputs to cortical
neurons and a large number of nonretinal excitatory afferents originate within
the same and nearby cortical columns, this implies that those stimuli will be
most effective in inducing modifications which match best the resonance prop-
erties of local arrays of excitatory cortical neurons. Since some of the excitatory
input to individual cortical neurons is recruited also from remote cortical and
subcortical centers, it can further be expected that the efficiency of stimuli to
induce modifications of cortical circuitry will increase to the extent that the
stimuli not only match the response properties of local neuron clusters in the
striate cortex, but also conform with the resonance properties of more distrib-
uted neuronal assemblies.
Even if these local and global requirements for cooperativity are fulfilled,
modifications may not occur unless the permissive modulatory systems—the
noradrenergic, the cholinergic, and probably also the intralaminar projections—
are in the appropriate state of activation. These systems in turn receive input
from most centers in the brain and hence can be expected to influence local
adaptive changes as a function of a very global evaluation of the adequacy of
just processed and potentially change-inducing activity. The probability that a
sensory pattern will induce a long-lasting modification of cortical circuitry thus
depends both on the nature of this pattern and the central state of the brain that
is maintained during sensory activation. The change will occur only if there is
sufficient cooperativity among the inputs converging onto the cell whose modi-
fiable inputs are to change; this in turn will be the case only if the pattern of
sensory activation matches sufficiently well with the resonance properties of
both local circuits in striate cortex and more distributed but interconnected neu-
ronal assemblies. The gating of local adaptive changes is thus a distributed func-

FIGURE 10.3. Factors influencing use-dependent modifications of neuronal circuits in the de-
veloping visual cortex. The sign of the change in modifiable connections A and B depends on the
temporal correlation of activity conveyed by the modifiable inputs. Whether a change occurs in
response to this activity depends on the state of activation of additional synaptic inputs to the same
neuron. These comprise feedback loops originating within the same cortical column, the same
cortical area, and probably also remote cortical and subcortical structures. Additional gating inputs
are derived from the more globally organized modulatory systems which originate in central core
structures of the brain.
226 REGULATION OF CORTICAL FUNCTION IN MEMORY

tion in which numerous centers of the brain participate. This distributed nature
of the gating process accounts well for the evidence that so many different inter-
ventions disrupt adaptive changes in response to sensory stimulation. From a
teleological point of view such a multilevel control of plasticity appears indis-
pensable in order to assure that local changes remain adapted to the global
requirements of the developing brain.

FACTORS LIMITING DEVELOPMENTAL PLASTICITY

A characteristic feature of developmental plasticity is its rapid decline with age.

The notion that cortical plasticity is influenced substantially by cholinergic
mechanisms suggested the possibility that the age-dependent decline of mallea-
bility is the result of selective changes in the cholinergic projection to striate
cortex. In this case, one expects a withdrawal of cholinergic afferents from par-
ticular subsets of neurons and/or a removal of muscarinic receptors. These
changes ought to be particularly pronounced in layer IV because the circuit mod-
ifications that lead to ocularity changes occur mainly within this layer. Experi-
mental data support this hypothesis. The density and laminar distribution of
high-affinity binding sites for muscarinic agonists change during development in
a way that suggests a relation to the time course of the critical period (Shaw,
Needier, & Cynader, 1984). However, this redistribution of receptors is not par-
alleled by the laminar pattern of cholinergic afferents. The density of fibers with
CHAT-like immunoreactivity increases with age, and there are no indications
for a selective reduction of cholinergic fibers or terminal boutons from layer IV
(Stichel & Singer, 1987). One reason for this could be that in layer IV, ACh acts
mainly on the nicotinic receptors, which are prominant in this layer and most
probably located on the thalamic afferents (Prusky, Shaw, & Cynader, 1987; Par-
kinson, Kratz, & Daw, 1988). However, a problem with this interpretation is
that so far no axo-axonic synapses have been identified in the neocortex, which
implies that the cholinergic influence on thalamic afferents would have to be
mediated through nonconventional ways. The morphological data thus suggest
that developmental changes of the cholinergic input to striate cortex are not
responsible for the temporal limitation of ocular dominance plasticity. The
interpretation that developmental changes in modulatory systems are not a lim-
iting factor of plasticity is supported by electrophysiological evidence. Combin-
ing ionophoretic application of NE and ACh with patterned light stimuli has
been found rather ineffective in producing substantial modifications of response
properties in neurons of the visual cortex of adult cats, although this procedure
has been effective in kittens (Greuel et al, 1988). Since ACh and NE were as
effective in adult cats as in kittens in facilitating responses to light, it appears
unlikely that the age-dependent decline in plasticity is a result of reduced avail-
ability of permissive modulatory signals. Recent evidence indicates, however,
that a redistribution of calcium-dependent mechanisms might account for
reduced plasticity in the adult. A first and indirect hint for this possibility came
from the observation that it is much more difficult to induce calcium fluxes from
 Ontogenetic Self-Organization and Learning 227

extracellular to intracellular compartments in the visual cortex of adult cats than

kittens.
More direct support for an age-dependent redistribution of calcium-depen-
dent mechanisms comes from investigations of calcium-binding proteins and of
calcium channels. An analysis of developmental changes in the distribution of
the calcium binding proteins parvalbumin and vitamin D~dependent calcium-
binding protein revealed that these calcium-buffering systems are particularly
rich in developing neurons (Stichel, Singer, Heizmann, & Norman, 1987).
We recently obtained evidence that binding sites for organic calcium channel
blockers of the dihydropyridin class are much more frequent in the visual cortex
of kittens than of adults. The highest concentrations of these binding sites have
been observed in layer IV at the peak of the critical period for ocular dominance
plasticity. In good correlation with the age-dependent decline in ocular domi-
nance plasticity, these binding sites disappeared from layer IV and eventually
became restricted essentially to supragranular layers. If one accepts the interpre-
tation that binding sites for organic calcium channel blockers reflect the location
of postsynaptic voltage-dependent high-threshold calcium channels, this obser-
vation suggests that voltage-dependent calcium conductances redistribute dur-
ing development and may even disappear from certain neuronal profiles alto-
gether.
A similar pattern of developmental redistribution has been found for binding
sites with a high affinity for APV, a potent antagonist of the NMDA receptor
(Bode-Greuel & Singer, 1988). These binding sites were also much more fre-
quent in kittens than in adults and more evenly distributed throughout the cor-
tical depth in the former than in the latter. In the adult they were again restricted
essentially to supragranular layers. If one assumes that high-affinity binding sites
for APV reflect NMDA receptors, this would indicate that transmitter-activated
calcium conductances also redistribute during development and become
reduced with age. Both the presumptive reduction of voltage-sensitive calcium
channels and that of transmitter-dependent calcium channels would be compat-
ible with the reduced calcium fluxes in the adult. Taken together, these results
suggest that certain neuronal profiles lose their calcium conductances and hence
are no longer exposed to activity-dependent calcium influx. Since the latter is
likely to serve as a messenger in activity-dependent long-term modifications of
neuronal response properties, removal of calcium conductances could be one
cause for the decline of plasticity with age.

SELF-ORGANIZATION AND LEARNING:

THE BRAIN AS AN EVER-CHANGING ORGAN

One interesting aspect of experience-dependent self-organization processes is

their formal similarity with learning. At a descriptive and perhaps even mech-
anistic level the activity-dependent refinement of connectivity patterns shares
certain characteristic features of associative learning. Two sets of variables, in
our special case the afferents from the two eyes, become permanently associated
228 REGULATION OF CORTICAL FUNCTION IN MEMORY

with a common effector, the cortical target cell. The criterion for this association
is coherency of activation, just as in classical or operant conditioning.
Another similarity between "developmental" self-organization and "adult"
learning is that the activity patterns have to match some predispositions of the
learning organism in order to establish associations. In our case the minimal
requirement was that the retinal signals conform with the response properties of
cortical neurons; that is, the visual stimuli had to contain spatial contrast gra-
dients. The same is true in adult learning. Stimulus configurations have to fulfill
some minimal criteria for saliency and unambiguity in order to be effective.
Finally, there is the parallel that selective associations occur only when the stim-
ulus patterns are contingent with internally generated gating signals that are
available only when the brain is awake and attentive. As has been established in
numerous behavioral experiments, a similar dependence on central states holds
true for adult learning. As elaborated later, comparison with other model sys-
tems of neuronal plasticity suggests that even the molecular mechanisms
involved in developmental self-organization and adult plasticity may be the
same.

THE INVOLVEMENT OF Ca2 + IONS

In all cases where long-term changes of excitability and/or synaptic efficacy can
be induced by neuronal activity, Ca2+ ions appear to serve as messengers trans-
lating membrane depolarization into biochemical signals. Both in Aplysia and
Hermissenda the surge of free Ca2+ in the cytosol appears to function as the
molecular equivalent of the conditional stimulus (Alkon, 1983; Carew, Haw-
kins, Abrams, & Kandel, 1984). The same is likely to be the case in the cerebel-
lum since the complex spike of Purkinje cells which follows climbing fiber acti-
vation is associated with the activation of high-threshold Ca2+ channels (Ito,
Sakurai, & Tongroach, 1982). Long-term potentiation (LTP) in the hippocam-
pus also appears to require free Ca2+ ions in the cytosol of postsynaptic dendrites
(Lynch & Baudry, 1984; Lynch, Larson, Kelso, Barrioneuvo, & Schottler, 1983).
Finally, a Ca2+-dependent process has also been implicated in activity-depen-
dent long-term modifications of excitability of motor cortex neurons (Woody,
1984). The evidence suggesting an involvement of Ca2+ ions in ocular domi-
nance plasticity was reviewed above. Recent evidence suggests that hippocam-
pal LTP is also reduced when NMDA receptors are blocked with APV, thus
adding a rather specific item to the list of similarities between developmental
learning and plasticity (Collingridge, Kehl, & McLennan, 1983).

THE INVOLVEMENT OF NEUROMODULATORS

AND RECEPTOR-DEPENDENT SECOND MESSENGER SYSTEMS

In Aplysia, serotonin has been suggested as the molecular signal mediating the
response to the conditional stimulus. It activates a membrane-bound cyclase
 Ontogenetic Self-Organization and Learning 229

and causes a surge of intracellular cAMP (for review see Carew et al., 1984). An
increase of cAMP also appears as a necessary step in hippocampal LTP (Lynch
& Baudry, 1984). In this case it is unclear whether the rise of cAMP is due to
activation of a receptor-coupled cyclase or whether it is secondary to activation
of a Ca2+-dependent cyclase. Since NE facilitates LTP, the former possibility
deserves consideration (Bliss, Goddard, & Riives, 1983; Hopkins & Johnston,
1984). Finally, NE has also been shown to facilitate the experience-dependent
reorganization of templates in the rabbit olfactory bulb (Gray, Freeman, & Skin-
ner, 1984). The evidence for a permissive role of NE and ACh in ocular domi-
nance plasticity was reviewed above.

THRESHOLD PHENOMENA, TEMPORAL CONTIGUITY,

AND COOPERATIVITY

There are indications from all vertebrate model systems that adaptive modifi-
cations have a threshold which is reached only if a sufficiently strong activation
is achieved. Available evidence suggests that this threshold process is located at
the posysynaptic side. In neurons of the motor cortex, changes in response to
synaptic activation are facilitated when the postsynaptic neuron is depolarized
by intracellular current injections, by antidromic stimulation, or by activation
of additional excitatory inputs (Baranyi & Feher, 1981; Woody, 1984; Woody,
Alkon, & Hay, 1984). In the cerebellum, changes in the efficiency of parallel fiber
synapses required concomitant activation of Purkinje cells by climbing fibers
(Ito et al., 1982). In the hippocampus, LTP is facilitated if converging excitatory
inputs are coactivated simultaneously (Lee, 1983; Levy & Steward, 1983;
McNaughton, Douglas, & Goddard, 1978), if glutamate is applied directly, or if
postsynaptic inhibition is reduced while the afferents that are to be potentiated
are stimulated (for references on cooperativity in LTP see Wigstrom & Gustafs-
son, 1985). All these manipulations are likely to enhance local dendritic depo-
larization. The evidence for threshold processes in ocular dominance plasticity
was reviewed above.
These numerous similarities between developmental and adult plasticity
prompt the hypothesis that learning in the adult might be considered as the con-
tinuation of the self-organizing processes that serve at first to adopt the building
blocks of our nervous system to each other and that subsequently serve to adapt
the whole system to its environment. This viewpoint predicts that there should
be a continuity between early self-organization and adult learning not only at
the level of formal description but also at the level of molecular mechanisms.

REFERENCES

Adrien, J., Blanc, G., Buisseret, P., Fregnac, Y., Gary-Bobo, E., Imbert, M., Tassin, J. P., &
Trotter, Y. (1985). Noradrenaline and functional plasticity in kitten visual cortex: A
reexamination. Journal of Physiology, 367, 73-98.
230 REGULATION OF CORTICAL FUNCTION IN MEMORY

Alkon, D. L. (1983). Learning in a marine snail. Scientific American, 249, 70.

Altmann, L., Eckhorn, R., & Singer, W. (1986). Temporal integration in the visual system:
Influence of temporal dispersion on figure-ground discrimination. Vision Research,
26, 1949-1957.
Altmann, L., Luhmann, H. J., Greuel, J. M., & Singer, W. (1987). Binocularity in kittens with
rapidly and alternating monocular occlusion. Journal of Neurophysiology, 16, 965-
980.
Artola, A., & Singer, W. (1987). Long-term potentiation and NMDA receptors in rat visual
cortex. Nature, 330, 649-652.
Baranyi, A., & Feher, O. (1981). Long-term facilitation of excitatory synaptic transmission in
single cortical neurones of the cat produced by repetitive pairing of synaptic poten-
tials following intracellular stimulation. Neuroscience Letters, 23, 303-308.
Bear, M. F., & Daniels, I. D. (1983). The plastic response to monocular deprivation persists in
kitten visual cortex after chronic depletion of norepinephrine. Journal of Neurosci-
ence, 3, 407-416.
Bear, M. F., Kleinschmidt, A., & Singer, W. (1988). Experience-dependent modifications of
kitten striate cortex are not prevented by thalamic lesions that include the intralam-
inar nuclei. Experimental Brain Research, 70, 627-631.
Bear, M. F., Paradiso, M. A., Schwartz, M., Nelson, S. B., Carnes, K. M., & Daniels, J. D.
(1983). Two methods of catecholamine depletion in kitten visual cortex yield differ-
ent effects on plasticity. Nature, 302, 245-247.
Bear, M. F., & Singer, W. (1986). Modulation of visual cortical plasticity by acetylcholine and
noradrenaline. Nature, 320, 172-176.
Blakemore, C, Garey, L. J., & Vital-Durand, F. (1978). The physiological effects of monocular
deprivation and their reversal in the monkey's visual cortex. Journal of Physiology,
282, 223-262.
Bliss, T. V. P., Goddard, G. V., & Riives, M. (1983). Reduction of long-term potentiation in
the dentate gyrus of the rat following selective depletion of monoamines. Journal of
Physiology, 334, 475-491.
Bode-Greuel, K. M., & Singer, W. (1988). Developmental changes of the distribution of binding
sites for organic Ca 2+ -channel blockers in cat visual cortex. Experimental Brain
Research, 70, 266-275.
Bonds, A. B., Silverman, M. S., Sclar, G., & Tootell, R. B. (1980). Visually evoked potentials
and deoxyglucose studies of monocularly deprived cats. Proceedings of the Associa-
tion for Research in Vision and Ophthalmology, 225.
Buisseret, P., Gary-Bobo, E., & Imbert, M. (1978). Ocular motility and recovery of orientational
properties of visual cortical neurones in dark-reared kittens. Nature, 272, 816-817.
Buisseret, P., & Singer, W. (1983). Proprioceptive signals from extraocular muscles gate expe-
rience-dependent modifications of receptive fields in the kitten visual cortex. Exper-
imental Brain Research, 51, 443-450.
Carew, T. J., Hawkins, R. D., Abrams, T. W., & Kandel, E. R. (1984). A test of Hebb's postulate
at identified synapses which mediate classical conditioning in Aplysia. Journal of
Neuroscience, 4, 1217-1224.
Collingridge, G. L., Kehl, S. J., & McLennan, H. (1983). Excitatory amino acids in synaptic
transmission in the Schaffer collateral-commissural pathway of the rat hippocampus.
Journal of Physiology (London), 334, 33-46.
Crewther, S. G., Crewther, D. P., Peck, C. K., & Pettigrew, J. D. (1980). Visual cortical effects
of rearing cats with monocular or binocular cyclotorsion. Journal of Neurophysiol-
ogy, 44, 97-118.
Cynader, M. (1983). Prolonged sensitivity to monocular deprivation in dark-reared cats: Effects
of age and visual exposure. Developmental Brain Research, 8, 155-164.
Cynader, M., & Mitchell, D. E. (1977). Monocular astigmatism effects on kitten visual cortex
development. Nature, 270, 177-178.
 Ontogenetic Self-Organization and Learning 231

Daw, N. W., Robertson, T. W., Rader, R. K., Videen, T. O., & Cosica, C. J. (1984). Substantial
reduction of noradrenaline by lesions of adrenergic pathway does not prevent effects
of monocular deprivation. Journal of Neuroscience, 4, 1354-1360.
Daw, N. W., Videen, T. O., Parkinson, D., & Rader, R. K. (1985). DSP-4 depletes noradrena-
line in kitten visual cortex without altering the effects of monocular deprivation.
Journal of Neuroscience, 5, 1925-1933.
Freeman, R. D., & Bonds, A. B. (1979). Cortical plasticity in monocularly deprived immobi-
lized kittens depends on eye movement. Science, 206, 1093-1095.
Freeman, R. D., & Olson, C. (1982). Brief periods of monocular deprivation in kittens: Effects
of delay prior to physiological study. Journal of Neurophysiology, 47, 139-150.
Geiger, H., & Singer, W. (1986). A possible role of Ca ++ -currents in developmental plasticity.
Experimental Brain Research, Ser. 14, 256-270.
Gray, C. M., Freeman, W. J., & Skinner, E. (1984). Associative changes in the spatial amplitude
patterns of rabbit olfactory EEG are norepinephrine dependent. Society for Neuro-
science Abstracts, 10, 121.
Greuel, J. M., Luhmann, H. J., & Singer, W. (1987). Evidence for a threshold in experience-
dependent long-term changes of kitten visual cortex. Developmental Brain Research,
34, 141-149.
Greuel, J. M., Luhmann, H. J., & Singer, W. (1988). Persistent changes of single-cell responses
in kitten striate cortex produced by pairing sensory stimulation with ionophoretic
application of neurotransmitters and neuromodulators. In C. D. Woody, D. L.
Alkon, & J. L. McGaugh (Eds.), Cellular mechanisms of conditioning and behavioral
plasticity (pp. 425-436). New York: Plenum.
Hebb, D. O. (1949). The organization of behavior. New York: Wiley.
Hopkins, W. F., & Johnston, D. (1984). Frequency dependent noradrenergic modulation of
long-term potentiation in the hippocampus. Science, 226, 350-351.
Hubel, D. H., & Wiesel, T. N. (1965). Binocular interaction in striate cortex of kittens reared
with artificial squint. Journal of Neurophysiology, 28, 1041-1059.
Ito, M., Sakurai, M., & Tongroach, P. (1982). Climbing fibre induced depression of both mossy
fibre responsiveness and glutamate sensitivity of cerebellar Purkinje cells. Journal of
Physiology, 324, 113-134.
Kasamatsu, T., & Pettigrew, J. D. (1979). Preservation of binocularity after monocular depri-
vation in the striate cortex of kittens treated with 6-hydroxydopamine. Journal of
Comparative Neurology, 185, 139-161.
Kasamatsu, T., Pettigrew, J. D., & Ary, M.-L. (1979). Restoration of visual cortical plasticity
by local microperfusion of norepinephrine. Journal of Comparative Neurology, 185,
163-182.
Kleinschmidt, A., Bear, M. F., & Singer, W. (1987). Blockade of "NMDA" receptors disrupts
experience-dependent plasticity in kitten striate cortex. Science, 238, 355-358.
Lee, K. S. (1983). Cooperativity among afferents for the induction of long-term potentiation in
the CAj region of the hippocampus. Journal of Neuroscience, 3, 1369-1372.
LeVay, S., Stryker, M. P., & Shatz, C. J. (1978). Ocular dominance columns and their devel-
opment in layer IV of the cat's visual cortex: A quantitative study. Journal of Com-
parative Neurology, 179, 223-244.
Levy, W. B., & Steward, O. (1983). Temporal contiguity requirements for long-term associative
potentiation/depression in the hippocampus. Neuroscience, 8, 791-797.
Llinas, R. (1979). The role of calcium in neuronal function. In F. O. Schmitt & F. G. Worden
(Eds.), The Neurosciences fourth study program (pp. 555-571). Cambridge, Mass.:
MIT Press.
Llinas, R., & Sugimori, M. (1979). Calcium conductances in Purkinje cell dendrites: Their role
in development and integration. Progress in Brain Research, 51, 323-334.
Lynch, G., & Baudry, M. (1984). The biochemistry of memory: A new and specific hypothesis.
Science, 224, 1057-1063.
232 REGULATION OF CORTICAL FUNCTION IN MEMORY

Lynch, G., Larson, J., Kelso, S., Barrionuevo, G., & Schottler, F. (1983). Intracellular injections
of EGTA block induction of hippocampal long-term potentiation. Nature, 305, 719-
721.
McNaughton, B. L., Douglas, R. M., & Goddard, G. V. (1978). Synaptic enhancement in fascia
dentata: Cooperativity among coactive afferents. Brain Research, 157, 277-293.
Mioche, L., & Singer, W. (1989). Chronic recordings from single sites of kitten striate cortex
during experience-dependent modification of receptive field properties. Journal of
Neurophysiology, 62.
Mower, G. D., Caplan, C. J., Christen, W. G., & Duffy, F. H. (1985). Dark rearing prolongs
physiological but not anatomical plasticity of the cat visual cortex. Journal of Com-
parative Neurology, 235, 448-466.
Mower, G. D., & Christen, W. G. (1985). Role of visual experience in activating critical period
in cat visual cortex. Journal of Neurophysiology, 53, 572-589.
Olson, C., & Freeman, R. D. (1980). Profile of the sensitive period for monocular deprivation
in kitten. Experimental Brain Research, 39, 17-21.
Parkinson, D., Kratz, K. E., & Daw, N. W. (1988). Evidence for a nicotinic component to the
actions of acetylcholine in cat visual cortex. Experimental Brain Research, 73, 553-
568.
Prusky, G. T., Shaw, C, & Cynader, M. S. (1987). Nicotinic receptors are located on lateral
geniculate nucleus terminals in cat visual cortex. Brain Research, 412, 131-138.
Rauschecker, J. P., & Hahn, S. (1987). Ketamine-xylazine anesthesia blocks consolidation of
ocular dominance changes in kitten visual cortex. Nature, 326, 183-185.
Rauschecker, J. P., & Singer, W. (1979). Changes in the circuitry of the kitten's visual cortex
are gated by postsynaptic activity. Nature, 280, 58-60.
Rauschecker, J. P., & Singer, W. (1981). The effects of early visual experience on the cat's visual
cortex and their possible explanation by Hebb synapses. Journal of Physiology, 310,
215-239.
Shatz, C. J., & Stryker, M. P. (1978). Ocular dominance in layer IV of the cat's visual cortex
and the effects of monocular deprivation. Journal of Physiology, 281, 267-283.
Shaw, C., Needier, M. C., & Cynader, M. (1984). Ontogenesis of muscarinic acetylcholine bind-
ing sites in cat visual cortex: Reversal of specific laminar distribution during critical
period. Developmental Brain Research, 14, 295-299.
Shaw, C., Wilkinson, M., Cynader, M., Needier, M. C, Aoki, C, & Hall, S. E. (1986). The
laminar distributions and postnatal development of neurotransmitter and neuro-
modulator receptors in cat visual cortex. Brain Research Bulletin, 16, 661-672.
Sillito, A. M., Kemp, J. A., & Blakemore, C. (1983). The role of GABAergic inhibition in the
cortical effects of monocular deprivation. Nature, 291, 318-322.
Singer, W. (1979). Central core control of visual cortex functions. In F. O. Schmitt & F. G.
Worden (Eds.), The Neurosciences Fourth Study Program (pp. 1093-1109). Cam-
bridge, Mass.: MIT Press.
Singer, W. (1982). Central core control of developmental plasticity in the kitten visual cortex:
I. Diencephalic lesions. Experimental Brain Research, 47, 209-222.
Singer, W., & Rauschecker, J. P. (1982). Central core control of developmental plasticity in the
kitten visual cortex. II. Electrical activation of mesencephalic and diencephalic pro-
jections. Experimental Brain Research, 47, 223-233.
Singer, W., Rauschecker, J., & Werth, R. (1977). The effect of monocular exposure to temporal
contrasts on ocular dominance in kittens. Brain Research, 134, 568-572.
Singer, W., Tretter, F., & Yinon, U. (1982). Central gating of developmental plasticity in kitten
visual cortex. Journal of Physiology, 324, 221-237.
Singer, W., von Grunau, M., & Rauschecker, J. (1979). Requirements for the disruption of
binocularity in the visual cortex of strabismic kittens. Brain Research, 171, 536-540.
Stichel, C. C., & Singer, W. (1987). Quantitative analysis of the choline acetyltransferase-immu-
noreactive axonal network in the cat primary visual cortex. II. Pre- and postnatal
development. Journal of Comparative Neurology, 258, 99-111.
 Ontogenetic Self-Organization and Learning 233

Stichel, C. C., Singer, W., Heizmann, C. W., & Norman, A. W. (1987). Immunohistochemicai
localization of calcium-binding proteins, parvalbumin and calbindin-D 28 K, in the
adult and developing visual cortex of cats: A light and electron microscopic study.
Journal of Comparative Neurology, 262, 563-577.
Tiemann, S. B. (1985). The anatomy of geniculocortical connections in monocularly deprived
cats. Cellular and Molecular Neurobiology, 5, 35-45.
Wiesel, T. N., & Hubel, D. H. (1965a). Comparison of the effects of unilateral and bilateral eye
closure on cortical unit responses in kittens. Journal of Neurophysiology, 28, 1029-
1040.
Wiesel, T. N., & Hubel, D. H. (1965b). Extent of recovery from the effects of visual deprivation
in kittens. Journal of Neurophysiology, 28, 1060-1072.
Wigstrom, H., & Gustafsson B. (1985). On long-lasting potentiation in the hippocampus: A
proposed mechanism for its dependence on coincident pre- and postsynaptic activity.
Acta Physiologica Scandinavica, 123, 519-522,
Woody, C. D. (1984). The electrical excitability of nerve cells as an index of learned behavior.
In D. L. Alkon & J. Farley (Eds.), Primary neural substrates of learning and behav-
ioral change (pp. 101-127). New York: Cambridge University Press.
Woody, C. D., Alkon, D. L., & Hay, B. (1984). Depolarization-induced effects of Ca2+-calmo-
dulin-dependent protein kinase injection, in vivo, in single neurons of cat motor cor-
tex. Brain Research, 321, 192-197.
11
The Dissection by Alzheimer's Disease
of Cortical and Limbic Neural Systems
Relevant to Memory
GARY W. VAN HOE SEN

Since the mid-1970s, Alzheimer's disease (AD) has been recognized as the most
common form of dementia (Katzman, 1976). The clinical manifestations of AD
are characterized by an unremitting decline in higher mental function, domi-
nated typically by a memory impairment. However, the tempo of the changes
vary and early onset AD is associated with a rapid decline (Corsellis, 1976; Tom-
linson & Corsellis, 1984). Historically, this contributed to the confusion sur-
rounding the term presenile dementia. Both this and senile dementia are now
considered behaviorally and pathologically as AD, which recognizes that
dementia can occur abnormally early, but that neither dementia nor extensive
brain pathology is a necessary concomitant of senility. This is not to say that
behavioral changes and structural changes to the brain do not occur in aging; of
course they do, but an aberrant pace in terms of cognitive decline and an aber-
rant quantity of structural changes are abnormal, and together they form much
of the essence of AD (Katzman, 1986).
Behaviorally, the cardinal change in AD is an impairment in memory, and it
characterizes the disorder throughout its course. Aside from a relatively well pre-
served capacity for skill memory (Eslinger & Damasio, 1986; Knopman & Nis-
sen, 1987) until late in the illness, few details are available regarding the course
and character of the memory disorder. Episodic or contextual memory is clearly
impaired and some believe that changes in their anterograde compartment are
among AD's first manifestations (Van Hoesen & Damasio, 1987). Forgetting
appointments, new names, new obligations, and news events are all examples.
Episodic or contextual memory loss in the retrograde compartment is also com-
mon and in many instances seems to be the trigger that sufficiently alarms the
victim to seek a medical opinion. An inability to remember a well-known route
or the face of a relative or good friend is an example of such a change.

234
 Alzheimer's Disease, Neural Systems, and Memory 235

Disorders of generic or semantic memory are seen frequently in AD, espe-

cially in its later stages. The conceptual class to which a stimulus belongs—
including the simple identity and function of common natural and man-made
objects—might be lost. For practical purposes, this disability is an agnosia
(Damasio, 1984).
Although this may strike a familiar chord with clinicians who examine
patients with AD, or families and health care professionals who attend to their
needs, the exact nature and sequence of the changes have not been studied fully.
Similarly, the course of the memory changes has not been documented fully. For
example, it is tempting to believe that the initial memory changes occur in the
anterograde compartments of memory and entail episodic or contextual events
and excessive forgetfulness. Later one might expect an involvement of the ret-
rograde compartments of memory in the episodic or contextual realms, so that
previously familiar stimuli, their relationships, and even previously overlearned
spatial aspects of memory are impaired. As incapacitation approaches, generic
and semantic memory are affected typically so that the identity and function of
common environmental objects are lost, as is the patient's self-identity.
If this is the sequence of behavioral changes that characterize AD, it has
important implications that would lead to predictions about the loci of patho-
logical changes that occur as the illness progresses. Theoretically, these would be
verifiable in longitudinal studies of AD where all data points, both behavioral
and pathological, throughout the course of the disease were represented. For
example, in the early stages where the memory impairment is for episodic or
contextual events in the anterograde compartment, one might predict a focus of
pathology in temporal lobe structures such as the hippocampus, amygdala, and
parahippocampal gyrus (Scoville & Milner, 1957). Such changes would also be
compatible with diencephalic pathology (Squire, 1986, 1987). Some of these
changes—particularly those involving excessive forgetfulness—might also be
related to pathological changes in basal forebrain structures (Damasio, Graff-
Radford, Eslinger, Damasio, & Kassell, 1985).
The neural correlates for episodic and contextual memory impairments in the
retrograde compartment and impairments for generic and semantic memory are
not entirely understood. Clearly, neither depends exclusively on medial tem-
poral or medial diencephalic structures. Damage to these areas can produce a
partial retrograde change for events prior to the injury or surgery, but it is either
incipient or lasts for only a few years (Damasio, 1984; Scoville & Milner, 1957;
Squire, 1986). Recent evidence from survivors of herpes simplex encephalitis
suggests that the posterior parahippocampal, lateral temporal, temporal polar,
and anterior insular cortices may play a role in the preservation of retrograde
memories (Damasio, Eslinger, Damasio, Van Hoesen, & Cornell, 1985).
Finally, generic memory is seldom affected with medial temporal injury even
if it is bilateral. One would expect that the additive features in AD, which lead
to a total breakdown in memory mechanisms, including all aspects of episodic
and contextual memory plus generic and semantic memory, would be a conse-
quence of additional pathological changes in the various association cortices,
236 REGULATION OF CORTICAL FUNCTION IN MEMORY

and particularly the multimodal areas of the frontal, parietal, and temporal lobes
(see Damasio, 1984; Goldman-Rakic, 1987; Squire, 1986; Tulving, 1987).
Given our knowledge of the neural correlates of memory, derived from ani-
mal experimentation and the study of human patients with surgical, accidental,
or infectious damage to selective parts of the brain, predictions can be made
about the anatomical changes expected in AD. Although this would necessarily
be skewed in the direction of end-stage AD where all forms of episodic and
generic memory are impaired, the neural correlates of separate forms of memory
might be highlighted in patients who die before end-stage AD of other illnesses.
Thus a careful analysis of the specific cortical and subcortical changes in AD is
warranted along with an analysis of the known neural connectivity of affected
areas. However, before doing this it is essential to highlight certain aspects of
the neural connectivity of the primate brain and, even more, the brain of a non-
human primate like the rhesus monkey, since this species provides our best
approximation to humans in terms of experimental knowledge.

SOME BASICS OF CORTICAL CONNECTIVITY IN PRIMATES

Experimental studies in nonhuman primates since the late 1960s have revealed
an orderly sequence of intrahemispheric corticocortical neural systems that link
the primary sensory cortices with both the association cortices and limbic sys-
tem structures. These form an intricate series of multisynaptic neural pathways
that provide both unimodal and multimodal sensory processing and that pro-
vide limbic structures with a refined digest of cortical activity (Van Hoesen,
1982).
The primary sensory areas all give rise to powerful, short intrahemispheric
corticocortical connections to the adjacent primary sensory association cortex
(Jones & Powell, 1970; Pandya & Kuypers, 1969). This connection begins a feed-
forward system (Maunsell & Van Essen, 1983; Van Essen & Maunsell, 1983)
that characterizes the visual (Rockland & Pandya, 1981; Van Essen & Maunsell,
1983), auditory (Galaburda & Pandya, 1983), and somatosensory cortices (Pan-
dya & Seltzer, 1982b). It is the first step in the dissemination of sensory infor-
mation to other parts of the cortex. After successive relays in other sensory asso-
ciation areas, these projections end in the nonprimary association cortex of the
forntal, temporal, and parietal lobes. Many also have as a target multimodal
association areas, in that their input is characterized by projections from mul-
tiple primary sensory association cortices each committed functionally to a sin-
gle modality (Chavis & Pandya, 1976; Jones & Powell, 1970; Pandya & Seltzer,
1982a; Seltzer & Pandya, 1976). Another target of the association cortices,
including those of a multimodal variety, is the limbic system (Goldman-Rakic,
Selemon, & Schwartz, 1984; Jones & Powell, 1970; Van Hoesen, Pandya, & But-
ters, 1972). For example, nearly the entire limbic lobe (posterior orbital and
medial frontal cortex, cingulate cortex, parahippocampal cortex, temporal polar
 Alzheimer's Disease, Neural Systems, and Memory 237

cortex, and anterior insular cortex) receives as a major input projections from
the association cortices. Additional direct limbic projections from the associa-
tion cortices converge on several nuclei of the amygdala (Aggleton, Burton, &
Passingham, 1980; Herzog & Van Hoesen, 1976; Turner, Mishkin & Knapp,
1980; Whitlock & Nauta, 1956). Together with powerful projections from the
cortex of the limbic lobe (Van Hoesen, 1982), these provide the amygdala with
a sizable cortical input. One of the final limbic system linkages of the neural
systems of the association cortices is with the allocortex that forms the hippo-
campal formation. These are mediated via projections from the limbic lobe and
association cortex to the entorhinal cortex, as illustrated in Figure 11.1 (Amaral,
1987; Amaral, Insausti, & Cowan, 1983; Goldman-Rakic et al, 1984; Van Hoe-
sen, 1982; Van Hoesen & Pandya, 1975a, 1975b; Van Hoesen et al., 1972; Van
Hoesen, Pandya, & Butters, 1975). This area (Brodmann's area 28) gives rise to
the perforant pathway that terminates along the distal dendrites of the dentate
gyrus granule cells and hippocampal pyramidal cells (Van Hoesen, 1985; Van
Hoesen & Pandya, 1975a, 1975b; Van Hoesen et al., 1975).
Although less information is available, it is clear that many of these connec-
tions are reciprocated by a feedback series of neural systems (Galaburda & Pan-

FIGURE 11.1. Major cortical afferent and efferent neural systems of the entorhinal cortex in the
rhesus monkey on medial and lateral views of the cerebral hemisphere. Identification of cortical
areas corresponds to those of Brodmann and Bonin and Bailey. Note that cortical neural systems
from the frontal, parietal, occipital, temporal, and limbic lobes converge on Brodmann's area 28,
the entorhinal cortex. Layer IV of the entorhinal cortex receives a powerful output from the
subiculum (Sub) and CA1 parts of the hippocampal formation (HP) and gives rise to neural systems
that feed back to widespread limbic and association cortical areas.
238 REGULATION OF CORTICAL FUNCTION IN MEMORY

dya, 1983; Maunsell & Van Essen, 1983; Pandya & Yeterian, 1985; Rockland &
Pandya, 1981; Van Essen & Maunsell, 1983). To conceptualize this, recall that
the major output of the hippocampal formation arises from its subicular sub-
division. This cortex has several projections to subcortical structures (Swanson
& Cowan, 1975, 1977) and to other parts of the limbic lobe, including a strong
one to the deeper cell layers (layer IV) of the entorhinal cortex (Rosene & Van
Hoesen, 1977; Shipley & Sorensen, 1975). The cortical areas of the limbic lobe
all project to various parts of the association cortices, which in turn project back
to the various primary sensory cortices (Mesulam, Van Hoesen, Pandya, &
Geschwind, 1977; Pandya, Van Hoesen, & Mesulam, 1981; Vogt & Miller,
1983).
In short, corticocortical intrahemispheric pathways clearly form neural sys-
tems that serve to transmit sensory information that ultimately reaches the old-
est of cortices, the allocortex of the hippocampal formation. On the other hand,
output from the allocortex begins a reciprocation of these projections that even-
tually can be traced back to the primary sensory area (Van Hoesen, 1982). Some
of these relationships are illustrated in Figures 11.1 and 11.2. In this context it

FIGURE 11.2. Major efferent neural systems of the hippocampal formation. The major cortical
target is the entorhinal cortex (EC). Both the subiculum and entorhinal cortex are major targets for
neurofibrillary tangles in AD. CS = collateral sulcus; FF = flmbria fornix; HF = hippocampal
fissure; ITS = inferior temporal sulcus; PAS = parasubiculum; PSB = presubiculum;
UHP = uncal hippocampus.
 Alzheimer's Disease, Neural Systems, and Memory 239

is also important that recent investigations have shown that the amygdala also
gives rise to widespread projections to the association cortices as well (Amaral
& Price, 1984; Porrino, Crane, & Goldman-Rakic, 1981).
The cells of origin that give rise to intrahemispheric neural systems are not
entirely known. However, available evidence (see Andersen, Asanuma, &
Cowan, 1985; Bugbee & Goldman-Rakic, 1983; Goldman-Rakic & Schwartz,
1982; Jones, 1981, 1984; Maunsell & Van Essen, 1983; Rockland & Pandya,
1979; Schwartz & Goldman-Rakic, 1982) indicates that the feedforward system,
involved in the outflow of sensory information and other intrahemispheric pro-
jections, arises from both supragranular and infragranular cortical layers,
whereas the feedback system arises largely from infragranular layers. More spe-
cifically, the small pyramids that form layer II, many of the medium-sized pyr-
amids of layer III, and a subset of the larger pyramids of layer V give rise to the
feedforward system of corticocortical connections. In contrast, the feedback sys-
tem arises largely from layers V and VI (Fig. 11.3).

TOPOGRAPHICAL DISTRIBUTION OF PATHOLOGICAL CHANGES

IN ALZHEIMER'S DISEASE IN RELATION TO CORTICAL TYPES

The observation of severe pathological changes in the cerebral cortex in Biel-

chowsky-stained tissue was the major pathologic feature that motivated Alzhei-
mer (1907) to publish his now classic report on "a singular disease of the cerebral
cortex." His findings have never been disputed seriously. Cortical changes quan-
titatively overshadow those in subcortical areas, although it is clear that the cor-
tex is not affected uniformly. All types of cortex are affected in AD, but the
changes within a given type of cortex are selective. A review of the major types
of cortex and their involvement in AD follows.

FIGURE 11.3. Cortical laminar origin (triangles)

for feedback and feedforward corticocortical neural
systems and laminar distribution of their terminals
within the cortex (dots). Neurofibrillary tangles in
AD are found in great abundance in layers III and V
and neuritic plaques are seen often in layers I and III.
There is a strong suggestion that corticocortical as-
sociation neural systems are compromised in AD.
240 REGULATION OF CORTICAL FUNCTION IN MEMORY

Allocortical Changes
The allocortices are a conspicuous component of the cerebral cortex. They are
composed of two to three layers, of which one or two are cellular. Pyramidal
neurons are the dominant cell type. Characteristically, polarized apical dendrites
extend toward the pia mater and form a wide molecular layer. The dentate gyrus,
hippocampus, subiculum, olfactory, and periamygdaloid cortices are examples.
In AD these fields are heavily but selectively compromised (Ball, 1976, 1977,
1978; Hyman, Damasio, Van Hoesen, & Barnes, 1984; Kemper, 1978; Wilcock,
1983). For example, within the hippocampal formation, the pryamids that form
the subiculum and CA1 sector of the hippocampus are usually invested heavily
with neurofibrillary tangles. Neuritic plaques are found among them and in their
molecular layer (Burger, 1983; Gibson, 1983). In contrast, the adjoining CA3
sector of the hippocampus and the dentate gyrus may contain only an occasional
neuron with neurofibrillary tangles (Hyman et al., 1984). There is also evidence
that the olfactory allocortices are damaged heavily in AD (Esiri & Wilcok, 1984).

Periallocortical Changes
The periallocortical parts of the cerebral cortex are characterized by atypical
laminar arrangements and cell types. The deep layers are related closely to the
adjacent allocortex, whereas the superficial layers are reminiscent of the isocor-
tex. The entorhinal, parasubicular, presubicular, and retrosplenial cortices are
the best examples of periallocortex. As with allocortices, the involvement of the
periallocortices in AD is not uniform. For example, the entorhinal cortex is
involved heavily with many neurofibrillary tangles and neuritic plaques (Hirano
& Zimmerman, 1962; Hyman et al., 1984; Van Hoesen, Hyman, & Damasio,
1986). Typically the changes are so extensive that the normal and unique cyto-
architecture of this cortex is unrecognizable. The same may be said for the adja-
cent and closely related parasubicular cortex. In contrast, the adjacent presu-
bicular cortex is largely spared in AD (Hyman, Van Hoesen, Kromer, &
Damasio, 1985). It may contain an occasional neuritic plaque and neurofibril-
lary tangle, but these are in random locations without a predilection for a specific
layer or cell type.

Proisocortical Changes
The proisocortices, a sizable component of the cerebral cortex, along with the
periallocortices and allocortices form the limbic lobe. The posterior parts of the
orbitofrontal cortex, the cingulate cortex, the retrocalcarine, posterior parahip-
pocampal, perirhinal, temporal polar, and anterior insular cortices constitute
major examples. These areas resemble the isocortex in that there are multiple
cellular layers. However, some of these may be atypical, because they are accen-
tuated or incipient, and generally layers V and VI contain large and poorly dif-
 Alzheimer's Disease, Neural Systems, and Memory 241

ferentiated neurons. These areas are the most uniformly affected cortices in AD.
For example, the posterior part of the parahippocampal cortex (Brodmann's
area 36) is a common site of pathological change and contains an abundance of
neurofibrillary tangles and neuritic plaques. These are continuous into the more
anteriorly located perirhinal cortex (Brodmann's area 35) and the temporal
polar cortex (Brodmann's area 38) (Braak & Braak, 1985; Kemper, 1978).
The dysgranular part of the insular cortex and the posterior orbital cortex are
likewise involved to a major degree, as is the subcallosal part of the cingulate
gyms (Brodmann's area 25). Some investigators report that the posterior part of
the cingulate gyrus (Brodmann's area 23) is one of the more heavily affected
cortical areas in AD (Brun & Englund, 1981). However, there is a major excep-
tion regarding the anterior parts of this gyrus (Brodmann's area 24). This cortex
does not appear affected until later in the disorder and may be spared entirely
(Brun & Englund, 1981; Brun & Gustafson, 1976). Interestingly, this part of the
proisocortices is related closely to the motor cortices (supplementary, premotor,
and primary motor cortices), which typically are spared in AD.

Isocortical Changes
The nonuniform involvement of the isocortex in AD often can be detected by
visual inspection of the gross brain (Tomlinson & Corsellis, 1984). In brief,
pathological changes in the form of neurofibrillary tangles, neuritic plaques, and
cell loss are abundant in the association areas of the temporal and parietal lobes
in many Alzheimer's brains, and additional changes in the frontal and occipital
lobes are common. In contrast, the primary sensory areas (Fig. 11.4) and the

FIGURE 11.4. Cross section of

the pattern and laminar distribution
of neurofibrillary tangles in the au-
ditory cortex from a case of AD
charted with an X-Y recorder cou-
pled to the axes of the microscopic
stage. Each dot represents a tangle.
Note that the primary auditory cor-
tex (Brodmann's areas 41 and 42)
that forms Heschl's gyrus is largely
spared. The dorsal and lateral parts
of area 22, the primary sensory as-
sociation cortex, are also relatively
spared. More distal auditory asso-
ciation areas in the upper bank of
the superior temporal sulcus con-
tain extensive pathology.
A

B
 Alzheimer's Disease, Neural Systems, and Memory 243

primary motor cortices are not affected greatly. The relative sparing of these sen-
sory and motor areas is consistent with the paucity of motor and sensory impair-
ment in AD. Nonetheless, the heavy and possibly early involvement of the olfac-
tory cortex in AD (Esiri & Wilcock, 1984) suggests that odor detection is an
exception (Corwin, Serby, Conrad, & Rotrosen, 1985; Doty, Reyes, & Gregor,
1987; Warner, Peabody, Flattery, & Tinklenberg, 1986).
The association cortices affected in AD are of the so-called homotypical vari-
ety (Pearson, Esiri, Hiorns, Wilcock, & Powell, 1985; Rogers & Morrison, 1985).
In the frontal lobe the homotypical cortical areas form Brodmann's areas 9, 10,
11, 12, 13, 14, and 46. In the parietal lobe, they form Brodmann's areas 7, 39,
and 40. In the occipitotemporal region, Brodmann's area 37 contains the major
expanse of homotypical isocortex. Areas 20, 21, and 22 represent the major
examples of this type of cortex in the temporal lobe (Fig. 11.5). These regions
are characterized by a similarity of architectural plan and form the largest part
of the cerebral cortex in the human brain. They represent the classic six-layered
pattern of isocortex. Layer II is well developed and has small neurons, mostly
of the pyramidal variety. Layer III is generally wide and only slightly differen-
tiated, with the major cell type being medium-sized pyramids. Layer IV is com-
posed of small neurons of pyramidal or stellate variety. Layer V is composed
largely of pyramidal neurons of varying sizes with prominent apical dendrites
and is distinguishable from layer VI, which is formed by multiform elements of
varying size and shape. These areas give rise to a majority of the long cortico-
cortical association projections that mediate interconnectivity between the var-
ious lobes and with the limbic cortical areas (Jones & Powell, 1970; Pandya &
Kuypers, 1969). They participate prominently in both feedforward and feedback
cortical connections (see Fig. 11.3).
The degree of pathological involvement of the association cortices in AD can
vary greatly, ranging from extensive involvement to only selective involvement
in one or two lobes. However, the involvement of the proisocortex of the limbic
lobe, the periallocortex of the anterior parahippocampal gyrus, and the allocor-
tex of the hippocampal formation is more consistent. Whether this represents
early involvement of these types of cortex or a differential degree of intensity
once the disease process starts is not yet apparent. Among isocortical association
areas the parietotemporal and occipitotemporal regions seem affected heavily in
AD (Brun & Englund, 1981).

FIGURE 11.5. Two X-Y chartings of neurofibrillary tangle distribution through the anterior (A)
and middle (B) parts of the temporal lobe in AD. Brodmann's numbers denote auditory (22) and
visual (20, 21, and 37) association areas. In each case these are so-called distal association areas
and are known to be involved in complex sensory processing. The cortex that lines the banks of
the superior temporal sulcus is multimodal and is also associated with higher order functions. Note
the layer III and V laminar pattern of tangles and their dense distribution in the distal and multimodal
association cortices.
 Alzheimer's Disease, Neural Systems, and Memory 245

LAMINAR SPECIFICITY AND PATTERNS OF CORTICAL PATHOLOGY

IN ALZHEIMER'S DISEASE

The topographical specificity of Alzheimer's pathology in terms of cortical types

is paralleled by a cellular and laminar specificity within affected areas (Figs. 11.5
and 11.6). This is clearly the case for neurofibrillary tangles. For example, neu-
rons that form predominantly one or two cortical layers contain this cytoskeletal
alteration, whereas their immediate neighbors in adjacent layers may be unal-
tered (Braak & Braak, 1985; Hyman et al., 1984; Lewis, Campbell, Terry, &
Morrison, 1987; Pearson et al., 1985; Van Hoesen et al., 1986). The same applies
to some extent for neuritic plaque distribution (Rogers & Morrison, 1985).
Although this pathological marker may have a broader distribution in terms of
laminar location (Pearson et al., 1985), there are many examples in the cerebral
cortex where plaques are found largely in selective cortical layers. It is less clear
whether cell loss, in addition to that attributable to neurofibrillary tangles, fol-
lows a laminar-specific pattern (Mann, Yates, & Marcyniuk, 1985; Terry, 1983;
Terry, Peck, DeTeresa, Schechter, & Horoupian, 1981), but it appears that larger
cortical neurons are lost to a greater degree than small neurons in both AD and
normal aging. Because large neurons are located primarily in layers III, V, and
VI, it would follow that at least some laminar specificity occurs in age-related
cell loss whether AD is present or not.
It is apparent that pathological changes in AD are decidedly skewed toward
the association and limbic cortices. As summarized in Figure 11.1, homotypical
cortical areas that form the association cortices are a major source of input to
the limbic system (Van Hoesen, 1982) and represent a major recipient of limbic
system output (Amaral & Price, 1984; Kosel, Van Hoesen, & Rosene, 1982; Por-
rino et al., 1981; Van Hoesen, 1982). Moreover, they provide a major source of
long cortical projections that link together the various parts of the cerebral cor-
tex. Their destruction of the magnitude noted in AD would suggest that devas-
tating cognitive change should ensue. Morever, the laminar distribution of cor-
tical pathology in AD correlates closely with the laminar origin of cortical
projections and the laminar distribution of cortical association efferents.
The critical nature of these changes can be appreciated if one considers the
connectional relationship between the hippocampal formation and the cerebral
cortex. The hippocampus receives few direct inputs from cortical association
areas. Instead, its major cortical input arises from the entorhinal cortex (Brod-
mann's area 28), which forms the anterior parahippocampal gyrus. The ento-
rhinal cortex gives rise to the perforant pathway, a major neural system of the
temporal lobe, which terminates on the outer parts of the dendritic trees of den-

FIGURE 11.6. (left) Nissl-stained cross section through the cortex of the inferior temporal gyrus
in AD. (right) The same microscopic field (note the vessel track in the lower left corner of each
photo at arrowhead) with the tissue processed for Alz-50 immunoreactivity. This monoclonal
antibody recognizes a tissue antigen found in high concentrations in certain somas and their proc-
esses in AD but not in age-compatible normals. Note the high density of Alz-50 immunoreactive
somas in layers III, V, and VI.
246 REGULATION OF CORTICAL FUNCTION IN MEMORY

tate gyrus granule cells and hippocampal pyramidal cells. It has been stressed
since the time of Ramon y Cajal's original description of the perforant pathway
in Golgi-stained material that it carries the strongest source of input to the hip-
pocampal formation and forms its exclusive link to the remainder of the cortex.
The input to the entorhinal cortex thus becomes a critical issue and many of its
sources have been described recently (see Amaral, 1987; Van Hoesen, 1982).
Most afferents arise from neighboring cortical areas such as the presubiculum,
parasubiculum, periamygdaloid, and prepiriform cortex and proisocortical areas
of the temporal lobe. The latter includes area 35, the perirhinal cortex; area 36,
the posterior parahippocampal cortex; area 37, the occipitotemporal cortex; and
area 38, the temporal polar cortex. All of these areas are the recipients of pow-
erful corticocortical association input from sensory association and multimodal
cortical association areas and in turn project powerfully to the superficial layers
of the entorhinal cortex. The critical relay neurons in areas 35, 36, 37, and 38
are the pyramids in layers III and V. Briefly, it is probably accurate to state that
a sizable component of the feedforward system of cortical sensory connections
converges onto the entorhinal cortex, where it is relayed to the hippocampal
formation by the perforant pathway.
Nearly all of the key components of these neural systems are affected in AD.
For example, in areas 35, 36, 37, and 38, the most common distribution of neu-
rofibrillary tangles is in cortical layers III and V. Neuritic plaques are abundant
in their terminal zone in the entorhinal cortex layers I to III, and particularly
layer III. Moreover, the cells of origin for the perforant pathway, the large layer
II stellate neurons and the more superficial layer III pyramids, are laden with
neurofibrillary tangles (Figs. 11.7 and 11.8) and often are destroyed totally in
AD (Hyman, Van Hoesen, Kromer, & Damasio, 1986). There is a strong like-
lihood that the specific cortical neurons that form neural systems conveying cor-
tical association input to the hippocampal formation are destroyed in AD. The
reciprocal of this relationship also seems compromised greatly (Hyman et al.,
1984). This feedback system arises primarily from the subicular and CA1 part
of the hippocampal formation and from layer IV of the entorhinal cortex, which
itself receives a large subicular projection (see Figs. 11.1 and 11.2). Both the
subiculum and layer IV of the entorhinal cortex typically contain neurofibrillary
tangles in AD. Such observations led researchers to state previously that the
pathological distribution of neurofibrillary tangles and neuritic plaques in AD
dissects neural systems of the temporal lobe with a cellular precision that effec-
tively isolates the hippocampal formation from the remainder of the cerebral
cortex (Hyman et al., 1984; Van Hoesen & Damasio, 1987; Van Hoesen et al.,
1986).
In conclusion, there seems to be little doubt that the cerebral cortex is dam-
aged heavily in AD, but in a very selective manner. In the allocortex, it takes
the form of selective manner. In the allocortex, it takes the form of selective
destruction of the subicular and CA1 zones of the hippocampal formation but
with preservation of the neighboring CA3 zone. In the periallocortex it can be
seen by conspicuous alterations of the entorhinal (see Fig. 11.7) and parasu-
FIGURE 11.7. Thioflavin S-stained neurofibrillary tangles in layer II of the entorhinal cortex in
AD. These neurons give rise to major component of the perforant pathway that links the cortex
with the dentate gyrus of the hippocampal formation.

FIGURE 11.8. Alz-50 terminal immunoreactivity in the outer two thirds of the molecular layer
of the dentate gyrus in an area that would correspond to the terminal zone of the perforant pathway.
This pattern of immunoreactivity suggests that the AD antigen recognized by Alz-50 is located in
the terminals of layer II entorhinal neurons. Note the presence of Alz-50 immunoreactive neuritic
plaques in the immunoreactive zone. The vessel profile marks the location of the hippocampal
fissure. The granule cells of the dentate gyrus, or stratum granulosum (SG), have been stained
with thionin.
248 REGULATION OF CORTICAL FUNCTION IN MEMORY

bicular cortex but preservation of the adjacent presubicular periallocortex. In the

proisocortex, Brodmann's cingulate areas 25 and 23 may be heavily damaged.
In the isocortex, the homotypical association cortices of all four lobes may con-
tain cell loss, neurofibrillary tangles, and neuritic plaques, whereas adjacent pri-
mary sensory and motor cortices remain well preserved.
Imposed on these selective changes is a remarkable degree of laminar speci-
ficity within affected areas, such that often only certain layers contain neurofi-
brillary tangles or an unusually large number of neuritic plaques. This is best
demonstrated in the entorhinal cortex, where layers II and IV are damaged heav-
ily, whereas neighboring layers are less heavily damaged. It is also manifested in
the homotypical isocortical association areas, where neurofibrillary tangles occur
primarily in cell layers III and V (see Figs. 11.4 to 11.6).
On the basis of studies of cortical connectivity in nonhuman primates such as
the monkey, it is plausible to argue that the pathological targets of AD especially
affect the long association systems that are responsible for linking the various
association areas of the hemisphere among themselves and with the cortex of
the limbic lobe. Both the feedforward systems of connections that disseminate
cortical information throughout the cortex and the feedback systems that recip-
rocate them are implicated (see Fig. 11.3). Indeed, the laminar origins and some
of the terminal fields of these systems correspond remarkably with the patterns
of pathology observed in the cortex in AD (Hyman et al., 1984; Lewis et al.,
1987; Pearson et al., 1985; Rogers & Morrison, 1985).
Many questions can be asked about the patterns of cortical pathology in AD.
The most obvious is whether the changes occur as invariant features of the dis-
order, and they seem to, provided that the illness lasts long enough. In such
cases, or end-stage AD, one might expect a major lesion of the cortical associa-
tion and related limbic connections combined with an overall devastating
impairment of cognition. Variations in both pathology and cognitive impair-
ment occur for victims who die of other causes at earlier stages after onset of the
illness (Wilcock & Esiri, 1982). The rate and amount of accrual of pathological
change also vary. Certain neural systems are more involved than others. Natu-
rally, different patterns of cognitive impairment should be expected in relation
to such differential pathological involvement. The question of whether there is
a single focus for the initiation of the process, or whether it starts in different
parts of the nervous structure simultaneously, should be asked. Observations
suggest that there is a single early focus that resides in the temporal lobe, partic-
ularly in the parahippocampal gyrus and hippocampal formation. The isolated,
early changes in episodic and contextual memory, which represent the hallmark
of the disease, are the cognitive counterpart to this pathological change.

ANATOMICAL ALTERATIONS IN SUBCORTICAL STRUCTURES

IN ALZHEIMER'S DISEASE

Although more lesions in AD involve the cortex than other regions, subcortical
changes should not be overlooked. Well-studied cognitive abnormalities have
 Alzheimer's Disease, Neural Systems, and Memory 249

been reported with purely subcortical lesions, and they can affect processes such
as memory (Damasio, 1984; von Cramon, Hebel, & Schuri, 1985). In addition,
many of the subcortical structures implicated in AD provide sources of input
for the cortex. Other affected subcortical structures include critical relays in neu-
ral systems that convey limbic and hypothalamic influences directly to the
cortex.

Telencephalic Structures
Among subcortical telencephalic areas, those that form the basal forebrain are
clearly implicated in AD. They include the gray matter masses located deep to
the neostriatum and globus pallidus and those that converge along the midline
at the base of the septum pellucidum. Structures such as the septum, the nuclei
of the horizontal and vertical limbs of the diagonal bands of Broca, the substan-
tia innominata and its associated nucleus basalis of Meynert, and the amygdala
are major parts. Anatomically these are characterized by reciprocal connections
with at least one part of the preoptic-hypothalamic area and reciprocal connec-
tions with at least one part of the cerebral cortex. In this sense they serve as
intermediaries between those parts of the brain that largely subserve and interact
with the internal environment and those that more prominently interact with
the external environment.

Nucleus Basalis of Meynert

The nucleus basalis of Meynert is a dispersed group of hyperchromatic neurons
that lies in a position ventral to the striatum and globus pallidus. The neurons
are multipolar and fusiform in shape and are distributed broadly in both medio-
lateral and anterior-posterior directions. In the mediolateral plane they occupy
a position a few millimeters from the midline anteriorly to a position lying
beneath the temporal limb of the anterior commissure posteriorly, where the
neurons are situated dorsal to the amygdala and ventral to the globus pallidus.
In the anterior-posterior plane they extend from the septum anteriorly to the
substantia nigra posteriorly. At some levels, scattered neurons of the nucleus
basalis may be noted within the internal and external medullary laminae of the
globus pallidus and within the lateral hypothalamic area. Although the neurons
of the nucleus basalis of Meynert are dispersed, clustering does occur and can
be observed in certain loci (Mesulam, Mufson, Levey, & Wainer, 1983). Some
of the larger clusters lie within the so-called substantia innominata.
The nucleus basalis of Meynert has attracted attention because of several fac-
tors. During the 1980s, with the advent of retrograde tracing procedures, several
investigators observed that the neurons that form the nucleus basalis send axons
to much of the cerebral cortex and especially to the somatomotor cortices,
including Brodmann's areas 6, 4, 3, 1, 2, and 5 (Divac, 1975; Jones, Burton,
Saper, & Swanson, 1976; Kievet & Kuypers, 1975; Mesulam & Van Hoesen,
1976; Pearson, Gatter, Bratal, & Powell, 1983). Using combined retrograde
labeling and histochemical methods, Mesulam and Van Hoesen (1976) dem-
onstrated that many of the neurons that form the nucleus baasalis of Meynert
250 REGULATION OF CORTICAL FUNCTION IN MEMORY

contain cholinergic enzymes. Subsequent studies using antibodies directed

against choline acetyltransferase left little doubt that nucleus basalis neurons
provide the major source of cholinergic input to the cerebral cortex (Mesulam
et al., 1983). Thus it was of interest to find that some neurons of the basal fore-
brain are greatly diminished in certain patients with AD (Arendt, Bigl, Arendt,
& Tennstedt, 1983; Candy et al., 1983; Whitehouse, Price, Clark, Coyle, &
DeLong, 1981; Wilcock, Esiri, Bowen, & Smith, 1983). This pathological change
offered a convenient structural correlate for the results of previous neurochem-
ical studies revealing decreased levels of cortical acetylcholine in cerebral cortex
in AD (Davies & Maloney, 1976). Moreover, on first impression this association
seemed compatible with the memory impairments of AD because of the well-
known effects of anticholinergic drugs on memory (Bartus, Dean, Beer, & Lippa,
1982; Coyle, Price, & DeLong, 1983; Drachman, 1977; Smith & Swash, 1978).
There is little doubt that the loss of neurons in the nucleus basalis of Meynert
is a notable feature in some cases of AD. Nevertheless, the extent of pathology
in this region can be as variable as that in other vulnerable brain areas (Arendt
et al., 1983; Mann, Yates, & Marcyniuk, 1984; Pearson et al., 1984), and many
cases show no change. It may range from mere shrinkage of perikaryon (Pearson
et al., 1985; Pearson, Gatter, & Powell, 1983; Sofroniew, Pearson, Eckenstein,
Cuello, & Powell, 1984), to minimal cell loss, to extensive cell loss (Saper, Ger-
man, & White, 1985). Neurofibrillary tangles may be present or absent, and the
number of neuritic plaques can vary widely. The issues here are not unlike those
raised for any other affected brain region: What contribution does the loss or
diminution of cholinergic input to the cerebral cortex make to the cognitive
changes in AD, and why are these neurons targeted when their immediate neigh-
bors survive? The latter is of special interest because of the cholinergic nature
and connections of these neurons and the fact that neighboring striatal neurons
and brain stem and spinal cord motor neurons, also acetylcholine positive, do
not seem affected in AD.
It is important to consider nucleus basalis pathology from a neural systems
viewpoint, because it figures prominently in many networks that undoubtedly
participate in a wide range of cognitive processes. Virtually nothing is known
about the subcortical projections of the nucleus basalis, although work in non-
human primates suggests that these may be widespread (Divac, 1975). The cor-
tical projections of this nucleus raise even greater questions. These projections
are organized topographically so that they innervate functionally diverse parts
of the cortex (Mesulam et al., 1983). Focal pathological changes in the nucleus
basalis might be correlated with one cognitive change more than another. How-
ever, some of the larger, more conspicuous clusters of nucleus basalis neurons
project largely to somatosensory and motor cortices (Jones et al., 1976). These
areas are some of the last cortices to be affected in AD; moreover, their destruc-
tion in other neurological diseases is not associated with changes in other than
somatomotor performance. Additionally, insofar as a relation to cognitive sys-
tems is concerned, the large hyperchromatic neurons that partially form the ver-
tical limb of the diagonal band of Broca also are magnocellular basal forebrain
 Alzheimer's Disease, Neural Systems, and Memory 251

nerve cells similar to if not identical to those that form the nucleus basalis of
Meynert (Jones et al., 1976). As is the case for the closely related cholinergic
neurons of the medial septum, they have projections directed not to the isocor-
tex but preferentially to the proisocortex, periallocortex, and allocortex of the
limbic lobe. An especially strong input is directed toward the hippocampal for-
mation via the fimbria fornix and forms the well-known septohippocampal cho-
linergic system. Evidence suggests that this pathway is affected in AD (Arendt
et al., 1983; Nakano & Hirano, 1982, 1983), and in such cases the hippocampal
formation would be deprived of another major afferent source. This is of special
interest because infarcts in this region cause a specific impairment of memory
(Damasio, 1985; Volpe & Hirst, 1983).
Unlike many neural systems of the cortex, direct reciprocity of connections
does not characterize the nucleus basalis. Although the nucleus has extremely
widespread and topographically organized cortical projections, it receives input
from only a subset of the cortical areas to which it projects. These inputs are
largely derived from those parts of the limbic lobe that are located in the tem-
poral lobe, anterior insula, and medial and orbital parts of the frontal lobe
(Mesulam & Mufson, 1984). Thus, for instance, nucleus basalis output, or feed-
back, to the visual cortex is reciprocated only after the entire feedforward
sequence of corticocortical outflow from sensory cortex to the limbic lobe is
traced.

Amygdala
The amygdala is the largest basal forebrain structure. The various nuclei of the
amygdala receive powerful cortical projections from all types of cortex including
the visual and auditory association cortex along the lateral part of the temporal
lobe. Other nuclei are more closely associated with proisocortices of the limbic
lobe. The superficial nuclei, the so-called cortical amygdaloid nuclei, receive
prominent projections from the olfactory and hippocampal allocortices. Perial-
locortical projections exist but have a limited distribution (Van Hoesen, 1981).
In nearly all instances, these consist of reciprocal connections that feed back to
parts of the cortex that originally projected to individual amygdaloid nuclei
(Amaral & Price, 1984; Porrino et al., 1981). Prominent subcortical input arises
from the hypothalamus, periaqueductal gray, peripeduncular nucleus, ventral
tegmental area, supramammillary nucleus, and various midline thalamic nuclei.
The involvement of the amygdala in AD has been studied by several research-
ers (Burger, 1983; Herzog & Kemper, 1980; Kemper, 1983), although little
emphasis has been given to the patterns of pathology with regard to individual
amygdaloid nuclei. Neutritic plaques can be found in at least some parts of the
structure, as can neurofibrillary tangles. Major cell loss has been noted in the
superficial amygdaloid nuclei. It would be desirable to correlate the distribution
of Alzheimer pathology with the neural systems that enter and leave the amyg-
dala. The cortical elements of these pathways are partially understood in non-
human primates. The origins of subcortical afferents are known, but their ter-
minal patterns are not yet fully defined. The subcortical efferent projections of
252 REGULATION OF CORTICAL FUNCTION IN MEMORY

the amygdala have been well studied in many mammals, and a substantial
amount is known regarding intrinsic amygdaloid connections. The role that
amygdaloid pathology plays in the cognitive defects of AD is not known, but
this structure is certainly related to memory at least in animals (Mishkin, 1978,
1982) and unquestionably is involved in many types of motivated and emo-
tional behavior. The knowledge of amygdaloid anatomy in higher mammals
invites correlative studies with the patterns of pathology in AD to gain insight
into the altered behavior of the victims of this disorder. A host of putative trans-
mitters and modulators has been localized in the amygdala, but little is known
about their fate in AD.

Basal Ganglia
Other subcortical components of the telencephalon (the caudate nucleus, puta-
men, and claustrum) do not escape the pathological processes of AD (Rudelli,
Ambler, & Wisniewski, 1984). For example, claustrum abnormalities in the
form of both neurofibrillary tangles and neuritic plaques have been noted by
several investigators. Recent studies have demonstrated that the caudate
nucleus and putamen contain neuritic plaques in AD. These occur in a mosa-
iclike pattern that is reminiscent of the pattern of certain afferents to these struc-
tures as well as the distribution of some receptors.

Diencephalic Structures
Unlike most other components of the neuraxis, the diencephalon has been
thought to be relatively spared in AD. It is difficult to assess the accuracy of this
opinion, because few systematic investigations have addressed the issue. More-
over, of those reports in which these areas have been examined, it is difficult to
appreciate the relative density of the associated abnormalities. In the absence of
more evidence to the contrary, apparently both the hypothalamus and thalamus
are generally spared in AD. Abnormalities may affect these structures but almost
certainly in lesser density than in limbic and cortical areas. When structural
changes do affect the diencephalon in AD, they are distributed in a nonrandom
pattern confined to certain cells in selective nuclei.
In the hypothalamus only a subset of the many hypothalamic nuclei has been
implicated, including the tuberal, anterior, paraventricular, dorsal, and lateral
nuclei. The mammillary complex and lateral preoptic area have also been impli-
cated (Ishii, 1966; Mann et al, 1985). A similar selectivity occurs for the thala-
mus. Included are the ventral anterior, reticular, ventral lateral, dorsomedial,
lateral posterior, lateral dorsal, centromedian, parafascicular, paracentral, and
central lateral nuclei (Rudelli et al., 1984). This array of nuclei is interesting
from many viewpoints. First, the changes entirely spare the major lemniscal
relay nuclei such as the medial and lateral geniculate nuclei and the ventral pos-
terior complex. Second, several nuclei that are damaged have direct and recip-
rocal connections with the limbic cortices. Third, the intralaminar and midline
 Alzheimer's Disease, Neural Systems, and Memory 253

thalamic nuclei seem affected, as does the reticular nucleus, the major associa-
tion nucleus of the thalamus.

Mesencephalic and Myelencephalic Structures

Brain stem abnormalities involving the mesencephalon and myelencephalon
have been reported in some individuals with AD (Ishii, 1966; Rossor, 1981).
The changes include reduced cell numbers and neurofibrillary tangles in selected
nuclei known to give rise to major monoaminergic neural systems that innervate
cerebral cortex.

Raphe Nuclei
Several investigations report abnormalities affecting the raphe complex, the
principal source of central serotonergic projections, with a sixfold increase in
neurofibrillary tangles when compared with age-matched controls (Curcio &
Kemper, 1984). At least in the dorsal raphe nucleus, it appears that these changes
may not be paralleled by a corresponding reduction of serotonin-producing neu-
rons. The percentage of raphe neurons invested with neurofibrillary tangles
appears to be less than 5%, although somewhat larger estimates have been
obtained (Yamamoto & Hirano, 1985).

Locus Coeruleus
The locus coeruleus, a major focus of noradrenergic neurons for the CNS, is also
damaged in some individuals with AD (Bondareff, Mountjoy, & Roth, 1982;
Forno, 1978; Mann, Yates, & Hawkes, 1982; Mann et al., 1984; Tomlinson,
Irving, & Blessed, 1981). This small pigmented nucleus, located in the pontine
tegmentum, has widespread cortical and subcortical projections; the cortical
projections have an especially broad distribution. Both the loss of noradrenergic
neurons and the presence of neurofibrillary tangles have been noted in the
nucleus locus coeruleus and are thought to correlate closely with a reduction in
noradrenergic cortical markers in AD (Perry & Perry, 1986). A recent report of
special interest shows that the locus coeruleus is not uniformly affected in AD
except for that part which projects to the temporal lobe (Marcyniuk, Mann, &
Yates, 1986).

CONCLUSIONS

Recent interest in the neuroanatomical correlates of AD has revealed many new

facts. Clearly, the pathological correlates of this disorder are multicentric and
are not found entirely in cortical areas. Discrete subcortical areas are affected
but, curiously, they (locus coeruleus, raphe complex, and nucleus basalis of Mey-
nert) are those nuclei that have close connectional relationships with the cortex.
The sensory relay and motor nuclei of the thalamus are exceptions in this regard
254 REGULATION OF CORTICAL FUNCTION IN MEMORY

and are not affected, but some association and midline nuclei are. The latter in
fact are those thalamic nuclei that project to the hippocampus and to other areas
of the limbic lobe.
Additionally, the pathological changes in AD are not distributed randomly in
all parts of the cortex. Instead, the association and limbic cortical areas are tar-
geted. The temporo-occipital and temporoparietal cortices and nearly the
entirety of the limbic lobe except area 24 are especially vulnerable. Within these
areas there is a further refinement such that only selected laminae and fields of
the cortex contain evidence of pathology. For example, in the allocortices such
as the hippocampal formation, the dentate gyrus, CA3 pyramids, and the pre-
subicular cortices are largely spared, whereas the subicular and parasubicular
cortices are heavily damaged. In the periallocortex, like the entorhinal area, lay-
ers II and IV contain abundant neurofibrillary tangles, whereas other layers are
spared. In the proisocortex and isocortex, layers III and V are targeted and in
many instances show alterations in Nissl-stained material. In contrast, other cor-
tical areas and laminae are seldom affected. There appears to be a strong likeli-
hood that the affected neurons give rise to corticocortical and corticolimbic neu-
ral systems.
The results reviewed here reiterate the fact that AD is largely a disease of the
cerebral cortex and related subcortical systems connected intimately with it.
However, the cerebral cortex is by no means a uniform and homogeneous struc-
ture. Differential projections in relation to cytoarchitecture have been known for
decades and in recent years it has become clear that different laminae give rise
to different association and corticofugal neural systems. This focuses a new light
on the cortical pathology of AD and offers opportunities to view these changes
in the perspective of behavioral abnormalities as well as experimental studies
regarding cortical connectivity in higher primates.
In AD the most persistent, and perhaps earliest, changes occur in the hippo-
campal formation and parahippocampal gyrus. In fact, there is abundant evi-
dence from a variety of methods to suggest that the temporal lobe is targeted
early in the disease and perhaps overall more heavily than others, excluding, of
course, the limbic lobe. Later in the illness the parietal and frontal lobes develop
pathology consistent with AD.
In many regards the initial focus and pathological course of AD correlate well
with the behavioral changes observed in this disorder. For example, one of the
earliest changes in AD occurs in the hippocampus, particularly its subicular sub-
division. Other changes early in the illness affect the anterior parahippocampal
cortices such as the entorhinal cortex. These lesions disconnect the hippocampal
formation from the association cortices and probably underlie the early signs of
AD such as changes in the anterograde compartments of episodic and contextual
memory. Later changes that affect the lateral temporal, posterior parahippocam-
pal, temporopolar, and anterior insular cortices further exacerbate the problem
and contribute to retrograde memory abnormalities in the episodic and contex-
tual memory realms. The involvement of the multimodal association cortices
would seem to mark the point in AD when generic and semantic memory
 Alzheimer's Disease, Neural Systems, and Memory 255

decline and where even self-identity is lost. Thus it is suggested that AD does
indeed dissect human memory into distinct compartments. It is, as many lay-
men can tell us, a disease of memory and it highlights the cortical and subcor-
tical neural systems that play a fundamental role in this cognitive ability. Since
its cellular correlates can be identified readily, it provides a rare opportunity to
understand the neural correlates of a human behavior in cellular terms.

Acknowledgment
Supported by NS 14944, PO NS 19632, and the Mathers Foundation.

REFERENCES

Aggleton, J. P., Burton, J. .1., & Passingham, R. E. (1980). Cortical and subcortical afferents to
the amygdala of the rhesus monkey (Macaca mulatto). Brain Research, 190, 347-
368.
Alzheimer, A. (1907). Uber eine eigenartige Erkrankung der Hirnrinde. Allgemeine Zeitschrift
fur Psychiatric, 64, 146-148.
Amaral, D. G. (1987). Memory: Anatomical organization of candidate brain regions. In F.
Plum, (Ed.), Handbook of physiology: Sec. 1. The nervous system (pp. 211-294).
Bethesda, Md.: American Physiological Society.
Amaral, D. G., Insausti, R., & Cowan, W. M. (1983). Evidence for a direct projection from the
superior temporal gyrus to the entorhinal cortex in the monkey. Brain Research, 275,
263-277.
Amaral, D. G., & Price, J. L. (1984). Amygdalo-cortical projections in the monkey (Macaca
fascicularis). Journal of Comparative Neurology, 230, 465-496.
Andersen, R. A., Asanuma, C, & Cowan, W. M. (1985). Callosal and prefrontal association
projecting associational projecting cell populations in area 7a of the macaque mon-
key: A study using retrogradely transported fluorescent dyes. Journal of Comparative
Neurology, 232, 443-455.
Arendt, T., Bigl, V., Arendt, A., & Tennstedt, A. (1983). Loss of neurons in the nucleus basalis
of Meynert in Alzheimer's disease, paralysis agitans and Korsakoff's disease. Ada
Neuropathologica, 61, 101-108.
Armstrong, D. M., Le Roy, S., Shields, D., & Terry, R. D. (1985). Somatostatin-like immuno-
reactivity within neuritic plaques. Brain Research, 338, 71-79.
Ball, M. J. (1976). Neurofibrillary tangles and the pathogenesis of dementia: A quantitative
study. Neuropathology and Applied Neurobiology, 2, 395-410.
Ball, M. J. (1977). Neuronal loss, neurofibrillary tangles and granulevacuolar degeneration in
the hippocampus with ageing and dementia. Acta Neuropathologica, 37, 11-18.
Ball, M. J. (1978). Topographic distribution of neurofibrillary tangles and granulovacuolar
degeneration in hippocampal cortex of ageing and demented patients: A quantitative
study. Acta Neuropathologica, 42, 73-80.
Bartus, R. T., Dean, R. L., Ill, Beer, B., & Lippa, A. S. (1982). The cholinergic hypothesis of
geriatric memory dysfunction. Science, 217, 408-4173.
Bondareff, W., Mountjoy, C. Q., & Roth, M. (1982). Loss of neurons of adrenergic projection
to cerebral cortex (nucleus locus ceruleus) in senile dementia. Neurology, 32, 164-
169.
Braak, H., & Braak, E. (1985). On areas of transition between entorhinal allocortex and tem-
porary isocortex in the human brain. Normal morphology and lamina-specific
pathology in Alzheimer's disease. Acta Neuropathologica, 68, 325-332.
256 REGULATION OF CORTICAL FUNCTION IN MEMORY

Brun, A., & Englund, E. (1981). Regional pattern of degeneration in Alzheimer's disease: Neu-
ronal loss and histopathological grading. Histopathology, 5, 549-564.
Brun, A., & Englund, E. (1986). Brain changes in dementia of Alzheimer's type relevant to new
imaging diagnostic methods. Progress in Neuropsychopharmacology and Biological
Psychiatry, 10, 297-308.
Brun, A., & Gustafson, L. (1976). Distribution of cerebral degeneration in Alzheimer's disease.
Acta Psychiatrica Nervenkrankheiten, 223, 15-33.
Bugbee, N. M. & Goldman-Rakic, P. S. (1983). Columnar organization of corticocortical pro-
jections in squirrel and rhesus monkeys: Similarity of column width in species of
differing cortical volume. Journal of Comparative Neurology, 220, 355-364.
Burger, P. C. (1983). The limbic system in Alzheimer's disease. In R. Katzman (Ed.), Banbury
report. Biological aspects of Alzheimer's disease (Vol. 15, pp. 37-44). Cold Spring
Harbor, N.Y.: Cold Spring Harbor Laboratory.
Candy, J. M., Perry, R. H., Perry, E. K., Irving, D., Blessed, G., Fairbairn, A. F., & Tomlinson,
B. E. (1983). Pathological changes in the nucleus of Meynert in Alzheimer's and Par-
kinson's diseases. Journal of Neurological Sciences, 59, 277-289.
Chavis, D. A., & Pandya, D. N. (1976). Further observations on corticofrontal connections in
the rhesus monkey. Brain Research, 117, 369-386.
Corsellis, J. A. N. (1976). Aging and the dementias. In W. Blackwood & J. A. N. Corsellis (Eds.),
Greenfield's neuropathology (pp. 796-848). London: Edward Arnold.
Corwin, J., Serby, M., Conrad, P., & Rotrosen, J. (1985). Olfactory recognition deficit in Alzhei-
mer's and Parkinsonian dementias. IRCS Medical Science Library Compendium, 13,
260.
Coyle, J. T., Price, D. L., & DeLong, M. R. (1983). Alzheimer's disease: A disorder of cortical
cholinergic innervation. Science, 219, 1184-1190.
Curcio, C. A., & Kemper, T. (1984). Nucleus raphe dorsalis in dementia of the Alzheimer type:
Neurofibrillary changes and neuronal packing density. Journal of Neuropathology
and Experimental Neurology, 43, 359-368.
Damasio, A. R. (1984). The anatomic basis of memory disorders. Seminars in Neurology, 4,
223-225.
Damasio, A. R. (1985). The frontal lobes. In K. Heilman & E. Valenstein (Eds.), Clinical neu-
ropsychology (pp. 339-375). New York: Oxford University Press.
Damasio, A. R., Eslinger, P. J., Damasio, H., Van Hoesen, G. W., & Cornell, S. (1985). Mul-
timodal amnesic syndrome following bilateral temporal and basal forebrain damage.
Archives of Neurology, 42, 252-259.
Damasio, A. R., GrafF-Radford, N., Eslinger, P. J., Damasio, H., & Kassell, N. (1985). Amnesia
following basal forebrain lesions. Archives of Neurology, 42, 263-271.
Davies, P., & Maloney, A. J. F. (1976). Selective loss of central cholinergic neurons in Alzhei-
mer's disease (letter to the editor). Lancet, 2, 1403.
Divac, I. (1975). Magnocellular nuclei of the basal forebrain project to neocortex, brain stem,
and olfactory bulb: Review of some functional correlates. Brain Research, 93, 385-
398.
Doty, R. L., Reyes, P. F., & Gregor, T. (1987). Presence of both odor identification and detec-
tion deficits in Alzheimer's disease. Brain Research Bulletin, 18, 597-600.
Drachman, D. A. (1977). Memory and cognitive function in man: Does the cholinergic system
have a specific role? Neurology, 27, 783-790.
Esiri, M. M., & Wilcock, G. K. (1984). The olfactory bulbs in Alzheimer's disease. Journal of
Neurology Neurosurgery and Psychiatry, 47, 56-60.
Eslinger, P. J., & Damasio, A. R. (1986). Preserved motor learning in Alzheimer's disease. Jour-
nal of Neuroscience, 6, 3006-3009.
Forno, L. S. (1978). The locus coeruleus in Alzheimer's disease. Journal of Neuropathology and
Experimental Neurology, 37, 614.
Galaburda, A. M., & Pandya, D. N. (1983). The intrinsic architectonic and connectional orga-
nization of the superior temporal region of the rhesus monkey. Journal of Compar-
ative Neurology, 221, 169-184.
 Alzheimer's Disease, Neural Systems, and Memory 257

Gibson, P. H. (1983). Form and distribution of senile plaques seen in silver impregnated sec-
tions in the brains of intellectually normal elderly people and people with Alzheimer-
type dementia. Neuropathology and Applied Neurobiology, 9, 379-389.
Goldman-Rakic, P. S. (1987). Circuitry of the primate prefrontal cortex and the regulation of
behavior by representational memory. In F. Plum (Ed.), Handbook of physiology:
Sec. 1. The nervous system: Vol. 5. Higher functions of the brain (pp. 373-417).
Bethesda, Md.: American Physiological Society.
Goldman-Rakic, P. S., & Schwartz, M. L. (1982). Interdigitation of contralateral and ipsilateral
columnar projections to frontal association cortex in primates. Science, 216, 755-
757.
Goldman-Rakic, P. S., Selemon, L. D., & Schwartz, M. L. (1984). Dual pathways connecting
the dorsolateral prefrontal cortex with the hippocampal formation and parahippo-
campal cortex in the rhesus monkey. Neuroscience, 12, 719-743.
Goodman, L. (1953). Alzheimer's disease. A clinicopathologic analysis of twenty-three cases
with a theory on pathogenesis. Journal of Nervous and Mental Disease, 117, 97-
130.
Herzog, A. G., & Kemper, T. L. (1980). Amygdaloid changes in aging and dementia. Archives
of Neurology, 37, 625-629.
Herzog, A. G., & Van Hoesen, G. W. (1976). Temporal neocortical afferent connections to the
amygdala in the rhesus monkey. Brain Research, 115, 57-69.
Hirano, A., & Zimmerman, H. M. (1962). Alzheimer's neurofibrillary changes. Archives of Neu-
rology, 7, 73-88.
Hyman, B. T., Damasio, A. R., Van Hoesen, G. W., & Barnes, C. L. (1984). Alzheimer's dis-
ease: Cell specific pathology isolates the hippocampal formation. Science, 225, 1168-
1170.
Hyman, B. T., Van Hoesen, G. W., & Damasio, A. R. (1987). Alzheimer's disease: Glutamate
depletion in the hippocampal perforant pathway zone. Annals of Neurology, 22, 37-
40.
Hyman, B. T., Van Hoesen, G. W., Kromer, L. J., & Damasio, A. R. (1985). The subicular
cortices in Alzheimer's disease: Neuroanatomical relationships and the memory
impairment. Society for Neuroscience Abstracts, 11, 458.
Hyman, B. T., Van Hoesen, G. W., Kromer, L. J., & Damasio, A. R. (1986). Perforant pathway
changes and the memory impairment of Alzheimer's disease. Annals of Neurology,
20, 472-481.
Ishii, T. (1966). Distribution of Alzheimer's neurofibrillary changes in the brain stem and hypo-
thalamus of senile dementia. Acta Nenropathologica, 6, 181-187.
Jamada, M., & Mehraein, P. (1968). Verteilungsmuster der senilen Veranderunger in Gehirn.
Archive fur Psychiatric and Nervenkrankheiten, 211, 308-324.
Jones, E. G. (1981). Anatomy of cerebral cortex: Columnar input-output organization. In F.
O. Schmitt, F. G. Worden, G. Adelman, & S. G. Dennis (Eds.), The organization of
the cerebral cortex (pp. 199-236). Cambridge, Mass.: MIT Press.
Jones, E. G. (1984). Laminar distribution of cortical efferent cells. In A. A. Peters & E. G. Jones
(Eds.), Cerebral cortex: Vol. L Cellular components of the cerebral cortex (pp. 521-
548). New York: Plenum.
Jones, E. G., Burton, H., Saper, C. B., & Swanson, L. W. (1976). Midbrain, diencephalic and
cortical relationships of the basal nucleus of Meynert and associated structures in
primates. Journal of Comparative Neurology, 167, 385-420.
Jones, E. G., & Powell, T. P. S. (1970). An anatomical study of converging sensory pathways
within the cerebral cortex of the monkey. Brain, 93, 793-820.
Katzman, R. (1976). The prevalence and malignancy of Alzheimer's disease: A major killer.
Archives of Neurology, 33, 217-218.
Katzman, R. (1986). Alzheimer's disease. New England Journal of Medicine, 314, 964-973.
Kemper, T. L. (1978). Senile dementia: A focal disease in the temporal lobe. In K. Nandy (Ed.),
Senile dementia: A biomedical approach (pp. 105-113). New York: Elsevier/North-
Holland.
258 REGULATION OF CORTICAL FUNCTION IN MEMORY

Kemper, T. L. (1983). Organization of the neuropathology of the amygdala in Alzheimer's dis-

ease. In R. Katzman (Ed.), Banbury report: Vol. 15. Biological aspects of Alzheimer's
disease (pp. 31-35). Cold Spring Harbor, N.Y.: Cold Spring Harbor Laboratory.
Kievet, J., & Kuypers, H. G. J. M. (1975). Basal forebrain and hypothalamic connections to
the frontal and parietal cortex in the rhesus monkey. Science, 187, 660-662.
Knopman, D. S., & Nissen, M. J. (1987). Implicit learning in patients with probable Alzhei-
mer's disease. Neurology, 37, 784-788.
Kosel, K. C, Van Hoesen, G. W., & Rosene, D. L. (1982). Nonhippocampal cortical projections
from the entorhinal cortex in the rat and rhesus monkey. Brain Research, 244, 201 -
214.
Lewis, D. A., Campbell, M. J., Terry, R. D., & Morrison, J. H. (1987). Laminar and regional
distributions of neurofibrillary tangles and neuritic plaques in Alzheimer's disease: A
quantitative study of visual and auditory cortices. Journal of Neuroscience, 7, 1799-
1808.
Mann, D. M. A. (1985). The neuropathology of Alzheimer's disease: A review with pathoge-
netic, aetiological and therapeutic considerations. Mechanisms of Ageing and Devel-
opment, 31, 213-255.
Mann, D. M. A., Yates, P. O., & Hawkes, J. (1982). The noradrenergic system in Alzheimer
and multi-infarct dementias. Journal of Neurology Neurosurgery and Psychiatry, 45,
113-119.
Mann, D. M. A., Yates, P. O., & Marcyniuk, B. (1984). A comparison of changes in the nucleus
basalis and locus coeruleus in Alzheimer's disease. Journal of Neurology Neurosur-
gery and Psychiatry, 47, 201-203.
Mann, D. M. A., Yates, P. O., & Marcyniuk, B. (1985). Some morphometric observations on
the cerebral cortex and hippocampus in presenile Alzheimer's disease, senile demen-
tia of Alzheimer type and Down's syndrome in middle age. Journal of the Neurolog-
ical Sciences, 69, 139-159.
Marcyniuk, B., Mann, D. M. A., & Yates, P. O. (1986). Loss of nerve cells from locus coeruleus
in Alzheimer's disease is topographically arranged. Neuroscience Letters, 64, 247-
252.
Maunsell, J. H. R., & Van Essen, D. C. (1983). The connections of the middle temporal visual
area (MT) and their relationship to a cortical hierarchy in the macaque monkey.
Journal of Neuroscience, 3, 2563-2586.
Mesulam, M.-M., & Mufson, E. J. (1984). Neural inputs into the nucleus basalis of the sub-
stantia innominata (CH4) in the rhesus monkey. Brain Research, 107, 253-274.
Mesulam, M.-M., Mufson, E. J., Levey, A. L, & Wainer, B. H. (1983). Cholinergic innervation
of cortex by the basal forebrain: Cytochemistry and cortical connections of the septal
area, diagonal band nucleus basalis (substantia innominata), and hypothalamus in
the rhesus monkey. Journal of Comparative Neurology, 214, 170-197.
Mesulam, M.-M., & Van Hoesen, G. W. (1976). Acetylcholinesterase-rich projections from the
basal forebrain of the rhesus monkey to neocortex. Brain Research, 109, 152-157.
Mesulam, M.-M., Van Hoesen, G. W., Pandya, D. N., & Gcschwind, N. (1977). Limbic and
sensory connections of the inferior parietal lobule (area PG) in the rhesus monkey:
A study using a new method for horseradish peroxidase. Brain Research, 136, 393-
414.
Mishkin, M. (1978). Memory in monkeys severely impaired by combined but not by separate
removal of the amygdala and hippocampus. Nature, 273, 297-298.
Mishkin, M. (1982). A memory system in the monkey. Philosophical Transactions of the Royal
Society (London), 298, 85-95.
Morel, F., & Wildi, E. (1955). Contribution a la connaissance des differentes alterations cere-
brales du grand age. Schweizer Archiv Neurologie und Psychiatric, 76, 195-223.
Nakano, L, & Hirano, A. (1982). Loss of large neurons of the medial septal nucleus in an
autopsy case of Alzheimer's disease. Journal of Neuropathology and Experimental
Neurology, 41, 341.
 Alzheimer's Disease, Neural Systems, and Memory 259

Nakano, I., & Hirano, A. (1983). Neuron loss in the nucleus basalis of Meynert in parkinson-
ism-dementia complex of Guam. Annals of Neurology, 13, 87-91.
Pandya, D. N., & Kuypers, H. G. J. M. (1969). Cortico-cortical connections in the rhesus mon-
key. Brain Research, 13, 13-36.
Pandya, D. N., & Seltzer, B. (1982a). Association areas of the cerebral cortex. Trends in Neu-
rosciences, 5, 386-390.
Pandya, D. N., & Seltzer, B. (1982b). Intrinsic connections and architectonics of posterior pari-
etal cortex in the rhesus monkey. Journal of Comparative Neurology, 204, 196-
210.
Pandya, D. N., Van Hoesen, G. W., & Mesulam, M.-M. (1981). The cortico-cortical projections
of the cingulate cortex in the rhesus monkey. Experimental Brain Research, 42, 319-
330.
Pandya, D. N., & Yeterian, E. H. (1985). Architecture and connections of cortical association
areas. In A. Peters & E. G. Jones (Eds.), Cerebral cortex: Vol. 4. Association and audi-
tory cortices (pp. 3-61). New York: Plenum.
Pearson, R. C. A., Esiri, M. M., Hiorns, R. W., Wilcock, G. K., & Powell, T. P. S. (1985).
Anatomical correlates of the distribution of the pathological changes in the neocortex
in Alzheimer disease. Proceedings of the National Academy of Sciences USA, 82,
4531-4534.
Pearson, R. C. A., Gatter, K. C., Brodal, P., & Powell, T. P. S. (1983). The projection of the
basal nucleus of Meynert upon the neocortex in the monkey. Brain Research, 259,
132-136.
Pearson, R. C. A., Gatter, K. C, & Powell, T. P. S. (1983). Retrograde cell degeneration in the
basal nucleus in monkey and man. Brain Research, 261, 321-326.
Pearson, R. C. A., Sofroniew, M. V., Cuello, A. C, Powell, T. S., Eckenstein, F., Esiri, M. M.,
& Wilcox, G. K. (1984). Persistence of cholinergic neurons in the basal nucleus in a
brain with senile dementia of the Alzheimer's type demonstrated by immunohisto-
chemical staining for choline acetyltransferase. Brain Research, 239, 375-379.
Perry, E. K., & Perry, R. H. (1986). A review of neuropathological and neurochemical correlates
of the Alzheimer's disease. Danish Medical Bulletin, 32 (Suppl. 1), 27-34.
Porrino, L. J., Crane, A. M., & Goldman-Rakic, P. S. (1981). Direct and indirect pathways from
the amygdala to the frontal lobe in rhesus monkeys. Journal of Comparative Neurol-
ogy, 198, 121-136.
Rockland, K. S., & Pandya, D. N. (1979). Laminar origins and terminations of cortical con-
nections of the occipital lobe in the rhesus monkey. Brain Research, 179, 3-20.
Rockland, K. S., & Pandya, D. N. (1981). Cortical connections of the occipital lobe in the rhesus
monkey: Interconnections between areas 17, 18, 19 and the superior temporal gyrus.
Brain Research, 212, 249-270.
Rogers, J., & Morrison, J. H. (1985). Quantitative morphology and regional and laminar dis-
tributions of senile plaques in Alzheimer's disease. Journal of Neuroscience, 5, 2801-
2808.
Rosene, D. L., & Van Hoesen, G. W. (1977). Hippocampal efferents reach widespread areas of
the cerebral cortex and amygdala in the rhesus monkey. Science, 198, 315-317.
Rossor, M. N. (198!). Parkinson's disease and Alzheimer's disease as disorders of the isoden-
driticcore. British Medical Journal, 283, 1588-1590.
Rudelli, R. D., Ambler, M. W., & Wisniewski, H. M. (1984). Morphology and distribution of
Alzheimer neuritic (senile) and amyloid plaques in striatum and diencephalon. Ada
Neuropathologica, 64, 273-286.
Saper, C. B., German, D. C, & White, C. L., Ill, (1985). Neuronal pathology in the nucleus
basalis and associated cell groups in senile dementia of the Alzheimer's type: Possible
role in cell loss. Neurology, 35, 1089-1095.
Schwartz, M. L., & Goldman-Rakic, P. S. (1982). Single cortical neurones have axon collaterals
to ipsilateral and contralateral cortex in fetal and adult primates. Nature, 299, 154-
155.
260 REGULATION OF CORTICAL FUNCTION IN MEMORY

Scoville, W. B.,& Milner, B. (1957). Loss of recent memory after bilateral hippocampal lesions.
Journal of Neurology Neurosurgery and Psychiatry, 20, 11-21.
Seltzer, B., & Pandya, D. N. (1976). Some cortical projections to the parahippocampal area in
the rhesus monkey. Experimental Neurology, 50, 146-160.
Shipley, M. T., & Sorensen, K. E. (1975). Some afferent and intrinsic connections in the guinea
pig hippocampal region and a new pathway from subiculum feeding back to parahip-
pocampal cortex. Experimental Brain Research (Suppl. 1), 188-190.
Smith, C. M., & Swash, M. (1978). Possible biochemical basis of memory disorder in Alzheimer
disease. Annals of Neurology, 3, 471-473.
Sofroniew, M. V., Pearson, R. C. A., Eckenstein, R., Cuello, A. C, & Powell, T. P. S. (1984).
Retrograde changes in cholinergic neurons in the basal forebrain of the rat following
cortical damage. Brain Research, 289, 370-374.
Sorenson, K. E., & Shipley, M. T. (1979). Projections from the subiculum to the deep layers of
the ipsilateral presubicular and entorhinal cortices in the guinea pig. Journal of Com-
parative Neurology, 188, 313-314.
Squire, L. R. (1986). Mechanisms of memory. Science, 232, 1612-1619.
Squire, L. R. (1987). Memory: Neural organization and behavior. In F. Plum (Ed.), Handbook
of physiology: Sec. I: The nervous system (pp. 295-372). Bethesda, Md.: American
Physiological Society.
Struble, R. G., Cork, L. C., Whitehouse, P. J., & Price, D. L. (1982). Cholinergic innervation
in neuritic plaques. Science Washington, DC, 216, 413-415.
Swanson, L. W., & Cowan, W. M. (1975). Hippocampo-hypothalamic connections: Origin in
subicular cortex, not Ammon's horn. Science, 189, 303-304.
Swanson, L. W., & Cowan, W. M. (1977). An autoradiographic study of the organization of the
efferent connections of the hippocampal formation in the rat. Journal of Comparative
Neurology, 172, 49-84.
Terry, R. D. (1983). Cortical morphometry in Alzheimer's disease. In R. Katzman (Ed.), Ban-
bury report. Biological aspects of Alzheimer's disease (Vol. 15, pp. 95-103). Cold
Spring Harbor, N.Y.: Cold Spring Harbor Laboratory.
Terry, R. D., & Katzman, R. (1983). Senile dementia of the Alzheimer type. Annals of Neurol-
ogy, 14, 497-506.
Terry, R. D., Peck, A., DeTeresa, R., Schechter, R., & Horoupian, D. S. (1981). Some morpho-
metric aspects of the brain in senile dementia of the Alzheimer type. Annals of Neu-
rology, 10, 184-192.
Tomlinson, B. E., Blessed, G., & Roth, M. (1968). Observations on the brains of demented old
people. Journal of the Neurological Sciences, 7, 331-356.
Tomlinson, B. E., & Corsellis, J. A. N. (1984). Ageing and the dementias. In J. Hume Adams,
J. A. N. Corsellis, & L. W. Duchen (Eds.), Greenfield's neuropathology (pp. 951-
1025). New York: Wiley.
Tomlinson, B. E., Irving, D., & Blessed, G. (1981). Cell loss in the locus coeruleus in senile
dementia of Alzheimer type. Journal of the Neurological Sciences, 49, 419-428.
Tulving, E. (1987). Multiple memory systems and consciousness. Human Neurobiology, 6, 67-
80.
Turner, B. H., Mishkin, M., & Knapp, M. (1980). Organization of the amygdalopetal projec-
tions from modality-specific cortical association areas in the monkey. Journal of
Comparative Neurology, 191, 515-543.
Van Essen, D. C., & Maunsell, J. H. R. (1983). Hierarchical organization and functional
streams in the visual cortex. Trends in Neurosciences, 6, 370-375.
Van Hoesen, G. W. (1981). The differential distribution, diversity and sprouting of cortical
projections to the amygdala in the rhesus monkey. In Y. Ben Ari (Ed.), The amyg-
daloid complex (pp. 77-90). Amsterdam: Elsevier.
Van Hoesen, G. W. (1982). The primate parahippocampal gyrus: New insights regarding its
cortical connections. Trends in Neurosciences, 5, 345-350.
 Alzheimer's Disease, Neural Systems, and Memory 261

Van Hoesen, G. W. (1985). Neural systems of the non-human primate forebrain implicated in
memory. In S. Corkin, E. Gramzu, & D. Olton (Eds.), Memory dysfunctions (pp. 97-
112) (Annals of the New York Academy of Sciences, Vol. 444). New York: New York
Academy of Sciences.
Van Hoesen, G. W., & Damasio, A. R. (1987). Neural correlates of cognitive impairment in
Alzheimer's disease. In F. Plum (Ed.), Handbook of physiology: Sec. 1: The nervous
system (pp. 871-898). Bethesda, Md.: American Physiological Society.
Van Hoesen, G. W., Hyman, B. T., & Damasio, A. R. (1986). Cell-specific pathology in neural
systems of the temporal lobe in Alzheimer's disease. In D. Swaab (Ed.), Progress in
brain research (Vol. 70, pp. 361-375). Amsterdam: Elsevier.
Van Hoesen, G. W., & Pandya, D. N. (1975a). Some connections of the entorhinal (area 28)
and perirhinal (area 35) cortices in the rhesus monkey. I. Temporal lobe afferents.
Brain Research, 95, 1 -24.
Van Hoesen, G. W., & Pandya, D. N. (1975b). Some connections of the entorhinal (area 28)
and perirhinal (area 35) cortices of the rhesus monkey. III. Efferent connections.
Brain Research, 95, 39-59.
Van Hoesen, G. W., Pandya, D. N., & Butters, N. (1972). Cortical afferents to the entorhinal
cortex of the rhesus monkey. Science, 175, 1471-1473.
Van Hoesen, G. W., Pandya, D. N., & Butters, N. (1975). Some connections of the entorhinal
(area 28) and perirhinal (area 35) cortices in the rhesus monkey. II. Frontal lobe affer-
ents. Brain Research, 95, 25-38.
Vogt, B. A., & Miller, M. W. (1983). Cortical connections between rat cingulate cortex and
visual, motor and postsubicular cortices. Journal of Comparative Neurology, 216,
192-210.
Volpe, B. T., & Hirst, W. (1983). Amnesia following the rupture and repair of an anterior com-
municating artery aneurysm. Journal of Neurology Neurosurgery and Psychiatry, 46,
704-709.
von Cramon, D. Y., Hebel, N., & Schuri, U. (1985). A contribution to the anatomic basis of
thalamic amnesia. Brain, 108, 993-1008.
Warner, M. D., Peabody, O. A., Flattery, J. J., & Tinklenberg, J. R. (1986). Olfactory deficits
and Alzheimer's disease. Biological Psychiatry, 21, 116-118.
Whitehouse, P. J., Price, D. L., Clark, A. W., Coyle, J. T., & DeLong, M. R. (1981). Alzheimer
disease: Evidence for selective loss of cholinergic neurons in the nucleus basalis.
Annals of Neurology, 10, 122-126.
Whitlock, D. G., & Nauta, W. J. H. (1956). Subcortical projections from the temporal neocortex
in Macaco, mulatta. Journal of Comparative Neurology, 106, 183-212.
Wilcock, G. K. (1983). The temporal lobe in dementia of Alzheimer's type. Gerontology, 29,
320-324.
Wilcock, G. K., & Esiri, M. M. (1982). Plaques, tangles and dementia, a-qualitativc study. Jour-
nal of the Neurological Sciences, 56, 343-356.
Wilcock, G. K., Esiri, M. M., Bowen, D. M., & Smith, C. C. T. (1983). The nucleus basalis in
Alzheimer's disease—Cell counts and cortical biochemistry. Neuropathology and
Applied Neurobiology, 9, 175-179.
Yamamoto, T., & Hirano, A. (1985). Nucleus raphe dorsalis in Alzheimer's disease. Neurofi-
brillary tangles and loss of large neurons. Annals of Neurology, 17, 573-577.
This page intentionally left blank
COMMENTARIES AND
ALTERNATIVE PERSPECTIVES
This page intentionally left blank
12
The Neocortex and Memory Storage
H E R B E R T P. K I L L A C K E Y

In this brief chapter, I will outline some of the major themes appearing in neo-
cortical research since the early 1960s, how these themes are currently being
modified, and how they relate to the involvement of the neocortex in learning
and memory. The focus will be on memory processes. I use this term in the lay
sense of facts or data about the everyday world that one can (consciously or
unconsciously) store and recall on demand. Good examples are telephone or
social security numbers or, most wondrously, the name to match the face of a
colleague seen but once a year at a scientific meeting.
I will begin with a rather naive statement. In my opinion, memory processing
of the sort just referred to is both central to the human experience and carried
out at the cortical level (i.e., the neocortex is a major memory storage site). Such
memory processes remove each of us from the precarious immediacy of our
environment and connect us to a wealth of direct and indirect experiences with
which we can interpret environmental events. Indeed, some of the major inven-
tions of human civilization can be regarded chiefly as devices that both greatly
extend and externalize our neocortical memory-processing abilities. The most
obvious of these are the development of written language, which made possible
a permanent and transferable record of events; the printing press and movable
type, which together assured an increase in accuracy of written records and
resulted in the much wider dissemination of written records than was previously
possible; and the computer, which has been with us for a relatively short time
but will most likely further extend these processes in ways that cannot be entirely
foreseen.
What is the evidence that memory processing of this sort is particularly depen-
dent on the neocortex? First, there is a large body of literature from the neuro-
logical clinic. One particularly relevant disorder is Alzheimer's disease, in which
memory loss is a major symptom and there are specific and devastating effects
on neocortical morphology. Second, there have been experimental studies of the
effects of neocortical lesions on behavior in a variety of nonhuman mammalian
species. Together, these bodies of literature provide strong circumstantial, if not
definitive, evidence for neocortical involvement in memory processes. As a

265
266 REGULATION OF CORTICAL FUNCTION IN MEMORY

comparative neuroanatomist, I would also advance the following somewhat

convoluted argument. The cerebral hemispheres are the portion of the verte-
brate brain which have undergone the greatest expansion in the course of evo-
lution. A major part of this expansion can be attributed to the portion of the
cerebral hemispheres known as the neocortex. In higher primates, particularly
humans, the six-layered neocortex has greatly increased in surface area relative
to the other portions of the cerebrum and is its dominant feature. I would
assume that this greatly expanded neocortex is correlated with the expanded
behavioral repertoire of the human species, including our extensive memory
abilities. At a more prosaic level, I would submit that it is only this expanded
neocortex that can possibly provide the expanded physical substrate necessary
to meet the storage requirements of human memory.
There is, of course, nothing novel in the suggestion that memories are stored
in the neocortex. Such an idea has played some role in theories of cortical func-
tion going back to the pioneering studies of cortical localization in the late nine-
teenth century. With the early broad functional characterization of neocortical
areas as either sensory, motor, or associational also comes the suggestion that
memories are stored in association areas. Such a view was also implicit in the
attempts of physiologists early in this century to view neocortical function in
terms of conditioned reflex circuits. At the cortical level, such circuits stretched
from sensory cortex through association cortex to motor cortex. It was also gen-
erally assumed that it was only in association cortex that such a circuit could be
modified. For example, Herrick could state in 1926 that
quite apart from such speculative considerations it is clear that the mnemonic func-
tions of the association cortex are fundamental. The impressionability of this cortex
and its retentiveness of patterns once impressed upon it underlie all of its mysterious
capacities which are clearly recognizable in the anthropoid apes and which are prob-
ably incipient in dogs. (1926/1963, p. 244)

It was in this intellectual context that Lashley set out on his search for the
engram, the locus of the memory trace. On the basis of a lifetime of experiments,
he concluded not only that association areas of cortex are not specific store-
houses for specific memories but also that it was not possible to localize memory
traces anywhere within the nervous system (Lashley, 1950). From today's per-
spective it is clear that the experimental approach used by Lashley was more
limited than he recognized, and considerable progress has been made in relating
specific parts of the nervous system to memory processes in a number of species.
However, the conclusion that no specific cortical area serves as a memory stor-
age site is still valid. In this light, memory storage is probably best regarded as
an integral part of more general processing operations, and as such inseparable
from these operations and distributed throughout the neocortex. Since Lashley,
experimental investigations of the neocortex have been largely approached from
a different intellectual vantage.
Since the early 1960s, the dominant approach to the study of neocortical orga-
 The Neocortex and Memory Storage 267

nization has been made through the major sensory systems (the visual system,
in particular, but also the somatosensory and auditory systems). At a functional
level, the analysis of receptive field properties of neurons at a number of levels
along the neural axis, as well as within multiple cortical areas, has greatly
expanded our knowledge of the way sensory information is processed within the
brain. In the visual system, processing of information is clearly multifaceted. On
the one hand, there is evidence for the serial and hierarchical processing of
visual information; on the other hand, there is compelling evidence for parallel
processing of visual information in the sense that different functional categories
of visual information may be processed in separate but overlapping channels at
both the subcortical and cortical levels. Similarly, anatomical studies carried out
with recently introduced methods of great sensitivity have significantly
increased our knowledge of cortical connectivity. For example, it was once
assumed that major regions of neocortex, particularly association areas, did not
receive afferent input from the dorsal thalamus. Association areas were thought
to be dependent on sensory cortical areas for their afferent inputs. It is now clear
that this is not the case. Indeed, it has become clear that the classical notions of
a primary cortical sensory area for the analysis of sensation and of cortical asso-
ciation areas subserving more complex perceptual functions are at best almost
too simplistic to be useful. The use of the term association area in the classical
sense has almost disappeared from the current literature and has been replaced
by either neutral terms or terms that relate a given cortical area to a major sen-
sory system on the basis of the neural response properties in this area. For exam-
ple, in the rhesus monkey over a dozen cortical areas which are largely devoted
to visual function have been identified, and undoubtedly future investigations
will uncover still more cortical areas that can be characterized as visual (Van
Essen & Maunsell, 1983).
There has also been some progress in our understanding of the general func-
tional properties of the neocortex. While much of the research referred to in the
preceding paragraphs focused on defining discrete and different cortical areas in
both a functional and morphological sense, there is increasing evidence that neo-
cortex may follow the same general operational and computational principles
across cortical areas. At the morphological level, the similarities in organization
across cortical areas far outweigh the differences. There are overall similari-
ties across cortical areas in the distribution of cell types, patterns of afferent
input, and distribution of efferent output neurons. Further, it has been hypoth-
esized that there is a basic functional unit of neocortical organization composed
of a vertically oriented column of neurons extending across all cortical layers
(Mountcastle, 1978). Thus it seems reasonable to suggest that similar processing
operations may be performed in diverse cortical areas. If this is indeed the case,
the task of understanding neocortical organization is potentially much simpler.
This hypothesis receives some general support from two very different
approaches to neocortical organization. First, there have been experimental
developmental studies in which visual input was diverted into the subcortical
268 REGULATION OF CORTICAL FUNCTION IN MEMORY

somatosensory system and the response properties of neurons in the adult soma-
tosensory cortex was assayed (Frost & Metin, 1985). Under these circumstances,
some neurons in the somatosensory cortex respond to visual stimulation and
their response properties are similar to normal visual cortex neurons. This sug-
gests that cortical processing operations are somewhat independent of sensory
modality and specific cortical area. Second, it has been reported that the absolute
number of neurons in a small volume of neocortex extending from the surface
to the white matter is invariant across both cortical areas and species as diverse
as rat, cat, and humans (Rockel, Hiorns, & Powell, 1980).
These findings have significant implications for our understanding of neocor-
tical phylogeny. First, the fundamental unit of cortical information processing
may be quite conservative and have remained relatively unchanged in the
course of mammalian evolution. Second, the phylogenetic expansion of mam-
malian neocortex may have been accomplished by the straightforward addition
of basic vertical processing units. Thus the varying number of cortical areas seen
in different mammalian species may be composed of the same basic functional
units which in different species are grouped into different cortical areas on the
bases of the adaptive pressures acting in a given species.
A major component of any such cortical processing unit must be the pyram-
idal cell, the most common neocortical neuron and the only one with an axon
leaving the neocortex. This neuron with its vertically organized apical dendrite
stretching across the cortical layers between its cell body and the cortical surface
is positioned to sample afferent inputs from diverse sources and richly encrusted
with dendritic spines. A single pyramidal cell may have as many as 30,000 of
these postsynaptic specializations. This is particularly important to bear in mind
as most conceptions of the neural basis of learning and memory are couched in
terms that seem to regard the neuron as being involved in a single, or at best a
few synapses. The reality, far more complicated, was perhaps hinted at by Lash-
ley (1950) in another of the conclusions he reached in "In Search of the
Engram":

From the numerical relations involved, I believe that even the reservation of indi-
vidual synapses for special associative reactions is impossible. The alternative is, per-
haps, that the dendrites and cell body may be locally modified in such a manner that
the cell responds differentially, at least in the timing of its firing, according to the
pattern of combination of axon feet through which excitation is received, (p. 482)

As outlined above, the major heuristic approach to the understanding of neo-

cortical organization has been based on the sensory systems. Further, the
approach can be characterized as one that begins at the periphery, at the sensory
receptor, and then proceeds inward station by station in an attempt to unravel
both processing operations and sequence. This approach has been very valuable
and has greatly increased our knowledge of neocortical organization. At the
same time, this focus on sensory-based information, and consequently the dorsal
thalamus, has overshadowed other major afferent sources to the neocortex.
 The Neocortex and Memory Storage 269

Indeed, it was often assumed until fairly recently that the dorsal thalamus was
the only source of afferent input to the neocortex. It is now clear that there are
a number of major nonthalamic afferent systems to the neocortex (Foote & Mor-
rison, 1987). Each of these extrathalamic afferent systems arises in a different
site and is associated with a different putative neurotransmitter. The best char-
acterized to date are noradrenergic afferents from the locus coeruleus, seroto-
nergic afferents from the raphe nuclei, dopaminergic projections from the ven-
tral tegmental area, and acetylcholinergic afferents from the nucleus basalis.
Each of these afferent systems projects to the neocortex in a distinct areal and
laminar fashion, although as a whole their projection patterns are more wide-
spread and less highly organized than thalamic projections. The functional role
played by these systems in neocortical processing is still obscure, but it is likely
to differ significantly from the highly topographically organized thalamic system.
The functions of these systems are broadly viewed as modulatory but in a local,
not a global, sense. That is, these systems are organized in such a fashion that
they may regulate activity in small groups of neurons during particular stages in
information processing. Whether these afferent systems play any particular role
in memory processing remains to be determined. The developmental gating
mechanisms discussed by Singer (Chapter 10, this volume) point to one
area where thalamic and extrathalamic afferent systems to the cortex may
overlap.
Our sensory system-based orientation toward neocortical organization has
also influenced how we view patterns of connections within the neocortex. Anal-
yses of corticocortical connections have mainly focused on projections that arise
in sensory cortical areas. Such studies are aimed largely at outlining the major
flow of information within the neocortex, which is presumed to begin in sensory
cortex and then continue in other cortical areas. Such assumptions are made
explicit in the terms "feed forward," which characterizes projections originating
from a primary sensory cortical area or a cortical area that is assumed to be
lower in a hierarchical processing sequence, and "feed backward," which char-
acterizes projections in the opposite direction. Again, there has been much heu-
ristic value in such an approach but the time may have come to take a wider
approach, recognizing other major influences besides sensory systems on neo-
cortical processing. Van Hoesen (Chapter 11, this volume) and Rolls (Chapter
9, this volume) stress the largely ignored projections from other telencephalic
regions onto major portions of the neocortex, including primary sensory areas.
Although many of the details of such projections have not yet been studied
nearly as well as other neocortical inputs, they include direct and indirect pro-
jections from both the hippocampus and amygdala to most regions of the neo-
cortex. The model presented by Rolls is particularly noteworthy in that it pos-
tulates several specific functions for these projections, including the recall of
memories stored in the neocortex. Such hypotheses will undoubtedly be tested
by future research and further enrich our still evolving ideas of neocortical
organization.
270 REGULATION OF CORTICAL FUNCTION IN MEMORY

REFERENCES

Foote, S. L., & Morrison, J. H. (1987). Extrathalamic modulation of cortical function. Annual
Review of Newoscience, 10, 67-95.
Frost, D. O., & Metin, C. (1985). Induction of functional retinal projections to the somatosen-
sory system. Nature, 317, 162-164.
Herrick, C. J. (1963). Brains of rats and men. New York: Hafner. (Original work published in
1926)
Lashley, K. S. (1950). In search of the engram. Symposium of the Society for Experimental
Biology, 4, 454-482.
Mountcastle, V. B. (1978). An organizing principle for cerebral function: The unit module and
the distributed system. In G. M. Edelman (Ed.), The mindful brain (pp. 7-50). Cam-
bridge, Mass.: MIT Press.
Rockel, A. J., Hiorns, R. W., & Powell, T. P. S. (1980). The basic uniformity in structure of the
neocortex. Brain, 103, 221-244.
Van Essen, D. C., & Maunsell, J. H. R. (1983). Hierarchical organization and functional
streams in the visual cortex. Trends in Neuroscience, 6, 370-375.
13
A Network Model for Learned Spatial
Representation in the Posterior
Parietal Cortex
R I C H A R D A. A N D E R S E N
DAVID ZIPSER

Anatomists and neurophysiologists who study the cerebral cortex generally

believe that they are studying a hard-wired, genetically determined circuitry. It
is usually assumed that it is only during brief "critical periods" that the cortex
is amenable to changes dictated by exposure to the environment which fine tune
its structure. However, recent evidence suggests that the cortex is very plastic
and that this plasticity extends into adult life (see Merzenich, Recanzone, Jen-
kins, Allard, & Nudo, in press, for review). In the experiments described in this
chapter, combined neurophysiological and computational approaches were used
to investigate how the posterior parietal cortex of macaque monkeys represents
the location of visual stimuli in craniotopic coordinates (Andersen & Zipser,
1988; Zipser & Andersen 1988). A computer-generated network model was
designed to learn stimulus locations in craniotopic space based on eye and ret-
inal position inputs that were modeled on similar signals derived from the
recording data and are assumed to be inputs to the posterior parietal cortex. The
units in the network that map from input to output were found to develop the
same response properties as a large subset of cells found in area 7a of the pos-
terior parietal cortex. These experiments suggest that the spatial representation
found in area 7a is in fact learned by associating eye position with retinal inputs.

SPATIAL REPRESENTATIONS AND THE ROLE

OF THE POSTERIOR PARIETAL CORTEX

There are several reasons to believe that the brain uses representations of visual
space that are nonretinotopic and are framed in head- or body-centered coor-
dinates. One can reach accurately to the location of visual targets without visual

271
272 REGULATION OF CORTICAL FUNCTION IN MEMORY

feedback and independent of eye position, head position, or the location of the
image on the retinas. Thus the motor system appears to use representations of
visual stimuli mapped in body-centered rather than retinal coordinates. There
is substantial evidence that the planning of eye movements involves a stage in
which the target of the eye movement is represented in craniotopic coordinates
(Hallet & Lightstone, 1976; Mays & Sparks, 1980; Robinson, 1975). We know
by introspection that the visual world appears perceptually stable in spite of the
fact that we are constantly making eye movements and subsequently shifting the
location of images on the retinas. All these results suggest that there exist neural
representations of space that are head or body centered.
A likely area of the brain to find nonretinotopic representations of visual space
is the posterior parietal cortex. Lesions in this area in humans and monkeys
produce visual disorientation, a syndrome in which the subjects cannot reach
accurately to visual targets and have difficulty navigating around seen obstacles
(see Andersen, 1987, for review). The patients are not blind and when tested
often have normal visual field functions. However, they appear to be unable to
associate what they see with the positions of their bodies.

RECORDING DATA FROM AREA 7

We examined the coordinate frame for visual space used by the posterior pari-
etal cortex by mapping visual receptive fields in this area with animals looking
in different directions (Andersen, Essick, & Siegel, 1985). The animals' heads
were fixed to simplify the coordinate space examined to a head-centered coor-
dinate frame. We reasoned that if the receptive fields moved with the eyes, then
the coordinate frame was retinotopic; if they remained static in space, then they
were coding in at least craniotopic coordinates. Figure 13.1 shows an example
of this experiment. The receptive field is first mapped with a flashed visual stim-
ulus while the animal fixates a small fixation spot located straight ahead at 0,0
in screen coordinates. Figure 13.IB shows a typical receptive field mapped in
this way where the axes represent screen coordinates and the contour lines dif-
ferent levels of neural response. Once the receptive field has been mapped, the
stimulus is then presented at the retinotopic location that gave the maximum

FIGURE 13.1. (A) Method of determining spatial gain fields of area 7a neurons. The animal
fixates point f at different locations on the screen with his head fixed. The stimulus, s, is always
presented in the center of the receptive field, rf. (B) Receptive field of a neuron plotted in coor-
dinates of visual angle determined with the animal always fixating straight ahead (screen coordinates
0,0). The contours represent the mean increased response rates in spikes per second. (C) Spatial
gain field of the cell in (B). The poststimulus histograms are positioned to correspond to the
locations of the fixations on the screen at which the responses were recorded for retinotopically
identical stimuli presented in the center of the receptive field (histogram ordinate, 25 spikes per
division, and abscissa, 100 msec per division; arrows indicate onset of stimulus flash). (From
Andersen et al., 1985)
A) C)

B)
274 REGULATION OF CORTICAL FUNCTION IN MEMORY

response, but with the animal gazing in different directions. If the response
changes with eye position for retinotopically identical stimuli, then the cells are
not coding in strictly retinotopic coordinates. Figure 13.1A shows how the visual
response is tested at different eye positions. The dashed line rf delineates the
receptive field that was mapped in Figure 13.IB. Stimulus s is presented at the
most responsive location in the receptive field, in this case the approximate cen-
ter of the response zone. On the left the animal is first required to fixate straight
ahead at 0,0 in screen coordinates. On the right the animal has been required to
fixate to the left of straight ahead by 20° at — 20,0 in screen coordinates. Since
the head is fixed the eyes are now in different positions in the orbits. The target
is again flashed in the same retinotopic location; however, since the eyes have
moved 20° to the left the stimulus has also been moved 20° to the left such that
the stimulus falls on the same retinotopic location.
In these experiments nine eye positions are tested in this manner. Results
from this cell are seen in Figure 13.1C. Each histogram is plotted at the corre-
sponding fixation location. The cell was most active for fixations up 20° and left
20°, less active for looking straight ahead, and not active at all when looking
down 20° and right 20°. The activity of the cell for retinotopically identical stim-
uli varied as a function of the angle of gaze. Plots for the nine fixation positions
as shown in Figure 13.1C are called spatial gain fields. Notice that the shape of
this particular gain field can be described by a plane tilted down and to the right.
We recorded complete spatial gain fields for 86 area la cells. The mean evoked
responses of these gain fields were further analyzed using a first-order linear
model with independent variables of horizontal and vertical eye position to
determine how many of these gain fields could be fit with a plane. The evoked
activity was obtained by subtracting the background activity before the stimulus
flash from the overall activity during and just after the stimulus flash. Three
types of gain fields were obtained by this analysis. Thirty-one percent of the gain
fields had a significant planar component and no significant lack of fit, indicating
that a plane was the best model for the data. Another 32% showed a planar com-
ponent but also a significant lack of fit, indicating that although a plane could
be fit to the data a plane was not the optimal model. Fully 75% of these neurons
looked very planar. Finally, another 37% showed no planar component and a
significant lack of fit. Thus a majority of cells showed planar or largely planar
gain fields (55%), but a significant number of gain fields (45%) are not planar.
When the same cells were analyzed for their overall activity rather than just the
evoked activity, it was found that a larger proportion of the cells (78%) had pla-
nar or largely planar gain fields. Thus the total signal of background activity and
evoked activity shows a greater degree of planarity than just the evoked
response. The total activity is the most likely signal used by the brain for spatial
localization since it is the output of the neurons.
A major concern is that the visual background, which is imaged at different
locations on the retinas at different angles of gaze, is influencing the responsive-
ness of the cells to the test flash. Two controls were performed to eliminate this
possibility. Many of the recordings were made in complete darkness except for
 Network Model for Learned Spatial Representation 275

the stimulus and the fixation point so that there was no visual background. The
second control was to change the angle of gaze using prisms; this requires the
animal to change eye position without changing the retinal locations of the
imaged background or visual stimulus. The cells showed the same gain fields
whether the eye positions were changed with prisms or without prisms by mov-
ing the fixation point. Thus we can conclude that the effect of eye position on
visual responses was not a result of background shifts.

EYE POSITION-DEPENDENT SPATIAL TUNING

The change in visual responses for retinotopically identical stimuli with eye
position could be a result of two mechanisms. First, the cells could have been
coding locations of targets in space independent of eye position. In this situation
the cells' receptive fields remain static in space and the retinal addresses of the
receptive fields change with eye movements to remain constant for spatial loca-
tion. Second, the receptive fields could remain retinotopic, with only the respon-
siveness of the cells varying as a function of eye position. In other words, eye
position gates the activity of the retinal receptive fields. To distinguish between
these two possibilities we mapped entire axes through the center of the receptive
fields. It was found that the responsiveness of neurons varies as a function of
eye position, but the peaks and symmetry of their receptive fields do not change.
Thus the receptive fields remain retinotopic and it is only the responsiveness of
the cells that is modulated by eye position.
The activity of parietal neurons was modeled as a multiplicative interaction
of eye position and retinal position using the equation A — G(ex,ey) X R(rx,ry)
where A is the cells' firing rate, G is a gain factor that is a function of horizontal
(ex) and vertical (ey) eye position, and R is the visual stimulus response profile,
which is a function of horizontal (rx) and vertical (ry) retinal locations. This mul-
tiplicative interaction produced a tuning for the location of visual stimuli in
craniotopic space, but it was dependent on eye position. A simple example of
this would be a cell that has a gain of 0 for all eye positions except looking 10°
to the right, in which case the gain is 1. This cell also has a narrow receptive
field centered at 10° to the right of the fovea. This cell will then be tuned to a
location 20° to the right in craniotopic coordinates, but only when the animal is
looking 10° to the right.

PROBLEMS TO BE ADDRESSED BY MODELS OF SPATIAL

REPRESENTATION IN AREA 7a

Cells have never been found that are spatially tuned in an eye position-inde-
pendent manner. From these experiments it must be concluded that the neural
representation of space in an eye position-independent manner is distributed.
This distributed coding presents a problem: How do you determine eye posi-
276 REGULATION OF CORTICAL FUNCTION IN MEMORY

tion-independent spatial location from the population response? One way

would be to map spatial location tuning systematically across the tangential
dimension of area 7a. Recording experiments have so far not revealed any obvi-
ous topography for spatial tuning, indicating that if such a map for craniotopic
space does exist, it will likely be crude. Moreover, the very large receptive fields
and spatial tuning fields would tend to mitigate against high-resolution mapping
like that found for retinotopy in VI.
Another unusual feature of the receptive fields of area 7a neurons is their com-
plexity. The fields are large, have approximately equal weighting to the fovea
and periphery, and can have multiple peaks. The fields generally have smoothly
varying levels for response for nearby sample points. No topographic organiza-
tion for retinal location of the receptive fields has been found. Any model for
spatial representation in area 7a should reproduce these unusual visual receptive
fields.
Another interesting aspect of the area 7a neurons is that the background activ-
ity of the cells also often varies as a function of eye position. In many cases the
background activity varies in the same direction as the gain on the visual
response. This result is not unexpected since the eye position input could
increase the visual response by depolarizing the membrane and in some cases
this depolarization would not only lower the cell's threshold to visual stimula-
tion but may also fire the cell, leading to an increase in its background activity.
However, other cells showed background effects that went in the opposite direc-
tion to the gain of the visual response. Again, any model of spatial coding in
area 7a would need to explain this behavior.

NETWORK MODEL FOR SPATIAL REPRESENTATION

Zipser and Andersen created a parallel network model that learns to map inputs
of retinotopic position and eye position to an output of location in head-cen-
tered space (Andersen & Zipser, 1988; Zipser & Andersen, 1988). This network
consists of three layers and uses the backpropagation learning algorithm
(Rumelhart, Hinton, & Williams, 1986). The units in the middle layer that
accomplish the spatial transformation show the same eye position-dependent
spatial tuning properties that are found for area 7a neurons. This model also
generates retinal receptive fields similar to those found for area 7a neurons and
reproduces similar background activities. The remarkable correspondence
between the model and experiment suggests that the distributed spatial coding
discovered in area 7a neurons is indicative of spatial transformations carried out
using the same computational algorithm discovered by the backpropagation
learning technique.
Figure 13.2 presents a schematic diagram of the network. The input layer con-
sists of a 10 by 10 retinal array and four eye position units. The retinal receptive
fields are Gaussian in shape with 1/e widths of 15°. This input is designed to be
similar to the receptive fields that are found in area 7a that do not show eye
 Network Model for Learned Spatial Representation 111

position effects and are assumed to be the retinal inputs to area 7a. The centers
of the 100 receptive fields are equally spaced over the 10 by 10 grid with 8°
spacings. The four eye position units consist of two units coding vertical position
and two horizontal position using opposite, symmetrical slopes. Each unit used
either a linear or, in later simulations, a squared function to approximate the
signal coming from eye position cells. The rationale for using a squared function
is to approximate the cumulative response of a group of eye position cells. Eye
position inputs are assumed to be those cells in area 7a that have only eye posi-
tion signals and no visual response. These cells generally code horizontal and/
or vertical position in a linear fashion (Andersen & Zipser, 1988). We ran sim-
ulations with both square functions and simple linear functions and got indis-
tinguishable results, indicating that the exact representation of the eye position
does not appear to be crucial as long as it is a monotonically increasing function.
The intermediate layer receives inputs from all 104 input units and in turn
projects to two or four output units. The output units code position in head-

FIGURE 13.2. Backpropagation network used to model area 7a. The input to the network consists
of retinal position and eye position information. The activity of the output units is a monotonic
function of the location of the visual stimulus in craniotopic coordinates. The middle or "hidden"
layer units map input to output. The details of the network are explained in the text.
278 REGULATION OF CORTICAL FUNCTION IN MEMORY

centered space as a linear function of spike rate. There are four output units with
pairs of opposite slope for horizontal and vertical position. As in the case with
the eye position inputs, the exact form of the function did not appear to matter
as long as it was monotonically increasing; also, whether only one unit of posi-
tive slope or two units of opposite slope were used did not seem to be important.
The rationale for using a monotonic function for the output was that the eye
position cells could be used as a teaching signal if the animal saccaded to fixate
the stimulus. In later experiments we also tried mapping to Gaussian represen-
tations of head-centered location; the interesting results are listed below.
The output of each cell in the network is calculated by first summing all
inputs, both inhibitory and excitatory, and then calculating the output as a sig-
moidal function of the input. A sigmoid is chosen as an output function since it
is similar to the operation performed by actual nerve cells that sum inputs, have
a threshold, and saturate at high levels of activity. There is also a threshold term
that can be either trained or set, and simulations using both of these options will
be discussed.
The network begins training with all the connections set to random weights
and completes training when the output units accurately indicate positions in
head-centered space for any pair of arbitrary retinal and eye position inputs. The
network learns by subtracting the output vector from the desired output vector
for each input pattern to generate an error. This error signal is then propagated
back through the network to change the weights in the network. The back-
propagation algorithm ensures that the weights will change to reduce error in the
performance of the network. The actual equations and derivation of the back-
propagation procedure are discussed in Rumelhart et al. (1986). This cycle is
repeated until the network reduces error to desired levels. The spatial transfor-
mation network learns quickly and always settles to very low error values.
Within 1000 trials the network shows accuracy that is better than the spacing of
the distance between the centers of the retinal receptive fields. For large numbers
of trials the network continues to show improvement to vanishingly small
errors.
After training is complete the middle-layer units have receptive fields that
remain retinotopic, but their activity becomes modulated by eye position in a
manner similar to that seen in the recording data from area 7a neurons. The
receptive fields remain retinotopic but the responsiveness changes with eye posi-
tion. The change in responsiveness is roughly planar and similar to a majority
of the gain fields recorded from area 7 a neurons. Moreover, the receptive fields
are large and can have peculiar shapes, not unlike the cells in area 7a.
In Figure 13.3 the retinotopic visual receptive fields from recording experi-
ments are compared with retinotopic receptive fields generated by the model. It
should be emphasized that these comparisons are intended to be qualitative and
show only that they are similar; obviously they will not be exactly the same, just
as no two area 7a neuron receptive fields will ever be identical. Surfaces have
been fit to the recording and model data using a Gaussian interpolation algo-
rithm. Each receptive field is 80° in diameter. The fields have been categorized
FIGURE 13.3. Visual receptive fields from the data and the model compared. The receptive
fields were divided into three classes: class I cells have a single, smooth peak of activity; class II
cells have one peak of activity but also other smaller peaks or depressions in the receptive field;
class III cells have multiple peaks of activity. Note the close correspondence between model- and
data-receptive fields.
280 REGULATION OF CORTICAL FUNCTION IN MEMORY

by complexity into three classes: class I has the simplest receptive fields with
each having a single peak of activity; class II fields are of intermediate complex-
ity with each field having a single greatest peak of activity and one or more
smaller peaks of activity; class III receptive fields are the most complex with
each having multiple peaks of greatest activity. The most complex fields are sim-
ilar to fields of the untrained model. The trained model seldom produces such
complex fields.
Next we compared gain fields generated by the model with data gain fields.
Figure 13.4 shows gain fields from recording data and Figure 13.5 gain fields

FIGURE 13.4. The spatial gain fields of nine cells recorded from area 7a using the technique
shown in Figure 13.1. The diameter of the inner, dark circle is proportional to the magnitude of
the visually evoked response. The outer circle diameter is proportional to the total rate (visual
response and background activity). The white annulus represents the firing rate in the absence of
visual stimulation.
 Network Model for Learned Spatial Representation 281

FIGURE 13.5. Hidden unit spatial gain fields generated by the network model, in the same
format as Figure 13.4. All the fields illustrated resulted from 10,000 training trials. Units g, h,
and i are from runs in which the craniotopic Gaussian format for the output layer was used. The
remainder of the cells were using a craniotopic monotonic format.

generated by the model's hidden unit. The nine pairs of circles for each data or
model unit represent the activity for the same retinotopic stimulus delivered at
nine different eye positions. The dark inner circle's diameter is proportional to
the response to the visual stimulus alone, and the outer circle's diameter is pro-
portional to the entire activity, both background and evoked.
Three basic types of planar gain fields were found from the recording data.
For 28%, the background and evoked activities changed in a parallel fashion
(Figure 13.4b,e,f). In the largest proportion of cells (43%), the evoked activity
changed while the background activity, if any, remained constant (Fig.
13.4a,c,d); 75% of these cells had low rates of background activity. For the
282 REGULATION OF CORTICAL FUNCTION IN MEMORY

remaining 28% of the neurons the background and evoked activities changed in
different directions, so the activity of either alone was grossly nonplanar, but the
overall activity was planar (Fig. 13.4g,h,i).
Figure 13.5 shows gain fields generated by the model; the arrangement shows
the close correspondence between recording and model gain fields. Both mono-
tonic and Gaussian output representations were used in the training that gen-
erated these fields. Either output representation generates total response gain
fields that are planar. However, when the training is made to monotonic output
function the visual response gain fields are generally planar (67%; Fig. 13.5a-f),
whereas training to the Gaussian format produces only 13% in this class. These
figures compare to the 55% found in the experimental data. If the thresholds are
trained along with the synaptic weights for monotonic outputs, then the back-
ground and evoked activities almost always change in a parallel fashion. How-
ever, if the threshold is held constant and at high values, the background activity
is low and does not change with eye position, similar to the largest proportion
of parietal cells. Finally, if training is done using a Gaussian format output, then
the visual evoked response is usually grossly nonplanar but the combination of
background and evoked responses is planar (Fig. 13.5g,h,i).

DISCUSSION OF THE MODEL

The simulation results show that training a parallel network to perform coor-
dinate transformations produces the same type of distributed code that is found
in area 7a. This similarity should be pursued to determine if it represents a fun-
damental outcome of using parallel networks to perform coordinate transfor-
mations. One line of research would be to make the model more complex by
incorporating features analogous to those found in the brain such as Hebb learn-
ing and reciprocal pathways for error feedback. It will be interesting to see if
these more complex and brainlike 'models still produce the same distributed
code. Another avenue would be to see if this model generalizes to three-dimen-
sional space- and body-centered coordinates by collecting data under these con-
ditions for parietal neurons and comparing the results to predictions made from
the model.
These results suggest that the posterior parietal cortex learns to associate body
position with respect to visual space. Thus the parietal lobe appears to form
associative memories for performing spatial transformations. A learning theory
for parietal spatial functions seems in order, since it would not be practical to
hard-wire spatial representations during development when the body is chang-
ing size. Moreover, distortions of space with prisms lead to rapid recalibration
in adults, suggesting that, unlike ocular dominance, there is no critical period
for spatial representations and they remain plastic in adults.
It is important to note that the model by definition does not have a topo-
graphic organization. Thus there is no requirement for topographic organization
in the brain. The reason topography is not necessary is that the organization of
 Network Model for Learned Spatial Representation 283

the network is distributed and the information is contained in the weights in the
synapses. It would be interesting to determine whether putting a crude topog-
raphy into the connections of the network would accelerate learning.
If the spatial representation in area 7a lacks topography, this does not of
course mean that there is not topographic organization in this area. We imagine
that learning spatial localization can occur within the dimensions of a typical
cortical column, 1 mm2. Recordings made within an area of parietal cortex of
this size contain a complete complement of receptive fields, eye position signals,
and gain fields necessary for a complete representation of craniotopic space
(Andersen, Essick, & Siegel, unpublished observation, 1985). Thus spatial loca-
tion can be mapped over and over again in many repeating units in the cortex
and may overlie some as yet unknown functional repetitive architecture that
would need in each of its modules the complete machinery for coordinate
transformations.
The rather large receptive fields suggest that every posterior parietal neuron
has access to the entire retina and the particular shape of each receptive field is
a result of competitive learning. This competition produces in both the parietal
neurons and the model units receptive fields that, although they are complex,
tend to coalesce so that they are smoothly varying, rather than random in struc-
ture. The large receptive fields are presumably due to the cascading divergence
that occurs in the multistage corticocortical projections from V1 to area 7a.
The fact that many fewer eye position synapses than retinal synapses were
requried at the convergence onto the hidden units in the model has interesting
parallels to the anatomy of the parietal lobe connections. It is believed that the
source of eye position information comes from lower brain stem centers and is
relayed through the intralaminar nuclei (Schlag-Rey & Schlag, 1984). However,
these nuclei are small compared to the cortical areas relaying visual information
to area 7a.
There is the question of where the output units of the model might exist in
the brain. Cells showing the expected eye position-independent behavior have
not been found in area 7a. Another possibility is that this distributed coding is
used throughout all brain regions that need spatial representations. The final
spatial output might be seen only in the eventual motor output. The final spatial
output may be pointing the eye or a finger accurately to a location in space and
no single cell in the brain might be found that codes the location of visual space
in an eye position-independent fashion.
Finally, there is the question of whether this form of model of distributed cod-
ing in parallel networks, which appears to explain the parietal data rather well,
will be useful in other brain regions. Recently the response of area V4 cells was
examined in a task in which the monkey must match the orientation of a visual
or somatosensory cue grating with the orientation of a visual test grating
(Haenny, Maunsell, & Schiller, in press). They find cells that respond to the cue
that are orientation tuned, cells that respond to the test stimulus that are ori-
entation tuned, and cells that show facilitated activity for a particular combi-
nation of cue and test stimulus. Thus the activity of some V4 neurons shows a
284 REGULATION OF CORTICAL FUNCTION IN MEMORY

multiplicative interaction between two inputs that is similar to the interaction

for eye and retinal inputs for area 7a neurons. It would be useful to construct a
similar network in which the inputs were the cue and test stimuli and the output
the correct match. Would the hidden units develop properties like those of V4
neurons?

Acknowledgments
We thank Carol Andersen for editorial assistance. R.A.A. was supported by
grants EY05522 and EY07492 from the National Institutes of Health, the Sloan
Foundation, and the Whitaker Health Sciences Foundation. D.Z. was supported
by grants from the System Development Foundation, ONR contract N0014-
87K-0671, and the AFOSR.

REFERENCES

Andersen, R. A. (1987). Inferior parietal lobule function in spatial perception and visuomotor
integration. In F. Plum (Ed.), Handbook of physiology (pp. 483-518). Bethesda, Md.:
American Physiological Society.
Andersen, R. A., Essick, G. K., & Siegel, R. M. (1985). Encoding of spatial location by posterior
parietal neurons. Science, 230, 456-458.
Andersen, R. A. & Zipser, D. (1988). The role of the posterior parietal cortex in coordinate
transformations for visual-motor integration. Canadian Journal of Physiology and
Pharmacology, 66, 488-501.
Haenny, P. E., Maunsell, J. H. R., & Schiller, P. H. (in press). State dependent activity in mon-
key visual cortex. Experimental Brain Research.
Hallet, P. E., & Lightstone, A. D. (1976). Saccadic eye movements toward stimuli triggered by
prior saccades. Vision Research, 16, 99.
Mays, L. E., & Sparks, D. L. (1980). Dissociation of visual and saccade-related responses in
superior colliculus neurons. Journal of Neurophysiology, 43, 207-232.
Merzenich, M. M., Recanzone, G., Jenkins, W. M., Allard, T. T., & Nudo, R. J. (in press).
Cortical representational plasticity. In P. Rakic & W. Singer (Eds.), Neurobiology of
neocortex. Dahlem konferenze. Chichester: Wiley.
Robinson, D. A. (1975). Oculomotor control signals. In P. Bach-y-Rita & G. Lernerstrand
(Eds.), Basic mechanisms of ocular motility and their clinical implications (pp. 337-
374). London: Pergamon.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning internal representations
by error propagation. In D. E. Rumelhart, J. L. McClelland, & PDF Research Group
(Eds.), Parallel distributed processing: Explorations in the microstructure of cognition:
Vol. 1. Foundations (pp. 318-362). Cambridge, Mass.: Bradford Books/MIT Press.
Schlag-Rey, M., & Schlag, J. (1984). Visuomotor functions of central thalamus in monkey. I.
Unit activity related to spontaneous eye movements. Journal of Neurophysiology, 51,
1149-1174.
Zipser, D., & Andersen, R. A. (1988). A back propagation programmed network that simulates
response properties of a subset of posterior parietal neurons. Nature. 331, 679-684.
14
Cortical Localization
of Working Memory
P A T R I C I A S. G O L D M A N - R A K I C

WORKING MEMORY

Working memory is the term applied by cognitive psychologists and theorists to

the type of memory that is relevant only for a short period of time, usually on
the scale of seconds. This criterion—relevant only transiently—distinguishes
working memory from almost every other form of memory in which the con-
tents are stable over time. Thus working memory can be distinguished easily
from what has been termed reference memory (Olton, Becken, & Handelmann,
1979), declarative memory (Squire & Cohen, 1984), semantic memory (Tulving,
1972), procedural memory (Squire & Cohen, 1984), and habit (Mishkin, Mala-
mut, & Bachevalier, 1984)—all of which have in common that their contents do
not depend on current experience or where or when the information was
learned.
A major idea about working memory is that it is a single common resource
with a limited capacity. Indeed, Baddeley (1986) argued that there is a process
of general working memory (his WMG) that can be used concurrently by a range
of cognitive tasks: learning, reading, comprehension, and so on. A significant
issue in current research on the psychology of working memory is whether there
may, in fact, be multiple memory systems, each dedicated to different informa-
tion-processing domains, or just one central executive system that is shared by
all concurrent tasks (Baddeley, 1986; Squire, 1987). As will be discussed later,
neuroscience has much to contribute toward elucidating this issue.
The distinction between working and other forms of memory has been
invoked less frequently by physiological psychologists. Olton can be credited as
the researcher who brought this distinction to the attention of neuroscientists
and demonstrated its usefulness in studies of maze behavior in rats (Olton et al.,
1979). Otherwise, the concept has not been given prominence in reviews of the
neuropsychological literature on the role of the limbic or diencephalic structures
in memory (Mishkin et al., 1984; Squire, 1982, 1986) and correspondingly little

285
286 REGULATION OF CORTICAL FUNCTION IN MEMORY

has been said about the neural structures or mechanisms underlying working
memory. This is surprising because several of the tasks used to study memory
in human and nonhuman primates have strong working memory components;
that is, matching-to-sample and nonmatching-to-sample tasks engage working
memory for object features.
In this chapter I will review old and new evidence that the brain obeys the
distinction between working and associative forms of memory and I will argue
that the prefrontal cortex has a specialized, perhaps even preeminent, role in
working memory processes. Moreover, there is already considerable evidence to
suggest that more than one working memory system exists in prefrontal cortex
and that still others may be disclosed. Perhaps through the study of experimental
animals more can be learned about the neurobiological principles underlying
these fundamental processes which govern our behavior and which decompose
in numerous neurological diseases.

DELAYED-RESPONSE TASKS: TESTS OF WORKING MEMORY

The delayed-response tests were designed and introduced to psychology by the

comparative psychologist Walter Hunter (1913). His intention was to devise a
test that would differentiate among animals (including humans) on their relative
ability to respond to situations on the basis of stored information, rather than
on the basis of "immediate" stimulation. The classical version of the delayed-
response test is essentially a hiding task in which the subject is shown the loca-
tion of a food morsel that is then hidden from view by an opaque screen (Fig.
14.1). Following a delay period of several seconds, the subject chooses the cor-
rect location out of two or more choices. In this situation the subject must
remember where the bait had been placed a few seconds earlier. In a variation
of this task, spatial delayed alternation, the subject is required to alternate
between left and right foodwells on successive trials that are separated by delay
periods (Fig. 14.1). Since in the latter task the animal is not permitted to observe
the experimenter baiting the foodwells in the delay periods, to be correct on any
given trial the subject must keep track of which response was made last. A cru-
cial feature of these tests is the lack of a fixed relationship between stimuli and
correct responses. Unlike the procedure employed in associative learning or con-
ditioning paradigms, although the stimulus arrangement at the time of response
is identical on each trial, the correct response varies depending on information
presented several seconds earlier. Thus the correct response is different on every
trial and information relevant to the response on one trial is irrelevant on the
next trial and actually best forgotten, lest it interfere with the current response.
The same underlying principle operates in other commonly used behavioral par-
adigms: object alternation and match-to-sample or nonmatch-to-sample
paradigms.
Delayed-response function has been associated with prefrontal cortex since
the early discovery that monkeys with prefrontal cortex damage were severely
 Cortical Localization of Working Memory 287

FIGURE 14.1. Two working memory tests, delayed spatial response and delayed-spatial alter-
nation. Dorsolateral prefrontal lesions produce impairments on these tasks, but not on associative
memory tasks such as visual pattern discriminations. (Modified from Friedman and Goldman-
Rakic, in press; see Goldman-Rakic, 1987b, for review.)

impaired on this and related tasks (Jacobsen, 1936). Although the association
between the capacity to perform delayed-response tasks and the functional integ-
rity of the dorsolateral prefrontal cortex is firmly established, the nature of the
process(es) tapped by these tasks has been the subject of continuing discussion
and seemingly contradictory interpretations. Jacobsen stressed the mnemonic or
temporal-sequential processes required by delayed-response tasks (and hence
mediated by prefrontal cortex), but this explanation somehow did not take hold.
In retrospect, it seems evident that in the thirties, and for several decades later,
memory was considered a unitary mechanism thought to have a single location.
288 REGULATION OF CORTICAL FUNCTION IN MEMORY

As patients and monkeys with prefrontal lesions displayed surprisingly adequate

performance on discrimination learning, object reversal, and conditional mem-
ory tasks and, in addition, patients displayed remarkably normal scores, at least
on conventional IQ tests, the impairment produced by prefrontal damage was
not global enough to fit the neuropsychologist's midcentury concept of a mne-
monic deficit. Thus various investigators sought other explanations—for exam-
ple, spatial perception and/or spatial orientation (e.g., Pohl, 1973; Pribram &
Mishkin, 1956), attentional functions (Bartus & LeVere, 1977; Malmo, 1942),
and motor control functions (Teuber, 1972)—and some considered disturbance
in the central representations of kinesthesis (Konorski, 1967) or proprioception
(Gentile & Stamm, 1972) as basic to the prefrontal deficit.

NEUROPHYSIOLOGICAL EVIDENCE FOR WORKING MEMORY

A major development in our understanding of delayed-response tasks and pre-

frontal cortex came in the early seventies when unit recording studies were per-
formed for the first time in awake behaving monkeys trained on delayed-
response tasks (Fuster & Alexander, 1970; Kubota & Niki, 1971). These studies
revealed that neurons in the prefrontal cortex became activated during the delay
period of a delayed-response trial. It was difficult to resist the hypothesis that
these prefrontal neurons were the cellular expression of a mnemonic process. A
similar conclusion had been reached in a paper that H. E. Rosvold and I pub-
lished one year earlier on the cortical localization of delayed-response deficits
and the nature of the difficulty experienced by monkeys with bilateral lesions on
delayed-response tasks. We suggested that cortex around the principal sulcal cor-
tex was involved in what we termed at the time "modality-specific" memory
(Goldman & Rosvold, 1970). In a series of studies (Goldman & Rosvold, 1970;
Goldman, Rosvold, Vest, & Galkin, 1971), we compared monkeys with various
circumscribed prefrontal lesions on spatial tasks (delayed-response and delayed
alternation) which contained temporal delays between spatial cues and spatial
responses and on other tasks (conditional position response and go-no go alter-
nation) which removed the delay or the spatial feature, respectively, from the
behavioral task. Our goal in these studies was to determine the importance of
the delay relative to the spatial features of tasks on which monkeys with pre-
frontal lesions were impaired. The results were clear-cut. Monkeys with princi-
pal sulcus lesions were impaired only on tasks that contained both spatial cues/
responses and a delay, but they performed well on tasks that contained only a
delay or only a spatial cue/response. The full significance of our experiments and
conclusions are clearer now than they were in the early seventies, for they pro-
vided evidence of a working memory system dedicated to spatial vision—that
did not overlap with memory mechanisms for other (e.g., nonspatial) domains
of knowledge.
The evidence from neurophysiological studies of prefrontal neurons has been
steadily accumulating since 1970. Since that time, numerous studies have pro-
vided evidence that certain classes of prefrontal neurons are particularly acti-
 Cortical Localization of Working Memory 289

vated during delay periods (for review, see Fuster, 1980). Yet the interpretation
that such activity reflected a memory process was still not widely appreciated
outside of a narrow circle of scientists working on the primate prefrontal cortex.
Perhaps one reason was that neurons activated during the delay of a delayed-
response trial were not necessarily coding specific information, but rather were
engaged in some sort of preparatory or motor set to respond. Further, prepara-
tion to respond could invoke postural mechanisms rather than central processes.
Whatever the reason, scientists did not seem to be impressed with the delay-
related activation of prefrontal neurons. However, some recent evidence from
studies that I have been conducting with Shintaro Funahashi and Charles Bruce
have provided the most convincing evidence yet of mnemonic processing in pre-
frontal cortex. We have been using an oculomotor delayed-response paradigm
to study prefrontal function (Funahashi, Bruce, & Goldman-Rakic, 1989). The
advantage of this paradigm over other methods of studying delayed-response
performance is that we require the animal to fixate a spot of light on a television
monitor and maintain fixation during the brief (0.5-sec) presentation of a light
followed by a delay period of variable length. Visual targets can be presented in
any part of the visual field and, importantly, we have complete control over the
specific information that the animal has to remember on any given trial. The
fixation point is turned off only at the end of the delay period and its offset is
our instruction to the animal to break fixation and direct its gaze to wherever
the target had been presented. Under these circumstances high levels of perfor-
mance can be attained only if the monkey uses mnemonic processing. In the
oculomotor task, as in manual delayed-response tasks, prefrontal neurons
increase (or often decrease) their discharge rate during the delay period of a trial.
The neuron displayed in Figure 14.2. is an example: its activity rises sharply at
the end of the stimulus, remains tonically active, and then ceases rather abruptly
at the end of the delay.
The activation of prefrontal neurons when a stimulus disappears from view
and the maintenance of that activation until a response is executed is highly
suggestive of a working memory process. However, other interpretations are
possible. For example, it could still be argued that the delay period activity
reflects simply the motor set of the animal to respond at the end of the delay, as
has been described for neurons in premotor fields (Mauritz and Wise, 1986;
Tanji, 1985). However, we and others showed that prefrontal neurons have
directional delay-period activity, that is, they exhibit increased or decreased dis-
charge only for specific targets. Our own studies testing behavioral performance
and single neuron activity with stimuli presented in eight different locations
showed that prefrontal neurons have "memory fields." Monkeys performed this
task with proficiency for every target location, but most neurons that we
recorded in prefrontal cortex altered their discharge rate only for one or a few
specific target locations. A general motor set explanation cannot easily account
for this result.
If prefrontal neurons are involved in mnemonic processing, one would expect
their activity to be sensitive to changes in the duration of the delay period.
Indeed, it has been shown that the activity of prefrontal neurons in the delay
290 REGULATION OF CORTICAL FUNCTION IN MEMORY

FIGURE 14.2. Single unit activity of a prefrontal neuron recorded during performance on an
oculomotor delayed-response task. The neuron's "memory field" is for the target at 270°. Activity
was enhanced during the delay period only when the target to be remembered was at that location
and not at any other tested. (Modified from Funahashi et al., 1989)

period expands and contracts as the delay duration is lengthened or foreshort-

ened (Kojima & Goldman-Rakic, 1982). Again, this would be expected if the
neuron is coding information that is to be retained until the end of the delay
period.

ANATOMICAL EVIDENCE

The principal sulcus in the prefrontal cortex is the anatomical focus for spatial
delayed-response function and knowledge of its connections with other struc-
tures is helping researchers to understand the circuit and cellular basis of work-
ing memory. It has become clear that this prefrontal subdivision has reciprocal
 Cortical Localization of Working Memory 291

connections with more than a dozen distinct association regions, with premotor
centers, with the caudate nucleus, superior colliculus, and brain stem centers (for
review see Goldman-Rakic, 1987b). One major goal of research in my laboratory
is to relate the different connections of prefrontal cortex to its capacity to medi-
ate delayed-response performance by decomposing the task into its composite
subfunctions. For this purpose, my colleagues and I proposed that in general the
ability to guide behavior by working memory requires mechanisms for (1) select-
ing or accessing pertinent information; (2) holding that information "on-line"
for the temporal interval over which a decision or operation is to be performed;
and (3) executing motor commands. In the case of the delayed-response task,
the animal at a minimum must (1) attend to the relevant visuospatial coordi-
nates; (2) remember the most recent preceding event (where the bait was placed
or which response was made last); and (3) select and execute the appropriate
manual or oculomotor response while inhibiting the inappropriate response.
Each of these component processes is presumably mediated by different path-
ways that comprise the neural circuitry of the principal sulcus. I have suggested
that the parietoprefrontal projection system is one that may supply the sensory-
representational information to the prefrontal cortex, and connections with the
caudate and superior colliculus, among other motor centers, play a role in relay-
ing motor commands (Goldman-Rakic, 1987b). However, the mechanism for
holding information in the short-term memory buffer is a total mystery at this
point and the role in this process of any connections that the prefrontal cortex
has with other centers, including memory centers, is not clear. The parietopre-
frontal circuit described above may be crucial. At the same time, we described
several multisynaptic connections between the prefrontal cortex and the hip-
pocampal formation and speculated that these connections imply a cooperative
functional relationship between the hippocampus and the prefrontal cortex
(Goldman-Rakic, Selemon, & Schwartz, 1984). As described below, our recent
examination of metabolic rate in the dentate gyms and hippocampus in mon-
keys performing working memory tasks supports this line of thinking (Friedman
& Goldman-Rakic, 1988).

METABOLIC STUDIES OF WORKING AND ASSOCIATIVE MEMORY

One strategy that we have been using to test the functional significance of vari-
ous prefrontal connections is the 2-deoxyglucose method of functional mapping.
We used this method to examine the metabolic rate in the dentate gyrus and
hippocampal subfields as well as other neural structures in monkeys trained to
perform working memory tasks and compared their brain metabolism with that
of monkeys performing other types of memory and/or control tasks. We showed
that metabolic rate increases about 20% in specific parts of the trisynaptic path
in monkeys performing a variety of working memory tasks including delayed-
response as compared with those performing visual pattern or sensorimotor
tasks (Friedman & Goldman-Rakic, 1988).
Our results not only implicate the hippocampus and dentate gyrus in a coop-
292 REGULATION OF CORTICAL FUNCTION IN MEMORY

erative relationship with the prefrontal cortex; they also allowed us to identify
other parts of the neural system or circuitry engaged by working memory. For
example, we recently showed that the mediodorsal nucleus and anterior nuclei
are selectively driven in terms of enhanced metabolic rate by working memory
requirements. However, the amygdala (Friedman & Goldman-Rakic, 1988) and
mammillary bodies (Friedman, Janas, & Goldman-Rakic, 1987) exhibit the
same metabolic rate for associative memory and sensorimotor tasks as they do
for working memory tasks. The 2-deoxyglucose studies in normal monkeys per-
forming tasks that vary in their mnemonic requirements should shed light on
the components of a neural system that are (or are not) crucial for selective
memory processes. Our preliminary results indicate that there is a network of
interconnected cortical and subcortical areas and hence an elaborate network of
structures that is activated by the working memory requirement in many mne-
monic tasks. Although the precise mechanisms are unknown, it seems that this
anatomical system includes at minimum the dorsolateral prefrontal cortex, the
trisynaptic pathway of the hippocampus, and selected medial and anterior tha-
lamic nuclei.

SIGNIFICANCE OF WORKING MEMORY: REGULATION OF

BEHAVIOR BY REPRESENTATIONAL KNOWLEDGE
VERSUS REGULATION BY IMMEDIATE STIMULATION

The significance of working memory for higher cortical function is widely rec-
ognized and attested to by cognitive psychologists. However, one reason it may
not have received the attention it deserves among neuroscientists is that its
importance is not self-evident. Perhaps even the quality of its transient nature
misleads us into thinking it is somehow less important than the more permanent
archival nature of long-term memory. However, the working memory function
of our brains may be its most inherently flexible mechanism and its evolution-
arily most significant achievement. Thus working memory confers the ability to
guide behavior by representations of the outside world rather than by immediate
stimulation. It is the cornerstone of the human capacity to base behavior on
ideas and thoughts and is critical for linguistic processing.
This capacity is not easily won: prosimian primates have much greater diffi-
culty learning delayed-response tasks than do rhesus monkeys, and some normal
galagos even fail to learn this task (Preuss & Goldman-Rakic, unpublished
observation, 1988). Cats and dogs have not been shown to perform a delayed-
response task given in the same manner as it is given to monkeys. Only if the
animals are allowed to keep the hidden object in view during the delay period,
as in a Nencki test apparatus, is delayed-response performance attained (War-
ren, Warren, & Akert, 1972). One may even question whether rhesus monkeys
are using working memory processes if allowed to keep the to-be-remembered
object in view during a delay period. The opportunity to view the target during
the delay period could obviate the whole purpose of the delayed-response task.
 Cortical Localization of Working Memory 293

As working memory is the process that allows behavior to be regulated by

representations, it may underlie our ability to form concepts—representations
of classes of stimuli of particular color, size, or form. It is significant, therefore,
that the neuropsychological deficit most widely recognized as being associated
with prefrontal dysfunction in humans is that revealed by a test of categoriza-
tion, the Wisconsin Card Sort Test. In this task, the subject is asked to sort a
deck of cards which bear stimuli that vary in number, color, and shape. As each
card in the stack comes up, the subject has to match it to a set of reference cards
on the basis of one dimension that is arbitrarily selected by the experimenter
(e.g., color). The experimenter then informs the subject if he or she is "right" or
"wrong" and the patient tries to get as many correct matches as possible. After
the patient achieves a specified number of consecutive correct matches, the sort-
ing principle is shifted without warning (e.g., to shape or number) and the
patients must modify their responses accordingly. Patients with right or left
hemisphere prefrontal damage exhibit difficulty in switching from one category
to another (Milner, 1963, 1964).
At first glance there would seem to be little similarity between the Wisconsin
Card Sort Test used with frontal lobe patients and the delayed-response task
used with lesioned monkeys. Commonalities between these two tasks have
rarely been claimed. However, the Wisconsin Card Sort Test does resemble the
delayed-response tasks in one essential respect. Although the relevant features
of the stimuli (color, size, shape) are all present at the time of response, there is
nevertheless no information given about the correct response; it must be pro-
vided from representational memory, perhaps in this case a second-order rep-
resentation (representation of a representation), that is, a concept for guiding
response choice. Although the capacity for this type of second-order represen-
tational system appears to be a unique acquisition of human intelligence,
undoubtedly linked to the emergence of language, it clearly must be built on a
first-order representational capacity shared with other mammals.
Another concept that may depend on working memory is object permanence:
the capacity to recognize that an object has continuity in time and space when
out of view. It may not be coincidental, therefore, that Piaget's Stage IV Object
Permanence Test (OPT) is formally similar to delayed-response. As in delayed
response, the child watches as a reward is hidden in one of two spatially sepa-
rated locations; then, after a delay of a few seconds, the child is allowed to find
the reward. We have shown that the performance of adult rhesus monkeys with
dorsolateral prefrontal lesions on the Piagetian test is remarkably similar to that
of human infants on the same test (Diamond & Goldman-Rakic, 1989). In mon-
keys, moreover, the capacity to perform the OPT has been shown to develop
concomitantly with the emergence of delayed-response ability during the first
half-year of life (Diamond & Goldman-Rakic, 1989). On the basis of their com-
mon dependence on prefrontal cortex and on what we know and can theorize
about prefrontal cortex, we conclude that both tasks measure a common process.
If our reasoning has merit, this process may be considered a building block of
concept development in humans. In any event, the elemental capacity to form
294 REGULATION OF CORTICAL FUNCTION IN MEMORY

representations of the outside world and especially to base responses on those

representations in the absence of the objects they represent is a nontrivial
achievement of the mammalian cortex.
In addition to the Wisconsin Card Sort Test and Piaget's Object Permanence
Test, a variety of other tests may depend to some extent on working memory
and thus reflect prefrontal cortical function. These include the Stroop Test, in
which the names of colors are printed in the ink of another color (the word yel-
low is printed in red) and subjects are instructed to report the color of the word
while suppressing the natural or prepotent tendency to report the written content
of the word. Here again subjects must guide their response by a representational
memory, in this case the memory of an instruction. Patients with prefrontal
lesions are impaired on the Stroop Test (Perret, 1974) and on tasks in which
they must keep track of the recency or order of their previous responses (Petrides
& Milner, 1982) as well as project sequences of future responses on "look-
ahead" puzzles, like the Tower of London test (Shallice, 1982). In the latter task,
a goal has to be decomposed into subgoals and the subgoals must be tackled in
the correct order. Such tasks, like delayed-response tasks, require an ability to
base current responses on representations of past or future events. Finally,
patients with prefrontal damage have difficulty in generating new responses.
Their well-documented impairments in generating word lists (Milner, 1964; Per-
ret, 1974) and in inventing abstract designs (Jones-Gotman & Milner, 1977) and
reproducing gestures might be expected and predicted if they suffer an impov-
erished ability to guide behavior on the basis of representations or ideas, and if
this ability itself depends on working memory. Thus it may be that a seemingly
diverse set of measures, rather than tapping into diverse processes, shares a com-
mon neurobiological mechanism—working memory.

EVOLUTIONARY SIGNIFICANCE OF WORKING MEMORY

The type of behavioral regulation subserved by prefrontal cortex can be con-

trasted with the formation of stimulus-response associations, which depend in
an obligatory manner on the repeated association of stimulus-response pairings
and/or the amplitude, duration, and frequency of stimuli in the environment.
Indeed, the evolution of a capacity for representational knowledge may provide
a mechanism for overriding the governance of behavior by conditioned, reflex-
ive, or innate prepotent responses, particularly when such behavior could be
maladaptive for the organism. The fact that delayed-response impairments pro-
duced by prefrontal lesions are absent at nil delays between stimulus and
response, yet are nearly as severe at delays of 1 to 2 sec as at much longer delays
(Goldman, 1971), suggests a sharp dissociation in the neural mechanisms that
mediate representational and strictly sensory guidance of behavior. Working
memory, as previously indicated, allows information coming in at one point in
time to be transiently associated with information encountered at a later time.
In contrast, the mechanisms of associative memory are precisely designed to
preserve or "stamp in" lasting relationships between stimuli and responses.
 Cortical Localization of Working Memory 295

Finally, efficient representational memory requires information to be dismissed

as efficiently as it is accessed; this type of memory demands a flexible, nonbind-
ing relationship between a given stimulus and a given response. Thus the mem-
ory mechanisms of prefrontal cortex could be expected to differ from those that
mediate associative learning.
The functional dissociation of working and associative memory mechanisms
is supported by anatomical localization studies. Decortication spares certain
forms of conditioned responses, and many classically conditioned responses
have been localized to the amygdala, diencephalon, brain stem, and spinal cord
(for review see Thompson, 1986). However, associative learning is completely
spared by prefrontal and also hippocampal lesions (for reviews see Goldman-
Rakic, 1987a; Squire, 1986). It therefore seems possible that the type of behavior
which utilizes memory in the representational, working memory, sense requires
participation of the prefrontal cortex, whereas that involving associative learn-
ing may rely more heavily on subcortical structures and/or other areas of neo-
cortex. The evolution of a capacity to guide behavior on the basis of represen-
tations of the outside world rather than on the trigger of immediate stimulation
introduces the possibility that concepts and plans may govern our behavior. One
way they may do so is by regulating, facilitating, or inhibiting the associatively
conditioned or reflexly organized responses of motor centers to which the pre-
frontal cortex projects (Goldman-Rakic, 1987c).
Although the mechanisms of working and associative memory are undoubt-
edly different, their contributions to behavior can be thought of as complemen-
tary. Associative mechanisms are the basis of our vast store of experience and
knowledge; those of working memory provide a means for utilizing that knowl-
edge in a flexible manner. Thus a complete understanding of behavioral regu-
lation would entail a detailed analysis of the relationships and interconnections
between mechanisms of associative and working memory.

THOUGHT DISORDERS: LOSS OF WORKING MEMORY

Consider what happens when working memory is lost, as it may be, in patients
and in experimental animals with prefrontal cortex damage. Studies of the loss
of this capacity may help researchers understand certain well-established fea-
tures of behavior expressed both by patients and by monkeys with prefrontal
lesions. One such feature is distractibility. In a hypothetical delayed-response
trial, a normal animal would form an internal representation of the position of
the bait during the cue period and maintain reference to that representation for
as long as is necessary to guide its response. Access to this knowledge during the
delay period bridges the temporal gap between cue and response and enables the
animal to select and execute the correct response; at the same time, inappropri-
ate responses to irrelevant or distracting external stimuli that may command the
animal's attention at the moment of response are, of necessity, suppressed.
Deprived of access to representational knowledge, the opposite would be true:
an animal would almost be forced to rely on external stimuli present at the time
296 REGULATION OF CORTICAL FUNCTION IN MEMORY

of response; it would be at the "mercy" of environmental cues. Indeed, a dom-

inant feature of the frontal lobe syndrome in both human and nonhuman pri-
mates is an excessive dependence on, or sensitivity to, external cues (e.g.,
humans: Lhermitte, 1986; Woods & Knight, 1986; monkey: Bartus & LeVere,
1977; Malmo, 1942). The monkey with prefrontal lesions seems to respond to
the first food well that catches its eye or to the just previously reinforced side
rather than to recently stored information about the location of the reward. The
description "out of sight, out of mind" (Jacobsen, 1936) or "animal of the
moment" (Bartus & LeVere, 1977) applies very well to the mindless perfor-
mance observed in prefrontally lesioned patients and experimental animals.
Thus it is possible that the capacity to guide behavior on the basis of internalized
representations marvelously accomplishes the dual task of initiating the correct
response and perforce disallowing or inhibiting the incorrect one that is presum-
ably mediated by lower centers. Although the mechanisms for issuing both facil-
itatory and inhibitory motor commands may be expected to be complex, the
prefrontal cortex has multiple connections with the motor centers by which
these complementary control functions could be exerted (Goldman-Rakic,
1987c).
Loss of a mechanism for guiding behavior by representational knowledge
might also explain the perseveration of incorrect responses that is so character-
istic of animals and patients with prefrontal lesions. If they are destined to
respond to situations on the basis of information available at the time of
response, they would be subject not only to control by external stimuli but to
conditioning by adventitious pairing of stimuli and responses. If a response on
any given delayed-response trial happened to be reinforced, there would be a
strong tendency to repeat that response, even in the face of intervening infor-
mation indicating that the same response is no longer appropriate. Perseveration
of previously rewarded responses is exactly what prefrontally damaged subjects
do. The integrity of the prefrontal cortex and access to representational memo-
ries may be necessary to override the tendency to behave strictly on the basis of
reinforcement contingencies or on the basis of stimulation present at the
moment of response and hence may enable the normal monkey to suppress such
tendencies. Viewed in this way, perseveration and distractibility are conse-
quences rather than primary causes of a more basic defect in the regulation of
behavior by representations. Since these are common symptoms of numerous
neurological conditions, understanding of the neural mechanisms subserving
working memory may be expected to shed light on human diseases that affect
cognition.

REFERENCES

Baddeley, A. D. (1986). Working memory. New York: Oxford University Press.

Bartus, R. T., & LeVere, T. E. (1977). Frontal decortication in rhesus monkeys: A test of the
interference hypothesis. Brain Research, 119, 223-248.
 Cortical Localization of Working Memory 297

Diamond, A., & Goldman-Rakic, P. S. (1989). Evidence for involvement of prefrontal cortex
in cognitive changes during the first year of life: Comparison of human infants and
rhesus monkeys on Piaget's AB task. Experimental Brain Research, 74, 24-40.
Friedman, H., & Goldman-Rakic, P. S. (1988). Activation of the hippocampus by working
memory: A 2-deoxyglucose study of behaving rhesus monkeys. Journal of Neurosci-
ence, 8, 4693-4706.
Friedman, H., Janas, J., & Goldman-Rakic, P. S. (1987). Metabolic activity in the thalarnus
and mammillary bodies of the monkey during spatial memory performance. Society
for Neuroscience Abstracts, 13, 206.
Funahashi, S., Bruce, C, & Goldman-Rakic, P. S. (1989). Mnemonic coding of visual space by
neurons in the monkey's dorsolateral prefrontal cortex revealed by an oculomotor
delayed-response task. Journal of Neurophysiology, 61, 1-19.
Fuster, J. M. (1980). The prefrontal cortex. New York: Raven.
Fuster, J. M., & Alexander, G. E. (1970). Delayed-response deficit by cryogenic depression of
frontal cortex. Brain Research, 20, 85-90.
Gentile, A. M., & Stamm, J. S. (1972). Supplementary cues and delayed-alternation perfor-
mance of frontal monkeys. Journal of Comparative and Physiological Psychology, 80,
230-237.
Goldman, P. S. (1971). Functional development of the prefrontal cortex in early life and the
problem of neuronal plasticity. Experimental Neurology, 32, 366-387.
Goldman, P. S., & Rosvold, H. E. (1970). Localization of function within the dorsolateral pre-
frontal cortex of the rheusus monkey. Experimental Neurology, 27, 291-304.
Goldman-Rakic, P. S. (1987a). Circuitry of the primate prefrontal cortex and the regulation of
behavior by representational memory. In F. Plum (Ed.), Handbook of physiology:
Sec. 1. The nervous system: Vol. 5. Higher functions of the brain (pp. 373-417).
Bethesda, Md.: American Physiological Society.
Goldman, P. S., Rosvold, H. E., Vest, B., & Galkin, T. W. (1971). Analysis of the delayed
alternation deficit produced by dorsolateral prefrontal lesions in the Rhesus monkey.
Journal of Comparative and Physiological Psychology, 77, 212-220.
Goldman-Rakic, P. S. (1987b). Development of cortical circuitry and cognitive function. Child
Development, 58, 642-691.
Goldman-Rakic, P. S. (1987c). Motor control functions of the prefrontal cortex. In R. Porter
(Ed.), Motor areas of the cerebral cortex (CIBA Foundation Symposium, No. 132, pp.
187-200). Chichester: Wiley.
Goldman-Rakic, P. S., Selemon, L. D., & Schwartz, M. L. (1984). Dual pathways connecting
the prefrontal cortex with the hippocampal formation and parahippocampal cortex
in the rhesus monkey. Neuroscience, 12, 719-743.
Hunter, W. S. (1913). The delayed reaction in animals and children. Behavioral Monographs,
2, 1-86.
Jacobsen, C. F. (1936). Studies of cerebral function in primates. Comparative Psychology Mon-
ographs, 13, 1-68.
Jones-Gotman, M., & Milner, B. (1977). Design Fluency: The invention of nonsense drawings
after focal lesions. Neuropsychologia, 15, 653-674.
Just, M. A., & Carpenter, P. A. (1985). Cognitive coordinate systems: Accounts of mental rota-
tion and individual differences in spatial ability. Psychological Review, 92, 137-172.
Kojima, S., & Goldman-Rakic, P. S. (1982). Delay-related activity of prefrontal cortical neu-
rons in rhesus monkeys performing delayed-response. Brain Research, 248, 43-49.
Konorski, J. (1967). Integrative activity of the brain. Chicago: University of Chicago Press.
Kubota, K., & Niki, H. (1971). Prefrontal cortical unit activity and delayed cortical unit activity
and delayed alternation performance in monkeys. Journal of Neurophysiology, 34,
337-347.
Lhermitte, F. (1986). Human anatomy and the frontal lobes. II. Patient behavior in complex
and social situations: The "environmental dependency syndrome." Annals of Neu-
rology, 19, 335-343.
298 REGULATION OF CORTICAL FUNCTION IN MEMORY

Malmo, H. P. (1942). Interference factors on delayed response in monkeys after removal of

frontal lobes. Journal of Neurophysiology, 5, 295-308.
Mauritz, M. H., & Wise, S. P. (1986). The premotor cortex of Rhesus monkeys: Neuronal activ-
ity before predictable environmental events. Experimental Brain Research, 61, 229-
244.
Milner, B. (1963). Effects of different brain lesions on card sorting. Archives of Neurology, 9,
100-110.
Milner, B. (1964). Some effects of frontal lobectomy in man. In J. M. Warren & K. Akert (Eds.),
The frontal granular cortex and behavior (pp. 313-334). New York: McGraw-Hill.
Mishkin, M., Malamut, B., & Bachevalier, J. (1984). Memories and habits: Two neural systems.
In G. Lynch, J. L. McGaugh, & N. M. Weinberger (Eds.), Neurobiology of learning
and memory (pp. 65-67). New York: Guilford.
Olton, D. S., Becken, J. T., & Handelmann, G. E. (1979). Hippocampus, space and memory.
Behavioral Brain Science, 2, 313-365.
Perret, E. (1974). The left frontal lobe of man and the suppression of habitual responses in
verbal categorical behaviour. Neuropsychologia, 12, 323-330.
Petrides, M., & Milner, B. (1982). Deficits on subject-ordered tasks after frontal- and temporal-
lobe lesions in man. Neuropsychologia, 20, 249-262.
Piaget, J. (1954). The construction of reality in the child. New York: Basic Books.
Pohl, W. (1973). Dissociation of spatial discrimination deficits following frontal and parietal
lesions in monkeys. Journal of Comparative Psychology, 82, 227-239.
Pribram, K. H., & Mishkin, M. (1956). Analysis of the effects of frontal lesions in monkey: III.
Object alternation. Journal of Comparative and Phsyiological Psychology, 49, 41-45.
Shallice, T. (1982). Specific impairments in planning. Philosophical Transactions of the Royal
Society of London B, 298, 199-209.
Squire, L. R. (1982). The neuropsychology of human memory. Annual Review of Neuroscience,
5, 241-273.
Squire, L. R. (1986). Mechanisms of memory. Science, 232, 1612-1619.
Squire, L. R. (1987). Memory and brain. New York: Oxford University Press.
Squire, L. R., & Cohen, N. J. (1984). Human memory and amnesia. In G. Lynch, J. L.
McGaugh, & N. M. Weinberger (Eds.), Neurobiology of memory and learning (pp. 3-
77). New York: Guilford.
Tanji, J. (1984). The neuronal activity in the supplementary motor area of primates. Trends in
Neuroscience, 7, 282-285.
Tanji, J. (1985). Comparison of neuronal activities in monkey supplementary and precentral
motor areas. Behavioural Brain Research, 18, 137-142.
Teuber, H.-L. (1972). Unity and diversity of frontal lobe functions.' Acta Neurobiologia Exper-
imentalis, 32, 615-656.
Thompson, R. D. (1986). The neurobiology of learning and memory. Science, 233, 941-947.
Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.),
Organization of memory (pp. 381-403). New York: Academic Press.
Warren, J. M., Warren, H. B., & Akert, K. (1972). The behavior of chronic cats with lesions in
the frontal association cortex. Acta Neurobiologia. Experimentalis, 32, 345-392.
Woods, D. L., & Knight, R. I. (1986). Electrophysiologic evidence of increased distractibility
after dorsolateral prefronta! lesions. Neurology, 36, 212-216.
III
Representations:
Beyond the Single Cell
This page intentionally left blank
Introduction
G O R D O N L. SHAW

Theoretical studies of the brain have been under way for several decades. For
example, the pioneering works of Turing (1936), McCulloch and Pitts (1943),
Hebb (1949), MacKay and McCulloch (1952), Cragg and Temperley (1954),
Beurle (1956), von Neumann (1956), Lettvin, Maturana, McCulloch, and Pitts
(1959), Rochester, Holland, Haibt, and Duda (1956), Caianiello (1961), Farley
and Clark (1961), Rail (1962), Rosenblatt (1962), and Uttley (1962) all were pub-
lished over 25 years ago. Over the years since, there have been relatively few
brain theorists, doing their research in a somewhat isolated fashion and publish-
ing in a range of journals (see Shaw & Palm, 1988). These factors kept brain
theory from playing as important a role as it should have.
An explosion occurred in the early 1980s as physicists, engineers, mathema-
ticians, computer scientists, and biologists from other fields started working in
theoretical neurobiology (or, perhaps more accurately, mostly in artificial intel-
ligence). At the same time, experiemental neurobiologists became critically
aware that in going "beyond the single cell" there needed to be a close, contin-
uing interplay of theory and experiment. That is not to say that theorists and
experimentalists had not spoken to each other, but they certainly had not in any
formal setting.
In some ways, brain theory has been ahead of its time. The brilliant concep-
tual principle for correlated pre-post synaptic plasticity put forward by Hebb in
his remarkable 1949 book, The Organization of Behavior, was almost immedi-
ately incorporated into mathematical models of memory in neural networks and
then into computer simulations of these models. The pioneering neurophysiol-
ogy experiments of Hubel and Wiesel in visual cortex were recognized and
referred to by brain theorists in the early 1960s (e.g., Uttley, 1962).
To understand memory and cortical function, I believe we will need to dis-
cover the fundamental principles of cortical function and incorporate them into
mathematical models. Studies connecting networks of "neurons" with a Hebb-
type learning rule, which contributed to large advances in the past, are clearly
not sufficient. Some of my favorite candidate principles of cortical function are
given below.

301
302 REPRESENTATIONS: BEYOND THE SINGLE CELL

HEBB SYNAPSE

Although there is increasing experimental evidence that the synaptic plasticity

dependent on correlated pre-post synaptic firing activity as predicted by Hebb
in 1949 is present, many questions remain to be answered: What are the exact
rules of this plasticity with respect to pre-post timing relations? Can the strength
of the synapse increase only as postulated by Hebb, or can it also decrease as
required by certain classes of mathematical models? What is the role of the net-
work? Interesting results related to this point are discussed by Singer (Chapter
10, this volume). What are the time scales of the changes? Do these Hebb-syn-
aptic changes occur after 50 msec, as suggested by Malsburg (Chapter 18) as
being crucial for short-term memory? What is the role of modulator
neuropeptides?

MOUNTCASTLE COLUMNAR ORGANIZATIONAL PRINCIPLE

Mountcastle (1978) proposed that the well-established cortical column (Gold-

man-Rakic, 1984), roughly 500 u in diameter, is the basic network in the cortex
and is comprised of small irreducible processing subunits (perhaps minicol-
umns, roughly 100 u in diameter). The subunits are connected into columns or
networks that have the capability of being excited into complex spatial-tempo-
ral firing patterns. The assumption is that higher, complex mammalian pro-
cesses involve the creation and transformation of complex spatial-temporal neu-
ronal firing patterns. This is in contrast to the assumption that the "code" for
information processing and memory recall involves only sets of neurons firing
with high frequency (the "grandmother" cell code). Perhaps there are several
codes in the CNS for communication among various regions, depending on the
sophistication of the information being analyzed and its urgency. In simple
terms, what are the next spatial and temporal scales for information processing
and memory recall beyond those associated with the single neuron? I suggest
that this next spatial scale is rough 100 u (corresponding to roughly 100 neu-
rons), and the temporal scale is roughly 50 msec (corresponding to firing bursts
of these spatial subunits). In Chapter 14, Goldman-Rakic has presented some
exciting new results on the complex, highly specific connections from prefrontal
and from posterior parietal cortex to many other cortical areas in correlated
columnar fashion.

EDELMAN SELECTION PRINCIPLE FOR LEARNING

Edelman (1978) argues that a network has a basic repertoire of firing patterns
inherently built into it. Through experience or learning, associations between
patterns and stimuli are made, and these patterns are enhanced and "selected"
out (via Hebb-type synaptic changes). Several mathematical models can be used
 Introduction 303

to implement this idea. The selection principle for learning should be contrasted
with models in which a network would learn from "scratch" or have a "teacher."

STATISTICAL FLUCTUATIONS

The reality of statistical fluctuations being present in neuronal spike trains and
in postsynaptic potentials has been clear for many years. The very early, brilliant
work of Katz (1969) and collaborators on the statistical nature of the release of
neurotransmitters from vesicles in the synapses was completely confirmed.
These synaptic fluctuations apparently play a crucial role in the functioning of
neural networks and are not simply a fact of nature that the brain must live with.
Cragg and Temperley (1954) took from physics the ideas of cooperative phe-
nomena among groups of atoms to give magnetism, and the role of temperature
T and interaction energy E in the Boltzmann probability distribution
[proportional to exp(—E/KT)] in determining the state of the material. They
suggested the neuronal analogy of cooperativity in the dynamical interactions of
a network of neurons and that a Boltzmann-like expression determined the time
evolution of the network firing behavior. Although they had no biological ana-
logues for E and T, and they did not develop their ideas mathematically, their
work contained many important concepts.
Some 20 years later, Little (1974), unaware of Cragg and Temperley's work,
developed his mathematical model of neuronal network firing behavior based
on an analogy with the modern statistical mechanics spin systems developments
used in understanding magnetism (and other strongly interacting many-body
systems such as the highly successful Bardeen, Cooper, and Schrieffer theory of
superconductivity). Vasudevan and I (Shaw & Vasudevan, 1974) showed that
the physiological basis for the "temperature" and the Boltzmann-like probability
equations in Little's model followed from the statistical fluctuations in synaptic
transmission found by Katz.
Present physical spin system analogies used to model neural networks, for
example, that of Hopfield (1982), derive from Little (1974). The full implica-
tions of the current, very popular theoretical neuronal models are yet to be deter-
mined and experimentally tested. It will be very interesting to see if, indeed,
statistical fluctuations as found in synaptic operation play an important role in
cortical function and memory on the more "global" level.

BALANCE BETWEEN EXCITATION AND INHIBITION

It has been made clear from many neurophysiological experiments that the bal-
ancing effects of excitation and inhibition play a vital role in determining neu-
ronal firing behavior in the CNS. For example, "lateral" inhibition shapes the
receptive fields of many sensory neurons with excitatory center and inhibitory
surround. Wilson and Cowan (1972) in their theoretical analysis showed inter-
304 REPRESENTATIONS: BEYOND THE SINGLE CELL

esting effects of interacting populations of excitatory and inhibitory neurons.

Key questions that need to be settled concern (1) spatial and strength distribu-
tions of excitory and inhibitory synaptic connections and (2) temporal scales for
excitatory and inhibitory postsynaptic potentials. Both of these points have cru-
cially important global network implications as well as obvious local neuronal
relevance. For example, consider the fact that inhibitory PSPs have considerably
longer time constants than excitatory PSPs. This can effectively translate in
models with discrete time steps to the network firing behavior being dependent
on the firing state of (at least) two previous time steps rather than (in most mod-
els) just the previous network firing state. The possible "big payoff" here is that
second-order difference equations in time easily generate sophisticated, oscillat-
ing firing solutions for the network.
Many more "potential cortical principles" could be added that are relevant
for brain theorists. Others might include the role of the layers in cortex and the
role of the many neurotransmitters. In my recent work I have concentrated on
building a testable mathematical model that incorporates the first five principles
listed above (Shaw, Silverman, & Pearson, 1985) in a specific manner. Clearly
there are many choices and many possible models. The excitement of the future
lies in close, interactive, collaborative efforts between theorists and experimen-
talists to understand the neurobiology of learning and memory. The principles
of cortical function are there to be discovered.

REFERENCES

Beurle, R. L. (1956). Properties of a mass of cells capable of regenerating pulses. Philosophical

Transactions of the Royal Society (London) B,240, 55-94.
Caianiello, E. R. (1961). Outline of a theory of thought processes and thinking machines. Jour-
nal of Theoretical Biology, 2, 204-235.
Cragg, B. G., & Temperley, H. N. V. (1954). The organisation of neurones: A cooperative anal-
ogy. Electroencephalography and Clinical Neurophysiology, 6, 85-92.
Edelman, G. M. (1978). Group selection and phasic reentrant signaling: A theory of higher
brain function. In G. M. Edelman & V. B. Mountcastle (Eds.), The mindful brain
(pp. 51-100). Cambridge, Mass.: MIT Press.
Farley, B. G., & Clark, W. A. (1961). Activity in networks of neuron-like elements. In C. Cherry
(Ed.), Information theory (pp. 242-251). Washington, D.C.: Butterworths.
Goldman-Rakic, P. S. (1984). The frontal lobes: Uncharted provinces of the brain. Trends in
Neuroscienses, 7, 425-429.
Hebb, D. O. (1949). The organization of behavior. New York: Wiley.
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective compu-
tional abilities. Proceedings of the National Academy of Sciences USA, 79, 2554-
2558.
Katz, B. (1969). The release of neural transmitter substances. Springfield, 111.: Charles C.
Thomas.
Lettvin, Y. J., Maturana, H. R., McColluch, W. S., & Pitts, W. H. (1959). What the frog's eye
tells the frog's brain. Proceedings of the IRE, 47, 1940-1951.
Little, W. A. (1974). The existence of persistent states in the brain. Mathematical Biosciences,
19, 101-120.
 Introduction 305

MacKay, D. M. & McCulloch, W. S. (1952). The limiting information capacity of a neuronal

link. Bulletin of Mathematical Biophysics, 14, 127-135.
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous
activity. Bulletin of Mathematical Biophysics, 5, 115-133.
Mountcastle, V. B. (1978). An organizing principle for cerebral function: The unit module and
the distributed system. In G. M. Edelman & V. B. Mountcastle (Eds.), The mindful
brain (pp. 7-50). Cambridge, Mass.: MIT Press.
Rail, W. (1962). Electrophysiology of a dendritic neuron model. Biophysical Journal, 2, 145-
167.
Rochester, N., Holland, J. H., Haibt, L., & Duda, W. L. (1956). Tests on a cell assembly theory
of the action of the brain, using a large digital computer, IRE Transactions, IT-2, 80-
93.
Rosenblatt, F. (1962). A comparison of several perception models. In M. C. Yovitts, G. T.
Jacobi, & G. D. Goldstein (Eds.), Self-organizing systems (pp. 463-484). Washington,
D.C.: Spartan Books.
Shaw, G. L., & Palm, G. (Eds.). (1988). Brain theory reprint volume. Singapore: World Scientific
Press.
Shaw, G. L., Silverman, D. J., & Pearson, J. C. (1985). Model of cortical organization embod-
ying a basis for a theory of information processing and memory recall. Proceedings
of the National Academy of Sciences USA, 82, 2364-2368.
Shaw, G. L., & Vasudevan, R. (1974). Persistent states of neural networks and the random
nature of synaptic transmission. Mathematical Biosciences, 21, 207-218.
Turing, A. M. (1936). On computable numbers with an application to the entscheidungsprob-
lem. Proceedings of the London Mathematical Society, 42, 230-265.
Uttley, A. M. (1962). Properties of plastic networks. Biophysics Journal, 2 (Suppl), 169-188.
von Neumann, J. (1956). Probabilistic logics and the synthesis of reliable organisms from unre-
liable components. In C. E. Shannon & J. McCarthy (Eds.), Automata studies (pp.
43-98). Princeton, N.J.: Princeton University Press.
Wilson, H. R., & Cowan, J. D. (1972) Excitatory and inhibitory interactions in localized pop-
ulations of model neurons. Biophysics Journal, 12, 1-24.
15
Neural Networks: Test Tubes
to Theorems
L E O N N. COOPER
M A R K F. B E A R
F O R D F. E B N E R
C H R I S T O P H E R SCOFIELD

When interest in neural networks revived around 1970, few people believed that
such systems would ever be of any use. Computers worked too well; it was felt
that they could be programmed to perform any desired task. But limitations of
current computers in solving many problems involving difficult to define rules
or complex pattern recognition now are widely recognized; if anything, expec-
tations for neural networks may be too high. The problem is no longer to con-
vince anyone that neural networks might be useful but rather to incorporate
such networks into systems that solve real-world problems economically.
Neural networks are inspired by biological systems where large numbers of
neurons, which individually function rather slowly and imperfectly, collectively
perform tasks that even the largest computers have not been able to match. They
are made of many relatively simple processors connected to one another by vari-
able memory elements whose weights are adjusted by experience. They differ
from the now standard Von Neumann computer in that they characteristically
process information in a manner that is highly parallel rather than serial, and
that they learn (memory element weights and thresholds are adjusted by expe-
rience) so that to a certain extent they can be said to program themselves. They
differ from the usual artificial intelligence systems in that (since neural networks
learn) the solution of real-world problems requires much less of the expensive
and elaborate programming and knowledge engineering required for such arti-
ficial intelligence products as rule-based expert systems.
To properly function, neural networks must learn in order that vast arrays of
memory elements (synapses) are assigned the proper strengths. A basic problem
becomes how this learning occurs: how these synapses adjust their weights so
that the resulting neural network shows the desired properties of memory stor-
age and cognitive behavior.

306
 Neural Networks: Test Tubes to Theorems 307

The problem can be divided into two parts. First, what type of modification
is required so that in the course of actual experience the neural network arrives
at the desired state? The answer to this question can be illuminated by mathe-
matical analysis of the evolution of neural networks using various learning
hypotheses. The second part of this problem is to find experimental justification
for any proposed modification algorithm. A question of extraordinary interest
is: What are the biological mechanisms that underlie the nervous system mod-
ification that results in learning, memory storage, and eventually cognitive
behavior?
One experimental model that appears to be well suited for the purpose of
determining how neural networks modify is the cat visual cortex. The modifi-
cation of visual cortical organization by sensory experience is recognized to be
an important component of early postnatal development (Sherman & Spear,
1982). Although much modifiability disappears after the first few months of life,
some of the underlying mechanisms are likely to be conserved in adulthood to
provide a basis for learning and memory. We have approached the problem of
experience-dependent synaptic modification by determining theoretically what
is required of a mechanism in order to account for the experimental observa-
tions in visual cortex. This process has led to the formulation of hypotheses,
many of which are testable with currently available techniques. In this chapter
we illustrate how the interaction between theory and experiment suggests a pos-
sible molecular mechanism for the experience-dependent modifications of func-
tional circuitry in the mammalian visual cortex.

THE EXPERIMENTAL MODEL

Neurons in the primary visual cortex, area 17, of normal adult cats are sharply
tuned for the orientation of an elongated slit of light and most are activated by
stimulation of either eye (Hubel & Wiesel, 1962). Both of these properties—
orientation selectivity and binocularity—depend on the type of visual environ-
ment experienced during a critical period of early postnatal development. We
believe that the mechanisms underlying the experience-dependent modification
of both receptive field properties are likely to be identical. However, for the sake
of clarity, we concentrate primarily on the modification of binocular connec-
tions in striate cortex.
The majority of binocular neurons in the striate cortex of a normal adult cat
do not respond with equal vigor to stimulation of either eye; instead they typi-
cally display an eye preference. To quantify this impression, Hubel and Wiesel
(1962) originally separated the population of recorded neurons into seven ocular
dominance (OD) categories. The OD distribution in a normal kitten or adult cat
shows a broad peak at group 4, which reflects a high percentage of binocular
neurons in area 17. This physiological assay of ocular dominance has proved to
be an effective measure of the state of functional binocularity in the visual
cortex.
308 REPRESENTATIONS: BEYOND THE SINGLE CELL

Monocular deprivation (MD) during the critical period (extending from

approximately 3 weeks to 3 months of age in the cat [Hubel & Wiesel, 1970])
has profound and reproducible effects on the functional connectivity of striate
cortex. Brief periods of MD will result in a dramatic shift in the OD of cortical
neurons such that most will be responsive exclusively to the open eye. The OD
shift after MD is the best known and most intensively studied type of visual
cortical plasticity.
When the MD is begun early in the critical period, the OD shift can be cor-
related with anatomically demonstrable differences in the geniculocortical axo-
nal arbors of the two eyes (LeVay, Wiesel, & Hubel, 1980; Shatz & Stryker,
1978). However, MD initiated late in the critical period (Presson & Gordon,
1982) or after a period of rearing in the dark (Mower, Caplan, Christen, & Duffy,
1985) will induce clear changes in cortical OD without a corresponding anatom-
ical change in the geniculocortical projection. Long-term recordings from awake
animals also indicate that ocular dominance changes can be detected within a
few hours of monocular experience, which seems too rapid to be explained by
the formation or elimination of axon terminals. Moreover, deprived-eye
responses in visual cortex may be restored within minutes under some condi-
tions (such as during intracortical bicuculline administration (Duffy, Snodgrass,
Burchfiel, & Conway, 1976]), which suggests that synapses deemed functionally
"disconnected" are nonetheless physically present. Therefore, it is reasonable to
assume that changes in function binocularity are explained not only by adjust-
ments of the terminal arbors of geniculocortical axons, but also by changes in
the efficacy of individual cortical synapses.
The consequences of binocular deprivation (BD) on visual cortex stand in
striking contrast to those observed after MD. Whereas 7 days of MD during the
second postnatal month leave few neurons in striate cortex responsive to stim-
ulation of the deprived eye, most cells remain responsive to visual stimulation
through either eye after a comparable period of BD (Wiesel & Hubel 1965,
1965). Thus it is not merely the absence of patterned activity in the deprived
geniculocortical projection that causes the decrease in synaptic efficacy after
MD.
Work over the past several years has led to a theoretical solution to the prob-
lem of visual cortical plasticity (Cooper, 1973). According to Cooper, Liberman,
and Oja (1979), when a visual cortical neuron is depolarized beyond a "modi-
fication threshold," 0M, synaptic efficacies change along lines envisaged by Hebb
(1949). However, when the level of postsynaptic activity falls below 0M> synaptic
strengths decrease. Thus in this model the sign of a synaptic change is a function
primarily of the level of postsynaptic activity. Analysis by Cooper et al. (1979)
confirmed that such a modification scheme could lead to the development of
selectivity that is appropriate for the input environment. However, these
authors also noted that a fixed modification threshold leads to certain technical
problems. For instance, if the postsynaptic response to all patterns of input
activity slipped below 0M (as might occur during binocular deprivation), then the
 Neural Networks: Test Tubes to Theorems 309

efficacy of all synapses would decrease to zero. Bienenstock, Cooper, and Munro
(1982) solved this problem by allowing the modification threshold to float as a
function of the averaged activity of the cell. With this feature, the theory can
successfully account for virtually all the types of modification that have been
observed in kitten striate cortex since 1970. This theory is outlined in more
detail in the next section; then we shall return to the question of possible
mechanisms.

THEORETICAL ANALYSIS

Cortical neurons receive synaptic inputs from many sources. In layer IV of

visual cortex the principal afferents are those from the lateral geniculate nucleus
(LGN) and from other cortical neurons. This leads to a complex network that
has been analyzed in several stages. In the first stage, consider a single neuron
with inputs from both eyes (Fig. 15.1 A). Here d represents the level of presyn-
aptic geniculocortical axon activity, m the synaptic transfer function (synaptic
strength or weight), and c the level of postsynaptic activity of the cortical neu-
ron. The output of this neuron (in the linear region) can be written

which means that the neuron firing rate (or dendritic depolarization) is the sum
of the inputs from the left eye multiplied by the left-eye synaptic weights plus
the inputs from the right eye multiplied by the right-eye synaptic weights. Thus
the signals from the left and right eyes are integrated by the cortical neuron and
determine its level of depolarization (output) at any instant.
The crucial question becomes: How does m change in time according to expe-
rience? According to Bienenstock et al. (1982), m modifies as a function of local,
quasi-local, and golbal variables. Consider the synaptic weight of the jth synapse
on a neuron, mj, (Fig. 15.IB). This synapse is affected by local variables in the
form of information available only through the jth synapse, such as the presyn-
aptic activity levels d and the efficacy of the synapse at a given instant in time
mj (t). Quasi-local variables represent information that is available to the jth
synapse through intracellular communication within the same cell. These
include the instantaneous firing rate (or dendritic depolarization) of the cell c,
the time-averaged firing rate c, and the potentials generated at neighboring syn-
aptic junctions (dm)k, I, m, .. . Finally, global variables represent information
that is available to a large number of cortical neurons, including the neuron
receiving the j'th synapse. These variables might include the presence or absence
of "modulatory" neurotransmitters such as acetylcholine and norepinephrine.
We can delay consideration of the global variables by assuming that they act
to render cortical synapses modifiable or nonmodifiable by experience. In the
310 REPRESENTATIONS: BEYOND THE SINGLE CELL

FIGURE 15.1. Pyramidal shaped cortical neurons and the proximal segments of their apical
dendrites. The shaded circles attached to the dendrites represent dendritic spines. (A) In the first
stage of the theoretical analysis, we consider only the inputs to the cell from the lateral geniculate
nucleus. The signals conveyed along these afferents arise from either the left retina (d1) or the right
retina (dr) and are transferred to the cortical neuron by the synaptic junctions m' and m'. The output
of the cortical neuron, as measured by the firing rate or the dendritic depolarization, is represented
as c, which is the sum of d1 • m1 and d' • m'. (B) The central question is how one of these afferent
synapses, mt, modifies in time as a function of both its level of presynaptic activity ds and the level
of postsynaptic depolarization.

"plastic state," the Bienenstock et al. algorithm for synaptic modification is

written

so that the strength of the jth synaptic junction mj changes its value in time as
a function 0 of the quasi-local states c and c, and as a linear function of the local
variable dj The crucial function, , is seen in Figure 15.2.
One significant feature of this model is the change of sign of at the modifi-
cation threshhold, 0M. When the input activity of the jth synapse d and are
 Neural Networks: Test Tubes to Theorems 311

both concurrently greater than zero, then the sign of the synaptic modification
is positive and the strength of the synapse increases; 0 > 0 when the output of
the cell exceeds the modification threshold (this type of synaptic modification is
"Hebbian"). When dj is positive and is less than zero, then the synaptic effi-
cacy weakens; < 0 when c < 0M. Thus "effective" synapses will be strength-
ened and "ineffective" synapses will be weakened, where synaptic effectiveness
is determined by whether the presynaptic pattern of activity is accompanied by
the simultaneous depolarization of the target dendrite beyond 0M. Since the
depolarization of the target cell beyond 0M normally requires the synchronous

FIGURE 15.2. The function at two values of the modification threshold 0M. According to
Bienenstock et al. (1982), active synapses (d > 0) are strengthened when 0 is positive and are
weakened when 0 is negative. 0 is positive when c, the postsynaptic depolarization, is greater
than 0M. The modification threshold, where 0 changes sign, is a nonlinear function of the average
activity of the postsynaptic neuron (c). Hence in this example 0M (1) would be expected when
cortical neurons have experienced a normal visual environment (A) while 0M(2) would result from
a prolonged period of binocular deprivation (B).
312 REPRESENTATIONS: BEYOND THE SINGLE CELL

activation of converging excitatory synapses, this type of modification will

"associate" those synapses that are concurrently active by increasing their effec-
tiveness together.
Another significant feature of this model is that the value of 0M is not fixed but
instead varies as a nonlinear function of the average output of the cell, c. In a
simple situation

By allowing 0M to vary with the average response in a faster-than-linear fashion,

the response characteristics of a neuron evolve to maximum selectivity starting
at any level within the range of the input environment. It is also this feature that
provides the stability properties of the model so that, for instance, simultaneous
presynaptic and postsynaptic activity at a continued high level does not con-
tinue to increase the synaptic strength.
Now consider the situation of reverse suture where the right, formerly open
eye is closed and the left, formerly sutured eye is reopened. The output of a
cortical neuron in area 17 approaches zero just after the reversal, since its only
source of patterned input is through the eye whose synapses had been function-
ally disconnected as a consequence of the prior MD. However as c diminishes,
so does the value of 0M. Eventually the modification threshold attains a value
below the small output that is evoked by the stimulation of the weak left-eye
synapses. Now the efficacy of these "functionally disconnected" synapses will
begin to increase because even their low response values exceed 0M. As these
synapses strengthen and the average output of the cell increases, 0M again slides
out until it overtakes the new left-eye response values. At the same time, the
efficacy of the right-eye synapses continually decreases because their response
values remain below the modification threshold. In its final stable state, the neu-
ron is responsive only to the newly opened eye, and the maximum output to
stimulation of this eye equals 0M. One particular consequence is that if 0M does
not adjust to the new average firing rate too rapidly, in the course of the rear-
rangement of response, the cell's response to the previously open eye will dimin-
ish before its response to the newly opened eye increases.
Thus far the discussion has been limited to an idealized single neuron whose
inputs arise only from the lateral geniculate nucleus. The second stage of the
theoretical analysis requires that relevant intracortical connections be incorpo-
rated into the model (Cooper & Scofield, 1988). Input from other regions of cor-
tex are considered part of a background excitation or inhibition contributing to
the spontaneous activity of the cell. In addition, the various time delays that
result in structure in the poststimulus time histogram are assumed to be inte-
grated over periods of the order of a second for purposes of synaptic modifica-
tion. This leads to a circuit as shown in Figure 15.3.
The output of the cells of the full network can be written
 Neural Networks: Test Tubes to Theorems 313

FIGURE 15.3. Network with in-

puts from left and right eyes, with
LGN-cortical and corticocortical
synapses.

where c* is a sigmoidal response function,

c, is the output firing rate of the ith cortical cell, and

where M'is and Mris are the 5th LGN "synapses" from the left and right eye to the
ith cortical cell.

are the time-averaged inputs from the left and right eye as described in Bienen-
stock, Cooper, and Munro (1982),

is the matrix of cortico-cortical synapses and Lij is the synapse from the jth cell
to the ith cell. (Notice that italicized boldface symbols always contain left and
right eye components.) In the monotonically increasing region above threshold
and below saturation, in a linear approximation,
314 REPRESENTATIONS: BEYOND THE SINGLE CELL

We consider a region of cortex for which the neural mapping of the input from
the visual field is constant (all of the cells, in effect, look at a given region of the
visual field). Under these conditions, for an input d constant in time, the equi-
librium state of the network would be

MEAN FIELD APPROXIMATION

For a given LGN-cortical vector of synapses mi (the ith row of M) and for a
given input from both eyes d, Eq. 5 for the firing rate of the ith cortical cell
becomes

where the first term is due to the input from LGN and the second due to input
from other cortical cells. We define c as the spatially averaged firing rate of all
of the cortical cells in the region defined above:

The mean field approximation is obtained by replacing cj in the sum in Eq. 5 by

its average value, so that c, becomes

Here, in a manner similar to that in the theory of magnetism, we have replaced

the effect of individual cortical cells by their average effect (as though all other
cortical cells can be replaced by an "effective" cell). It follows that

where

and

so that
 Neural Networks: Test Tubes to Theorems 315

If we assume that the lateral connection strengths are a function of i — j only

(not dependent on the absolute position of a cell in the network, therefore depen-
dent only on the distance of two cells from one another), Lij becomes a circular
matrix so that

and

In the mean field approximation we can therefore write

where the mean field

with

and we assume that L0 < 0 (the network is, on average, inhibitory).

THE CORTICAL NETWORK

The behavior of visual cortical cells in various rearing conditions suggests that
some cells respond more rapidly to environmental changes than others. In mon-
ocular deprivation, for example, some cells remain responsive to the closed eye
in spite of the very large shift of most cells to the open eye. Singer (1977) found,
using intracellular recording, that geniculo-cortical synapses on inhibitory inter-
neurons are more resistant to monocular deprivation than are synapses on
pyramidal cell dendrites. In dark rearing some cells become nonresponsive to
visual stimuli, although most cells retain some responsiveness (Sherman &
Spear, 1982). Recent work suggests that the density of inhibitory GABAergic
synapses in kitten striate cortex is also unaffected by MD during the critical
period (Bear, Schmechel, & Ebner, 1985; Mower, White, & Rustad, 1986).
These results suggest that some LGN cortical synapses modify rapidly,
whereas others modify relatively slowly, with slow modification of some cortico-
cortical synapses. Excitatory LGN cortical synapses onto excitatory cells may be
those that modify primarily. (Since these synapses are formed exclusively on
dendritic spines, this raises the possibility that the mechanisms underlying syn-
aptic modification exist primarily in axospinous synapses.) To embody these
facts we introduce two types of LGN cortical synapses: those (m i ) that modify
316 REPRESENTATIONS: BEYOND THE SINGLE CELL

(according to the modification rule discussed in Bienenstock et al., 1982) and

those (zk) that remain relatively constant. In a simple limit we have

where c denotes the spatial average over cortical cells, and ci denotes the time-
averaged activity of the ith cortical cell. The function 0 appears in Figure 15.2.
We assume for simplicity, and consistent with the foregoing physiological inter-
pretation, that these two types of synapses are confined to two different classes
of cells and that both left and right eye have similar synapses (both mi or both
zk) on a given cell. We therefore can write

Further, in what follows, we assume for maximum simplicity that there is no

modification of cortico-cortical synapses, although experimental results suggest
only that modification of inhibitory cortico-cortical synapses is slow (Bear et al.,
1985; Mower et al., 1986). The consequences of a theory including cortico-cor-
tical synapse modification for the full network was briefly discussed in Bienen-
stock et al (1982) and will be discussed more fully in the mean field approxi-
mation elsewhere.
In a cortical network with modifiable and nonmodifiable LGN cortical syn-
apses and nonmodifiable cortico-cortical synapses, the synaptic evolution equa-
tions become

This leads to a very complex set of coupled nonlinear stochastic evolution equa-
tions that have been simulated and partially analyzed by Cooper and Scofield
(1988). The mean field approximation permits dramatic simplification of these
equations, leading to analytic results and a fairly transparent understanding of
their consequences in various conditions. In this approximation Eqs. 14 become

So that we can now write

and
 Neural Networks: Test Tubes to Theorems 317

where al(r} contains terms from modifiable and nonmodifiable synapses:

and a is defined in Eq. 12a. (N = Nm + Nnm, where Nm is the number of cells

with modifiable synapses and Nnm is the number of cells with nonmodifiable
synapses.) Since it is assumed that neither L nor z changes as the network
evolves, only ml(r) is time dependent.
Having defined a mean field approximation that greatly simplifies the equa-
tions for the response and evolution of cortical cells, we have obtained a fun-
damental result: the stability and position of the fixed points in this network are
related to the fixed points in the absence of mean field (a'r = 0) by

where m* (a) is a fixed point of Eq. 15 in the mean field a and m* (0) is a fixed
point of this equation in zero mean field.
Thus if m* (a) is restricted to the first quadrant (positive values for all of its
components due to the excitatory nature of LGN cortical synapses), as long as
a is large enough and nonspecific (there is sufficient inhibition for all pattern
inputs), m,(a) can still reach all of the fixed points that would have been reached
by m,(0) (not restricted to the first quadrant). This means that if network inhi-
bition is sufficient, the selective stable fixed points can be reached even though
LGN cortical synapses are excitatory. Once reached, the fixed points m*(a) have
the same stability characteristics as the corresponding m*(Q).
We find, consistent with previous theory and with experiment, that most
learning can occur in the LGN cortical synapses; inhibitory (cortico-cortical)
synapses need not modify and some nonmodifiable LGN cortical synapses are
required. It becomes useful to know whether these could be associated with
some anatomical feature. (E.g., might these be synapses into shafts rather than
spines?)
As in the zero mean field theory, zero cell output is an unstable fixed point.
Thus in binocular deprivation the cell output could be on average above or
below spontaneous activity (depending on the level of inhibition). Some "non-
visual" cells would reappear if excitation were enhanced or inhibition
diminished.
In monocular deprivation the closed-eye response goes to

Therefore, LGN cortical synapses do not go to zero. Rather

318 REPRESENTATIONS: BEYOND THE SINGLE CELL

Thus if inhibition is suppressed, one would expect some response from the
closed eye. This is in agreement with experiment.

A POSSIBLE PHYSIOLOGICAL MECHANISM

One of the consequences of the network theory discussed in the previous section
is that the experimental results obtained in visual cortex over the last generation
can be explained by modification of excitatory synapses, with minimal changes
in intracortical inhibition. The balance of available experimental evidence sup-
ports this conclusion. For example, using intracellular recording Singer (1977)
found that geniculocortical synapses on inhibitory interneurons are more resis-
tant to monocular deprivation than are synapses on pyramidal cell dendrites.
And recent work suggests that the density of inhibitory GABAergic synapses in
kitten striate cortex is also unaffected by MD during the critical period (Bear et
al, 1985). Taken together, these theoretical and experimental results indicate
that the search for mechanisms should be focused on the excitatory synapses
that impinge on excitatory cells in visual cortex. Interestingly, this type of syn-
apse is formed exclusively on dendritic spines, a feature that distinguishes it
from other types of cortical synapse (Colonier, 1981). This suggests that expe-
rience-dependent modifications in striate cortex occur primarily at axospinous
synapses.
What mechanisms support the experience-dependent modification of axospi-
nous synapses? Recall that, according to the theory, when the postsynaptic cell
is depolarized beyond the modification threshold 0M, active synapses will be
strengthened. Depolarization beyond 0M minimally requires the synchronous
activation of converging excitatory afferents. When postsynaptic activity fails to
reach 0M, the active synapses will be weakened. The identification of the physi-
ological basis of 0M is therefore central to an understanding of the modification
mechanism.
Work on long-term potentiation (LTP) in the hippocampal slice preparation
has provided an important insight into the nature of the modification threshold
that may be applicable to the visual cortex. LTP is a long-lasting increase in the
synaptic strength of excitatory afferents that have been tetanically stimulated.
The induction of LTP depends on the coactivation of converging excitatory
afferents (input cooperativity [McNaughton, Douglas, & Goddard, 1978]), the
depolarization of the postsynaptic neuron (Kelso, Ganong, & Brown, 1986), the
activation of NMD A receptors (Collingridge, Kehl, & McLennan, 1983; Harris,
Ganong, & Cotman, 1984), and the postsynaptic entry of calcium ions (Lynch,
Larson, Kelso, Barrionuevo, & Schottler, 1983). A current working hypothesis
is that the synchronous tetanic activation of converging afferents depolarizes the
target dendrite beyond the threshold for postsynaptic Ca2+ entry through gates
linked to the NMDA receptor (Wigstrom & Guslafsson, 1985). Elevated den-
dritic calcium then triggers the intracellular changes that lead to enhanced syn-
aptic efficacy (Lynch & Baudry, 1984).
 Neural Networks: Test Tubes to Theorems 319

NMDA receptors are a subclass of excitatory amino acid receptor, and glu-
tamic acid or a closely related substance is thought to be the transmitter of excit-
atory axospinous synapses at many locations in the forebrain (Cotman, Foster,
& Lanthorn, 1981). These receptors are widely distributed in the cerebral cortex,
including the visual areas (Monaghan & Cotman, 1984). It appears that NMDA
receptors normally coexist postsynaptically with quisqualate and/or kainate
receptors (Foster & Fagg, 1985). The "non-NMDA" receptors are thought to
mediate the classical excitatory postsynaptic potential that normally results
from electrical stimulation of axospinous synapses. NMDA receptors, on the
contrary, appear to be linked to a membrane channel that will pass calcium ions.
Dingledine (1983) first reported that NMDA receptor activation leads to cal-
cium flux only when the cell is concurrently depolarized. This is apparently due
to a blockage of the NMDA channel by Mg2+ ions, which is alleviated only when
the membrane is depolarized sufficiently (Nowak, Bregostovski, Ascher, Her-
bert, & Prochiantz, 1984). Thus calcium entry through channels linked to the
NMDA receptor could specially signal presynaptic and postsynaptic
coactivation.
Recently Kleinschmidt, Bear, and Singer (1987) obtained results suggesting
that NMDA receptor-mediated Ca2+ entry also contributes to the synapse mod-
ifications that underlie ocular dominance plasticity in striate cortex. Specifically,
they found that intracortical infusion of 2-amino-5-phosphonovaleric acid
(APV), a selective antagonist of the NMDA receptor (Watkins & Evans, 1981),
prevents the ocular dominance shift that would normally occur after MD. More-
over, this pharmacological treatment also resulted in a striking loss of orienta-
tion selectivity.
Our theoretical analyses lead us to suggest the following:
1. 6M is linked to the membrane potential at which NMDA receptor activation
by sensory fiber activity results in dendritic calcium entry.
2. Increased calcium flux across the dendritic spine membrane results specifi-
cally in an increase in the synaptic gain.
3. Activated synapses accompanied by no postsynaptic calcium signals will be
weakened over time.
4. The membrane potential at which Ca2+ enters through NMDA channels or
the efficacy of the entering Ca2+ in producing synaptic enhancement should
vary depending on the history of prior cell activity.
This physiological model is consistent with the Bienenstock, Cooper, and
Munro (1982) theory. According to the model, the value and sign of 0 is linked
to the calcium ion movement into dendritic spines. Synaptic efficacy will
increase when presynaptic activity evokes a large postsynaptic calcium signal (0
> 0). This will occur only when the membrane potential exceeds the level
required to open the NMDA receptor-activated calcium channels (c > 0M).
When the amplitude of the evoked Ca2+ signal falls below a certain critical level,
corresponding to 0 = 0 and c = 0M, then active synapses will be weakened over
time. Application of an NMDA receptor blocker theoretically would increase the
320 REPRESENTATIONS: BEYOND THE SINGLE CELL

value of 0M, such that it would take a greater level of depolarization to achieve
the critical calcium concentration. In accordance with the experimental obser-
vations of Kleinschmidt et al. (1987), the theoretical consequence would be a
loss of orientation selectivity and a prevention of ocular dominance plasticity.
This model makes some explicit predictions about the regulation of the cal-
cium messenger system that is linked to the NMDA receptors on cortical den-
drites. Recall that 0M depends on the average activity of the cell. If the average
activity decreases, as it does during binocular deprivation, then 0M decreases,
and it should take less dendritic depolarization to maintain synaptic efficacy.
One way this could occur in our model would be to alter the voltage or trans-
mitter sensitivity of the NMDA receptors with the result that less synaptic activ-
ity (depolarization) would be required to evoke the necessary calcium signal. It
is well documented that receptor supersensitivity occurs as a consequence of
postsynaptic inactivity at many locations in the nervous system (Cannon &
Rosenblueth, 1949). Alternatively, a weak calcium signal could be amplified at
points further downstream, for example, by increasing the activity or genetic
expression of calcium-activated enzymes.

CONCLUSIONS

We have presented an algorithm for synaptic modification that reproduces clas-

sical experimental results in visual cortex. These include the relation of cell tun-
ing and response to various visual environments experienced during the critical
period: normal rearing, binocular deprivation, monocular deprivation, and
reverse suture. A molecular model for this form of modification has been pro-
posed based on the NMDA receptors. In this model the Bienenstock et al. (1982)
modification threshold 0M is linked with the voltage-dependent unblocking of
the NMDA receptor channels. A consequence of this relationship is that the
membrane potential at which Ca2+ enters through NMDA channels or the effi-
cacy of the entering Ca2+ in producing synaptic enhancement should vary
depending on the history of prior cell activity.
Stated in this language, many interesting questions are generated: How long
does it take 0M to adjust to a new average firing rate? What is the molecular basis
for this adjustment? How do the putative global modulators of cortical plastic-
ity, such as acetylcholine and norepinephrine, interact with the second-messen-
ger systems linked to NMDA receptors on cortical dendrites? Can we provide
direct evidence that those cells that modify are or are not those acted on by the
modulators? Are the known morphological features of dendritic spines causally
related to the modifiability of synaptic strength? Do the same rules apply to reor-
ganization in adults and in the developmental period?
There has been much discussion in recent years about possible modification
of synapses between neurons as the physiological basis of learning and memory
storage. Molecular models for learning at the single-synapse level have been pre-
sented, various learning algorithms have been proposed that show some indi-
 Neural Networks: Test Tubes to Theorems 321

cation of appropriate behavior, and a mathematical structure for networks of

neurons is rapidly evolving. We have begun a concerted effort to unite these
approaches, and we believe that the close interaction between theory and exper-
iment has greatly enriched both endeavors. Theory has been anchored to exper-
imental observations and experiments have been focused onto those issues most
relevant to sorting out the various possible hypotheses. Further, this interaction
has enabled us to pose new questions with precision and clarity.

Acknowledgment
This work was supported in part by the Office of Naval Research Accelerated
Research Initiative on Learning and Memory.

REFERENCES

Andersen, P. O. (1987). Long term potentiation: Outstanding problems. In J.-P. Changeux &
M. Konishi (Eds.), The neural and molecular bases of learning, (pp. 239-262). New
York: Wiley.
Bear, M. F., Schmechel, D. M., & Ebner F. F. (1985). Glutamic acid decarboxylase in the striate
cortex of normal and monocularly deprived kittens. Journal of Neuroscience, 5,
1262-1275.
Bienenstock, R. L., Cooper, L. N., & Munro, P. W. (1982). Theory for the development of
neuron selectivity: Orientation specificity and binocular interaction in visual cortex.
Journal of Neuroscience, 2, 32-48.
Bliss, T. V. P., & Lomo, T. (1973). Long-lasting potentiation of synaptic transmission in the
dentate area of the anaesthetized rabbit following stimulation of the perforant path.
Journal of Physiology (London), 232, 331.
Cannon, W. B., & Rosenblueth, A. (1949). The supersensitivity of denervated structures. New
York: Macmillan.
Collingridge, G. L., Kehl, S. L., & McLennan, H. (1983). Excitatory amino acids in synaptic
transmission in the Schaffer collateral-commisural pathway of the rat hippocampus.
Journal of Physiology, 334, 33-46.
Colonier, M. (1981). The electron-microscope analysis of the neuronal organization of the cere-
bral cortex. In F. O. Schmitt, F. G. Worden, G. Adelman, S. G. Dennis (Eds.), The
organization of the cerebral cortex (pp. 125-152). Cambridge, Mass.: MIT Press.
Cooper, L. N. (1973). A possible organization of animal memory and learning. In B. Lindquist
& S. Lindquist (Eds.), Proceedings of the Nobel symposium on collective properties of
physical systems (Vol. 24, pp. 252-264). New York: Academic Press.
Cooper, L. N., Liberman, F., & Oja, E. (1979). A theory for the acquisition and loss of neuron
specificity in visual cortex. Biological Cybernetics, 33, 9-28.
Cooper, L. N., & Scofield, C. (1988). Mean field approximation in neural networks, Proceedings
of the National Academy of Sciences USA, 85, 1973-1977.
Cotman, C. W., Foster, A., & Lanthorn, T. (1981). In G. DiChiara & G. Gessa (Eds.), Gluta-
mate as a neurotransmitter (pp. 1-27). New York: Raven.
Dingledine, R. (1983). TV-methylaspartate activates voltage-dependent calcium conductance in
rat hippocampal pyramidal cells. Journal of Physiology, 343, 385-405.
Duffy, F. H., Snodgrass, S. R., Burchfiel, J. L., & Conway, J. L. (1976). Bicuculline reversal of
deprivation amblyopia in the cat. Nature, 260, 256-257.
Foster, A. C., & Fagg, G. E. (1985). Amino acid binding sites in mammalian neuronal mem-
branes: Their characteristics and relationship to synaptic receptors. Brain Research
Review, 7, 103.
322 REPRESENTATIONS: BEYOND THE SINGLE CELL

Harris, E. W., Ganong, A. H., & Cotman, C. W. (1984). Long-term potentiation involves acti-
vation of N-methyl-D-aspartate receptors. Brain Research, 232, 132-137.
Hebb, D. O. (1949). The organization of behavior. New York: Wiley.
Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interactions and functional
architecture in the cat's visual cortex. Journal of Physiology (London), 160, 106-154.
Hubel, D. H., & Wiesel, T. N. (1970). The period of susceptibility to the physiological effects
of unilateral eye closure in kittens. Journal of Physiology, 206, 419-436.
Kelso, S. R., Ganong, A. H., & Brown, T. H. (1986). Hebbian synapses in hippocampus. Pro-
ceedings of National Academy Science, USA, 83, 5326(1986).
Kleinschmidt, A., Bear, M. F., & Singer, W. (1987). Blockade of "NMDA" receptors disrupts
experience-dependent plasticity of kitten striate cortex. Science, 238, 355-358.
LeVay, S., Wiesel, T. N., & Hubel, D. H. (1980). The development of ocular dominance col-
umns in normal and visually deprived monkeys. Journal of Comparative Neurology,
191, 1-52.
L0mo, T. (1966). Frequency potentiation of excitatory synaptic activity in the dentate area of
the hippocampal formation. Ada Physiologica Scandinavica, 68 (Suppl. 277), 128.
Lynch, G., & Baudry, M. (1984). The biochemistry of memory: A new and specific hypothesis.
Science, 224, 1057-1063.
Lynch, G., Larson, J., Kelso, S., Barrionuevo, S., Schottler, F. Intracellular injections of EGTA
block induction of hippocampal long-term potentiation. (1983). Nature (London),
305, 719.
McNaughton, B. L., Doughlas, R. M., & Goddard, G. V. (1978). Synaptic enhancement in fas-
cia dentata: Cooperativity among coactive afferents. Brain Research, 157, 111.
Monaghan, D. T., & Cotman, C. W. (1984). Distribution of N-methyl-D-aspartate-sensitive L-
[3H] glutamate binding sites in rat brain, Journal of Neuroscience, 11, 2909-2919.
Mower, G. D., Caplan, C. J., Christen, W. G., & Duffy, F. H. (1985). Dark rearing prolongs
physiological but not anatomical plasticity of the cat visual cortex. Journal of Com-
parative Neurology, 235, 448-466.
Mower, G. D., White, W. F., & Rustad, R. (1986). [3H] Muscimol binding of GABA receptors
in the visual cortex of normal and monocularly deprived cats. Brain Research, 380,
253-260.
Nicholls, D. G., & Sihra, T. S. (1986). Synaptosomes possess an exocytotic pool of glutamate.
Nature (London), 321, 772-773.
Nowak, L., Bregostovski, P., Ascher, A., Herbert, A., & Prochiantz, A. (1984). Magnesium gates
glutamate-activated channels in mouse central neurones. Nature (London), 307, 462-
465.
Presson, J., & Gordon, B. (1982). The effects of monocular deprivation on the physiology and
anatomy of the kitten's visual system. Society of Neuroscience Abstracts, 8, 5-10.
Shatz, C. J., & Stryker, M. P. (1978). Ocular dominance in layer IV of the cat's visual cortex
and the effects of monocular deprivation. Journal of Physiology, 281, 267-283.
Sherman, S. M., & Spear, P. D. (1982). Organization in visual pathways in normal and visually
deprived cats. Physiological Review, 62, 738-855.
Singer, W. (1977). Effects of monocular deprivation on excitatory and inhibitory pathways in
cat striate cortex. Experimental Brain Research, 134, 508-518.
Watkins, J. C., & Evans, R. H. (1981). Excitatory amino acid transmitters. Annual Review of
Pharmacology and Toxicology, 21, 165-204.
Wiesel, T. N., & Hubel, D. H. (1965). Comparison of the effects of unilateral and bilateral eye
closure on cortical unit responses in kittens. Journal of Neurophysiology, 28, 1029-
1040.
Wigstrom, H., & Gustafsson, B. (1985). Ada Physiologica Scandinavica, 123, 59.
16
Notes on Neural Computing
and Associative Memory
TEUVO K O H O N E N

There is no doubt that the brain consists of discrete operational units (cells or
cell groups) that are selectively activated to perform specific functions. Another
salient feature is the vast number of feedback loops of local as well as global
scale, which gives rise to various kinds of collective effects. A further indisput-
able fact is that the interconnections (the structures and/or strengths) are adap-
tive, a property without which learning would be impossible. Any remaining
discrepancies then concern the conceptualization and characterization of these
properties.
Consider the issue of localization of neural functions. I doubt whether any-
body, at least nowadays, wants to suggest that the mental processes are derivable
from the intrinsic properties of the neural tissue, or that, for instance, the neural
matters of the left and right hemispheres are different, although they respond
differently in mental tasks. The specialization of functions results partly from
different input and partly from the computations made thereby. The coarse
localization of computing functions in sensory, motor, and associative areas
may result from the most probable course of ontogenesis in which these func-
tions are specialized, whereas, for instance, in hydrocephalic patients, other
parts of the cell mass have become responsible for the same functions. This
seems to show that a particular type of localization represents one possible stable
state in the developmental process. After that, however, any neural network
organization must rely on its components, which are consolidated and confined
to some places.
The main difficulties in understanding seem to arise from the fact that in a
collectively operating system, the spatial distribution of the activities of the units
may vary (temporally, even for constant-stimulus conditions), whereas the
material components need not move anywhere from their place. Unless the dis-
tinction between function and activity is made, the nature of the internal rep-
resentations in the brain will remain obscure.
A high-level function that seems to be very difficult to conceptualize and con-
cretize is memory, especially associative memory. Depending on the scientific

323
324 REPRESENTATIONS: BEYOND THE SINGLE CELL

paradigm, one may thereby understand (1) basic (adaptive, distributed) memory
mechanisms, (2) adaptive transformations of signals in a neural network, (3)
detailed representations ("coding") of information that are actually stored, (4)
dynamic processes or temporal recollections that are synthesized in such a feed-
back system referring to a global scale, (5) the semantic content and relations
between the representations stored in memory, or (6) possibly other aspects. In
the scientific tradition, especially in the explanation of basic mechanisms, one
has to concentrate on one aspect at a time; if the phenomena are then studied
in reduced form, simplification is necessary to quantify the processes, although
any claim that the basic models do not explain general mental effects would be
unjustified.
This chapter begins with the categorization of neural models, and simple
structures and system equations for some of them are suggested. This is followed
by an example of a collective effect which shows that neural activity can be local-
ized in a discrete place while neural functions are not; this effect is utilized in
self-organization. After that, one possibility for the formation of internal repre-
sentations is devised. The associative memory connections and associative
recall of signal patterns can thereby be related to these representations. Finally,
a few issues concerning the extension of modeling to other aspects of the neural
functions are considered.

CATEGORIZATION OF NEURAL MODELS

It is obvious that the main purpose of the nervous system is to monitor and
control the living conditions of the organism relative to its environment. In
other words, sensory functions and motor control are the most basic tasks imple-
mented in all species. In higher animals, including humans, some kind of logic
inference and decision making must be involved with sensory and intellectual
tasks, but one important question is whether these operations occur at every
neuron and in the lowest level neural networks. I am inclined to believe that
such inference and decision operations do not result until collective actions in
rather large (hierarchical) assemblies of neurons have been realized at the last
stages of phylogenesis; at least these are not the functions with which the study
of the neural networks should be started.
In an attempt to qualify the basic theoretical models set up to explain the
operations of the nervous system, it may be necessary to consider them describ-
ing functions at different levels of organization and abstraction. First of all, one
may distinguish the following categories of models:
1. Neuronal-level (decoupled) models, where the most careful studies concen-
trate on the dynamic and adaptive properties of each neuron, in order to
describe the neuron as a unit that gives selective responses to patterns and
has some elementary control and memory functions.
2. Network-level models, where identical neurons are interconnected to exhibit
 Notes on Neural Computing and Associative Memory 325

one of the following functions in idealized form: filter function (e.g., projec-
tion operator, optimal estimator, or associative memory) or nonlinear feed-
back, which makes the system activity converge to one of its eigenstates.
Because these models are computationally heavier, they must apply simpli-
fied neuronal-level models for their basic units and concentrate on the
description of the basic collective effects.
3. Nervous system-level (organizational) models, in which usually two or more
networks of type 2 with different properties are combined to demonstrate
more complex or abstract functions of sensory perception, such as automatic
classification, concept formation, motor functions, or global stability control.
The latter is an important aspect, because the overall system must be kept in
a globally stable state all the time, but it must on demand exhibit the control
functions identifiable as attention, priming, and other psychological effects.
The primary objective is to study more qualified systemic phenomena than
those mentioned above.
4. Mental operation-level models, which are the most abstract models, usually
describing operations, procedures, algorithms, and strategies on the so-called
human information processing level. These models are meant to describe the
basic processes of cognition, thinking, problem-solving, and the like, and they
are very closely related to the approaches made in artificial intelligence
research.

STRUCTURES AND FUNCTIONS OF SOME BASIC

NEURAL NETWORKS

Elementary Network Structures

The basic structure in most neural networks is the crossbar switch (Fig. 16.la),
which is formed when a bundle of axons or their branches pierces the dendritic
trees of a set of neurons and makes synaptic connections with them. The next
development of neural functions results from lateral feedback (Fig. 16.1b). This
kind of connectivity has long been known in neuroanatomy. Many interesting
neural functions are derivable from its behavior, as discussed later in more
detail.
In an attempt to devise theoretical operation modules for neural systems, it
was found by the mid-1970s that it is useful to define the basic network structure
containing adaptive crossbars at input as well as for feedback (Fig. 16.2; see also
Kohonen, 1984; Kohonen, Lehtio, & Rovamo, 1974).
The most natural topology of the network is two-dimensional (Fig. 16.2 shows
only a one-dimensional section for simplicity), and the distribution of lateral
feedbacks within this subsystem is the same around every neuron; however, the
distribution of the interconnections could strongly depend on the distance
between two points in the network. Within an artificial module, all unit neurons
could receive the same set of input signals in parallel. Effects similar to those
326 REPRESENTATIONS: BEYOND THE SINGLE CELL

FIGURE 16.1. Elementary neural network structures. (a) Crossbar switch. (b) Complete feedback
matrix; input one-to-one to the neurons.

described below are demonstrable, although the input connections are more ran-
dom and incomplete. None of these simplifying assumptions is unnatural, how-
ever, even when relating to biological structures.

Functional Laws
The system equations for one module of Figure 16.2 may be written in the fol-
lowing general way:

Here x is the vector of all external inputs to the module, y the vector of all output
activities, and M and N are two (adaptive) connectivity matrices, respectively.
In biological systems, Eq. la might be called the relaxation equation, and its
time constants are small, on the order of 10 msec; y describes electrical activities,
and among other things involves the effect of diffusion of light ions. Equations
1b and 1c have much larger time constants, say, weeks, because M and N
describe changes in proteins, other macromolecules, and anatomical structures.
Both of these equations are adaptation equations; Eq. 1c further seems to
describe the basic function of associative memory.
The transfer properties of individual biological neurons should actually be
derived from the so-called membrane-triggering equations. Even in simplified
form, physical description of an active cell membrane may contain about two
dozen process variables and 15 parameters. Based on certain simplifying argu-
 Notes on Neural Computing and Associative Memory 327

merits, I suggested (Kohonen, 1984) the following differential equation for the
electrical activity (average triggering frequency) ni of neuron i':

where £ij is the activity at the input j of cell i, and ¥>,/£„) represents the effect of
input on membrane potential, and thereby on output activity too; ri(n) is a non-
linear loss term which combines all refractory time, leakage, shunting, and sat-
uration effects taking place at the membranes. The steady-state solution derived
from Eq. 2 will then lead to the popular static relation between the signals, called
the sigmoidal function <j( •). If the synaptic control effects were then linearized,
we would have for one type of modeling equation

j=1

where a has two saturation limits and 0, is an offset parameter, often termed
"threshold". The uij- correspond to synaptic efficacies.
For the adaptation equations, a simple but nonetheless theoretically very
influential equation could then read (Kohonen, 1984)

FIGURE 16.2. Operational neural module.

328 REPRESENTATIONS: BEYOND THE SINGLE CELL

where the first term on the right corresponds to the classical law of Hebb (1949)
widely used in modeling, a is a positive constant, and b(ni) a scalar function of
ni,-, with certain rather mild restrictions (e.g., the constant term in its Taylor
expansion must be zero). The second term represents active forgetting, being
nonzero only for nonzero ni. This term seems to be necessary to prevent the
synaptic strengths increasing monotonically with use, which would soon lead to
an inoperative state. The form of this term may be justified in different ways.
Active forgetting, of course, would be a natural physical effect caused by, say,
turnover of the molecules that are responsible for synaptic efficacy; proportion-
ality to cell activity is a natural feature since the chemical reactions may be facil-
itated or strongly enhanced at the triggering of the cell. One may notice that an
essentially similar form for Eq. 4 results if it is assumed that the synapses com-
pete on postsynaptic resources.
Some experimental evidence for the particular form of Eq. 4 also exists. For
instance, Levy and Burger (1987) showed by hippocampal recordings that no
changes in synaptic efficacies occur at zero postsynaptic activity; the efficacies
are increased at high simultaneous presynaptic and postsynaptic activities
and decreased at low presynaptic activity coinciding with high postsynaptic
activity.

DYNAMIC LOCALIZATION OF ACTIVITY IN A DISTRIBUTED

FEEDBACK NETWORK

The response from a neural network to external stimuli, although seemingly

localized at a particular neuron, is actually defined by rather complicated collec-
tive interactions resulting from abundant feedback connections. There are many
kinds of coupling in nervous systems, but it seems that a commonplace type of
local interaction between neighboring neurons is one whereby the closest neigh-
bors (say, up to a distance of 0.5 mm) excite each other, and the neighbors at a
somewhat longer distance (say, 1 to 2 mm) inhibit each other, respectively. In
the first approximation, interactions beyond this range may be ignored. This
type of local interaction is frequently termed the "Mexican hat function." It will
first be shown that such a network tends to concentrate its activity into local
clusters, "bubbles," and the locations of such bubbles depend on the external
signals in a particular way; the signals need not be spatially localized.
Consider again Figure 16.2 in which the ej, are components of vector x, ni; are
components of vector y, and nkl are elements of matrix M. The relaxation equa-
tion is written in the form

where Ii, is the net control effect of all external inputs to cell i, ukl the intercon-
 Notes on Neural Computing and Associative Memory 329

nection strength or kernel between cells k and /, K the number of cells, and -yi
the nonlinear loss term. We take uki = ^(\k — i \ ) , which has the form of the
Mexican hat (exemplified in Fig. 16.3a), and for simplicity let yi = const, n2i,
although the exact form of this function is not important. If the width of the u
function is not much less than the width of the cell array, and if /,, on average,
is a smooth function of i (although a random, even time-variable background
input is allowed), then the activity values ni, start to concentrate into a one-
dimensional bubble, that is, a stabilized, bounded activity cluster, as seen in
Figure 16.3b. Apparently the bubble is formed at a place where the initial activ-
ity, due to stimulation /,, was maximum. (To be more exact, it is the maximum
of the smoothed downward curvature of the ni that defines the position of the
bubble.) Remember that the Ii are also proportional to the inner product of x
with the various input weights mi.

FIGURE 16.3. (a) The "Mexican-hat" interaction kernel, (b) Formation of a bubble of activity
over a one-dimensional network, with 51 units (horizontal coordinate) indexed by i. Development
of activity ni as a function of time, t = 0,1,2, . . . , 10.
330 REPRESENTATIONS: BEYOND THE SINGLE CELL

SELF-ORGANIZED FORMATION OF SENSORY MAPS

The issues about internal representations of abstract knowledge in the brain may
be very difficult to treat in any detailed form. This section describes a rather
primitive case in which internal representations of primary sensory signals are
formed in a basic physical network; even this example, however, is far from
trivial and is very difficult to explain theoretically.
It seems that the two essential effects that lead to spatially organized maps are
(1) spatial concentration or clustering of the network activity into a bubble
whose location depends on the input information and (2) sensitization to the
prevailing input of those network units that lie within the bubble.

Simplification of the Clustering Effect for Computer Simulations

For numerical simulations by computer, without essentially losing any of the
general physical behavior, we shall "quantitize" the clustering process, or for-
mation of the bubble, in a very mild way. At the same time the basic effect may
become more lucid.
Let us recall that in the vector notation x = [£i1,-,, Ei2, ... , £in]T is the input
vector, which, for notational simplicity and computational efficiency, is
assumed to be connected to all the neurons i in this network. The synaptic
weight vector of neuron i was denoted by mt = [u i ,, ui2., ... , nin]T. Thus finding
the best match of x with the m, on the basis of the simplest measure would be
identical to comparing the inner products m x, since they occur as additive
terms in the monotonically increasing sigmoidal functions in Eq. 3. The thresh-
olds 0i must thereby be assumed constant with i. The m x are also proportional
to the Ii terms in Eq. 5, the relative magnitudes of which determined at which
location the bubble was formed.
At this time we shall make use of a result that has been justified elsewhere. It
seems, that due to the "active forgetting" in Eq. 4 the input weight vectors mi
in the adaptive process tend to become normalized to constant lengths, or norms
that are independent of the signal values. Then, instead of using inner products,
the matching of x and the mi can also be based on the comparison of their norms
of the vectorial differences x — mi. If this expression is used in the computerized
simulation process from the beginning to indicate the match, an even more reli-
able self-organizing result usually is obtained.
Let us recall that in reality the neurons formed a sheet. The topology of the
interactions in the network defines which neurons are neighbors. In the follow-
ing simulations the neurons were arranged in a two-dimensional hexagonal grid
and we can easily define six or more immediate neighbors to each unit. One
possible definition of the neighborhood can be seen in Figure 16.4: In the sim-
plest simulations we take it for granted that a bubble is formed corresponding
to such a neighborhood Nc and concentrated around the best matching neuron
c. For good ordering results it has further been found advantageous that the bub-
ble, or Nc be large in the beginning and shrink with time. How this corresponds
to the physical or physiological reality cannot be discussed here, but it has some-
 Notes on Neural Computing and Associative Memory 331

FIGURE 16.4. The mapping array and examples of topological neighborhood (ti < t2 < t3,).

thing to do with the balance between excitatory and inhibitory interconnections

(Kohonen, 1984).

Computational Simplification of the Adaptive Process

Each synaptic weight vector mi, which is a column of the input matrix M in
Figure 16.2, is assumed to change in relation to the signals. For the adaptation
law we use Eq. 4.
Earlier we referred to a result that further contributes to simplification. We
may assume that if the input vector x is changing slowly in time, at least in
relation to the electrical relaxation time constants of activity, n, attain either the
low or the high saturation limit value (inside or outside of the bubble, respec-
tively) whereby /3(ui) is binary, too. With no loss in generality we can rescale the
variables such that Ui e {0, 1} and B(Ui) e {0, «} whereby we obtain, in vector
notation,

The Self-Organizing Algorithm

All the results derived thus far are now collected into a single algorithm that can
effectively be computed on general-purpose digital computers, special array pro-
cessors, or even neural network computers. If the bubble is represented by a
neighborhood set Nc, depicted in Figure 16.4, which is centered around the best
matching neuron with index i = c, and Nc = Nc(t) as well as the factor a - a(f)
are both made time variable, as specified later in more detail, we can convert
Eq. 6 into the discrete-time formalism (t = 0, 1, 2, .. .):
332 REPRESENTATIONS: BEYOND THE SINGLE CELL

Here Nc = Nc(t) is centered at cell c, which is defined by the condition

During computation, Eqs. 8 and 7 alternate. Notice that x is a stochastic variable

with a well-defined statistical density function p(x), from which the successive
input values x(t) are drawn. In real-world observations, such as speech recog-
nition, the x(f) can simply be successive samples of the observation x in their
natural order of occurrence.
The process may be started by choosing completely random initial values for
the mi = mi(0), and the only restriction is that they should be different.
No attempt is made to give the mathematical proof of the ordering of the map
in this chapter. The problem is extremely difficult and its treatment lengthy.
Interested readers can find the proof in the references at the end of this chapter.
Numerical examples of efficient process parameters with the simulation exam-
ples are given here. It may be helpful to emphasize the following general
conditions.
1. Since learning is a stochastic process, the final accuracy of the mapping
depends on the number of steps, which must be reasonably large; there is no
way to circumvent it. A rule of thumb states that for good statistical accuracy
the number of steps needed is at least 500 times the number of network units.
On the other hand, the number of components in x has no effect on the num-
ber of iteration steps, and if neural computers were used, a very high dimen-
sionality of input might be allowed. Typically up to 100,000 steps are used in
simulations, but for "fast learning," as in speech recognition, 10,000 steps
may be enough. Notice that the algorithm is computationally extremely light.
If only a small number of samples is available, they must be recycled for the
wanted total number of steps.
2. During the first 1000 steps or so o = a(t) should start with a value close to
unity, thereafter decreasing gradually but staying above 0.1. The ordering of
the mi occurs during this initial period, while the rest of the steps are needed
only for the fine adjustment of the map. For final accuracy, a = a(f) should
attain small values (e.g., of the order of or less than 0.01) over a long
time.
3. A special precaution relates to the choice of Nc = Nc(t). If the neighborhood,
or the radius of the bubble, were too small in the beginning, the map would
not be ordered globally. Instead one would discern various kinds of parceling
of the map at the borders of which the ordering direction would change dis-
continuously. This phenomenon can be avoided by starting with a fairly wide
Nc - Nc and letting it shrink with time. The initial radius of Nc could even
be more than half the diameter of the network. During the first 1000 steps or
so, when the proper ordering takes place, and a = a(t) is fairly large, the
radius of Nc can shrink linearly to, say, one unit; during the fine-adjustment
phase, Nc should still contain the nearest neighbors of neuron c.
 Notes on Neural Computing and Associative Memory 333

A Numerical Example: Taxonomy (Hierarchical Clustering) of Abstract Data

Although the more practical applications of self-organizing maps can be found
in pattern recognition and modeling of the sensory systems, it may be intriguing
to apply this principle to abstract data vectors consisting of hypothetical attri-
butes or characteristiccs. In the example considered here, there are implicitly
defined structures in the primary data which the map algorithm is able to reveal.
Although this system is a single-level network, it can be shown to be able to
produce on it a hierarchical representation of the relations between the primary
data.
The most central result in self-organization is that if the input signals have a
well-defined probability density function, then the weight vectors of the neurons
are trying to imitate it, however complex its form is. It is even possible to per-
form a kind of numerical taxonomy on this model. As there are no restrictions
to the semantic content of the input signals, they can be regarded as any attri-
butes, discrete or continuous valued. In Table 16.1, a total of 32 items, with 5
hypothetical attributes each, were recorded in a data matrix. (This example is
completely artificial.) Each of the columns represents one item, and for later
inspection it is labeled A through 6, although these labels were not referred to
during learning.
The attribute values (£1; £2>• • • > £5) constitute the pattern vector x, which acts
as a set of signal values at the inputs of the network of the type of Figure 16.2.
This time the vectors x were picked from Table 16.1 by a random choice, and
the state of the network was changed in each representation. Sampling and adap-
tation were continued iteratively until one could regard the asymptotic state as
stationary. Such a "learned" network was then calibrated using the items from
Table 16.1 and labeling the units according to the maximum response caused by
the different items. Figure 16.5 is such a calibrated map. It is discernible that the
"images" of different items are related according to a taxonomic graph where
the different branches are visible. Figure 16.6 illustrates, for comparison, what
kind of minimal spanning tree would describe the similarity relations of the
items of Table 16.1. The following system parameters are engaged in this
process.

a = aft):
During the first 1000 steps, a decreased linearly with time from 0.5 to 0.04 (the
initial value could have been closer to unity, say, 0.9). During the subsequent
10,000 steps, a decreased from 0.04 to 0 linearly with time. [Exponential laws
for a = a(t) may yield a faster convergence; if the problem is not familiar, how-
ever, it is safer and more robust to use the piecewise linear laws.]

Nc = Nc(t):
The lattice was hexagonal, 7 by 10 units, and during the first 1000 steps the
radius of Nc decreased from the value 6 (encompassing the majority of cells in
the network) to 1 (encompassing neuron c and its six neighbors) linearly with
time, thereafter staying at the value 1.
 334 REPRESENTATIONS: BEYOND THE SINGLE CELL

TABLE 16.1 Input Data Matrix

Item
Attribute A B C D E F G H I J K L M N O

£l 1 2 3 4 5 3 3 3 3 3 3 3 3 3 3
h 0 0 0 0 0 1 2 3 4 5 3 3 3 3 3
{3 0 0 0 0 0 0 0 0 0 0 1 2 3 4 5
£> 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
|5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

DISTRIBUTED ASSOCIATIVE MEMORY

A model of autoassociative memory can be derived from the general scheme of

Figure 16.1. For that discussion we need not consider details of the input con-
nections; any activity pattern, either an image or activity distribution referring
to the sensory maps, may be imposed on the cells in one way or another, and
this becomes the information to be stored and recalled.
Assume that the feedback connections of Figure 16.1 are now of the long-
range type and time variable and they are distributed all over the network. Mod-
ification of their strengths will be approximated by Hebb's law, and this time
the forgetting effects are ignored. If the strength of connection between cells /
and j is denoted vy, then

where the ??, are components of the output vector y and of is another paramete
The state of the network—the set of synaptic interconnections represented by
the matrix TV—is thus changed, and these changes represent the memory traces
stored in the network. This memory can be read associatively only; for instance,

FIGURE 16.5. Mapping of the

items of Table 16.1 onto the neural
network.
 Notes on Neural Computing and Associative Memory 335

p Q R S T u V W X Y Z 1 2 3 4 5 6
3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
6 7 8 0 3 3 3 6 6 6 6 6 6 6 6 6 6
0 0 0 1 2 3 4 1 2 3 4 2 2 2 2 2 2
0 0 0 0 0 0 0 0 0 0 0 1 2 3 4 5 6

one may start with some partial activation of the network (key), which then
spreads into the other parts (recollection). Numerous studies have demonstrated
the ability of these networks to recall a previously stored pattern if the initial
activation is a part of it. This possibility was pointed out long ago.
For brevity, I am citing only a result that was obtained long ago (Kohonen,
1984; Kohonen et al., 1977). Figure 16.7a represents one sample of 500 different
photographic images used as the input patterns. Here each dot corresponds to
one cell in the network, and the size of the dot is proportional to its activity
value. Figure 16.7b is the key activation, and Figure 16.7c is the recollection that
was obtained after spreading of activation into the inactive cells. Notice that the
memory traces left by all 500 images were really superimposed on the same
network.

DISCUSSION OF SOME EXTENSIONS OF THE MODELS

After the theoretical approaches made above one might expect that comments
are possible on certain other phenomena known to occur in the neural realms.

Brain Waves
Many researchers have expressed the wish that the models exhibit certain
dynamical phenomena that then could be related to the observed neural signals.

FIGURE 16.6. The minimal

spanning tree of the items of Table
16.1.
336 REPRESENTATIONS: BEYOND THE SINGLE CELL

FIGURE 16.7. Demonstration of associative recall in a distributed network: (a) one of the 500
images stored; (b) key pattern used for excitation; (c) recollection. (Adapted from Kohonen, 1984)

The general mechanisms by which oscillations and waves in distributed media

are generated are, of course, well understood, but the problem lies in our incom-
plete knowledge of the neural tissue as a wave-propagating medium. The phase
velocities may depend on signal propagation in the excitable membranes, chem-
ical diffusion phenomena, and the like, and the signals may "skip" long dis-
tances. It seems unrealistic and unnecessary to describe such phenomena by net-
work models.

Structured Sequences of Signals

Especially in the discussion of memory, one might expect that the production of
temporal associations should be described in more detail. It has been shown in
the classical theories of automata that structured sequences of system states
directly follow from recurrent feedback connections that have a sufficient delay.
Detailed modeling is difficult, due to the complexity of brain networks.

Short-Term and Long-Term Memory

It has often been maintained that short-term memory (STM) should be identi-
fied with reverberating signal patterns and long-term memory (LTM) with
changes in network connectivity. At least the view held of STM ought to be
revised; it seems that due to feedback some activity state may persist after sim-
ulation, but maintenance of signal patterns in neural lines for any significant
period seems impossible. On the other hand, the connections may be altered
temporarily, having shorter time constants than with LTM. The strengths and
ranges of the STM and LTM connectivities may be different, too. Another prob-
lem seems to be more important: How can the memory traces be formed quickly
but stay permanent? Obviously very complicated chains of chemical reactions,
proceeding automatically after initial triggering, may take place at the synapses.
 Notes on Neural Computing and Associative Memory 337

REFERENCES

Hebb, D. (1949). Organization of behavior. New York: Wiley.

Kohonen, T. (1984). Self-organization and associative memory. Heidelberg: Springer-Verlag.
Kohonen, T., Lehtio, P., & Rovamo, J. (1974). Modelling of neural associative memory.
Annales Academiae Scientarium Fennicae (Series A: V Medica, No. 167).
Kohonen, T.. Lehtio, P., Rovamo, J., Hyvarinen, J., Bry, K., & Vainio, L. (1977). A principle
of neural associative memory. Neuroscience (IBRO), 2, 1065-1076.
Levy, W., & Burger, B. (1987, June 21-24). Electrophysiological observations which help
describe an associative synoptic modification rule. Paper presented at the First Inter-
national Conference on Neural Networks, San Diego.
17
Building Network Learning Algorithms
from Hebbian Synapses
T E R R E N C E J. S E J N O W S K I
G E R A L D TESAURO

In 1949 Donald Hebb published The Organization of Behavior, in which he

introduced several hypotheses about the neural substrate of learning and mem-
ory, including the Hebb learning rule, or Hebb synapse. We now have solid
physiological evidence, verified in several laboratories, that long-term potenti-
ation (LTP) in some parts of the mammalian hippocampus follows the Hebb
rule (Brown, Ganong, Kariss, Keenan, & Kelso, 1989; Kelso, Ganong, & Brown,
1986; Levy, Brassel, & Moore, 1983; McNaughton, Douglas, & Goddard, 1978;
McNaughton & Morris, 1987; Wigstrom and Gustafsson, 1985). The Hebb rule
and variations on it have also served as the starting point for the study of infor-
mation storage in simplified "neural network" models (Hopfield & Tank, 1986;
Kohonen, 1984; Rumelhart & McClelland, 1986; Sejnowski, 1981). Many types
of networks have been studied—networks with random connectivity, networks
with layers, networks with feedback between layers, and a wide variety of local
patterns of connectivity. Even the simplest network model has complexities that
are difficult to analyze.
In this chapter we will provide a framework within which the Hebb rule serves
as an important link between the implementation level of analysis, the level at
which experimental work on neural mechanisms takes place, and the algorith-
mic level, on which much of the work on learning in network models is being
pursued.

LEARNING MECHANISMS, LEARNING RULES,

AND LEARNING ALGORITHMS

Long-term potentiation has been found in a variety of preparations. The com-

mon denominator is a long-lasting change in synaptic efficacy or spike coupling
following afferent stimulation with a high-frequency tetanus. There are probably

338
 Building Network Learning Algorithms from Hebbian Synapses 339

several different molecular mechanisms underlying LTP in different prepara-

tions. For example, LTP in the CA1 region of the rat hippocampus can be
blocked with 2-amino-5-phosphonovaleric acid (APV), but the same application
of APV does not block LTP from the mossy fibers in the CA3 region of the
hippocampus (Chattarji, Stanton, & Sejnowski, 1988). How LTP is imple-
mented in these different regions is a question at the level of the learning
mechanism.
The Hebb synapse, in contrast, is a rule rather than a mechanism. That is,
many mechanisms can be used to implement a Hebb synapse, as we will show
in a later section. A learning rule specifies only the general conditions under
which plasticity should occur, such as the temporal and spatial relationships
between the presynaptic and postsynaptic signals, but not the locus of plasticity,
or even the neuronal geometry. As an extreme example, we show how the Hebb
rule can be implemented without synaptic plasticity.
A learning algorithm is more general than a learning rule since an algorithm
must also specify how the learning rule is to be used to perform a task, such as
storing information or wiring up a neural system. Thus a description of the task
to be performed and the type of information involved are essential ingredients
of a learning algorithm. Several examples of learning algorithms will be dis-
cussed in later sections.

IMPLEMENTATIONS OF THE HEBB RULE

The Hebb Rule

Before considering the various possible ways of implementing the Hebb rule,
one should examine what Hebb actually proposed:
When an axon of cell A is near enough to excite cell B or repeatedly or persistently
takes part in firing it, some growth process or metabolic change takes place in one or
both cells such that A's efficiency, as one of the cells firing B, is increased, (p. 62)
This statement can be translated into a precise quantitative expression as fol-
lows. We consider the situation in which neuron A, with average firing rate VA,
projects to neuron B, with average firing rate VB. The synaptic connection from
A to B has a strength value TBA, which determines the degree to which activity
in A is capable of exciting B. (The postsynaptic depolarization of B due to A is
usually taken to be the product of the firing rate VA times the synaptic strength
value TBA.) Now the preceding statement by Hebb states that the strength of the
synapse TBA should be modified in some way that is dependent on both activity
in A and activity in B. The most general expression that captures this notion is

which states that the change in the synaptic strength at any given time is some
as yet unspecified function F of both the presynaptic firing rate and the postsyn-
340 REPRESENTATIONS: BEYOND THE SINGLE CELL

aptic firing rate. Strictly speaking, we should say that F( VA, VB) is a functional,
since the plasticity may depend on the firing rates at previous times as well as
at the current time. Given this general form of the assumed learning rule, it is
then necessary to choose a particular form for the function F( VA, VB). The most
straightforward interpretation of what Hebb said is a simple product:

where e is a numerical constant usually taken to be small. There are many other
choices possible for the function F( VA, FB); the choice depends on the particular
task at hand. Equation 2 might be appropriate for a simple associative memory
task, but for other tasks one would need different forms of the function F( VA,VB)
in Eq. 1. For example, in classical conditioning, as we shall see in the following
section, the precise timing relationships of the presynaptic and postsynaptic sig-
nals are important, and the plasticity must then depend on the rate of change of
firing, or on the "trace" of the firing rate (i.e., a weighted average over previous
times), rather than simply depending on the current instantaneous firing rate.
Once the particular form of the learning algorithm is established, the next step
is to decide how the algorithm is to be implemented. We shall describe here
three possible implementation schemes. This is meant to illustrate the variety
of schemes possible.

Three Implementations
The first implementation scheme, seen in Figure 17.la, is the simplest way to
implement the proposed plasticity rule. The circuit consists solely of neurons A
and B and a conventional axodendritic or axosomatic synapse from A to B. One
postulates that there is some molecular mechanism operating on the postsyn-
aptic side of the synapse which is capable of sensing the rate of firing of both
cells and which changes the strength of synaptic transmission from cell B to cell
A according to the product of the two firing rates. This is in fact similar to the
recently discovered mechanism of associative LTP that has been studied in rat
hippocampus (Brown et al, 1989). (Strictly speaking, the plasticity in LTP
depends not on the postsynaptic firing rate, but instead on the postsynaptic
depolarization. However, in practice these two are usually closely related [Kelso
et al., 1986].) Even here, many different molecular mechanisms are possible. For
example, even though the induction of plasticity occurs at a postsynaptic site,
the long-term structural change may well be presynaptic (Dolphin, Errington, &
Bliss, 1982).
A second possible implementation scheme for the Hebb rule is seen in Figure
17.1b. In this circuit there is now a feedback projection from the postsynaptic
neuron, which forms an axoaxonic synapse on the projection from A to B. The
plasticity mechanism involves presynaptic facilitation: one assumes that the
strength of the synapse from A to B is increased in proportion to the product of
the presynaptic firing rate times the facilitator firing rate (i.e., the postsynaptic
 Building Network Learning Algorithms from Hebbian Synapses 341

FIGURE 17.1. Three implementations of the Hebb rule for

synaptic plasticity. The strength of the coupling between cell
A and cell B is strengthened when they are both active at the
same time, (a) Postsynaptic site for coincidence detection, (b)
Presynaptic site for coincidence detection, (c) Interneuron de-
tects coincidence.

firing rate). This type of mechanism also exists and has been extensively studied
in Aplysia (Carew, Hawkins, & Kandel, 1983; Kandel et al, 1987). Several
authors have pointed out that this circuit is a functionally equivalent way of
implementing the Hebb rule (Gelperin, Hopfield, & Tank, 1985; Hawkins &
Kandel, 1984; Tesauro, 1986).
A third scheme for implementing the Hebb rule—one that does not specifi-
cally require plasticity in individual synapses (Tesauro, 1988)—is seen in Figure
17. 1c. In this scheme the modifiable synapse from A to B is replaced by an inter-
neuron / with a modifiable threshold for initiation of action potentials. The
Hebb rule is satisfied if the threshold of 7 decreases according to the product of
the firing rate in the projection from A times the firing rate in the projection from
B. This is quite similar, although not strictly equivalent, to the literal Hebb rule,
because the effect of changing the interneuron threshold is not identical to the
effect of changing the strength of a direct synaptic connection. A plasticity mech-
anism similar to the one proposed here has been studied in Hermissenda (Alkon,
1987; Farley & Alkon, 1985).
The three methods for implementing the Hebb rule seen in Figure 17.1 are by
no means exhaustive. There is no doubt that nature is more clever than we are
at designing mechanisms for plasticity, especially since we are not aware of most
evolutionary constraints. These three circuits can be considered equivalent cir-
cuits since they effectively perform the same function even though they differ in
the way that they accomplish it. There also are many ways that each circuit
could be instantiated at the cellular and molecular levels. Despite major differ-
ences between them, we can nonetheless say that they all implement the Hebb
rule.
Most synapses in cerebral cortex occur on dendrites where complex spatial
interactions are possible. For example, the activation of a synapse might depo-
larize the dendrite sufficiently to serve as the postsynaptic signal for modifying
an adjacent synapse. Such cooperativity between synapses is a generalization of
the Hebb rule in which a section of dendrite rather than the entire neuron is
considered the functional unit (Finkel & Edelman, 1987). Dendritic compart-
342 REPRESENTATIONS: BEYOND THE SINGLE CELL

merits with voltage-dependent channels have all the properties needed for non-
linear processing units (Shepherd et al., 1985).

USES OF THE HEBB RULE

Conditioning
The Hebb rule can be used to form associations between one stimulus and
another. Such associations can either be static, in which case the resulting neural
circuit functions as an associative memory (Anderson, 1970; Kohonen, 1970;
Longuet-Higgins, 1968; Steinbuch, 1961); or they can be temporal, in which case
the network learns to predict that one stimulus pattern will be followed at a later
time by another. The latter case has been extensively studied in classical con-
ditioning experiments, in which repeated temporally paired presentations of a
conditioned stimulus (CS) followed by an unconditioned stimulus (US) cause
the animal to respond to the CS in a way that is similar to its response to the
US. The animal has learned that the presence of the CS predicts the subsequent
presence of the US. A simple neural circuit model of the classical conditioning
process that uses the Hebb rule is illustrated in Figure 17.2. This circuit contains
three neurons: a sensory neuron for the CS, a sensory neuron for the US, and a
motor neuron, R, which generates the unconditioned response. There is a strong,
unmodifiable synapse from US to R, so that the presence of the US automati-
cally evokes the response. There is also a modifiable synapse from CS to R,
which in the naive untrained animal is initially weak.
One might think that the straightforward application of the literal interpreta-
tion of the Hebb rule, as expressed in Eq. 2, would suffice to generate the desired
conditioning effects in the circuit of Figure 17.2. However, there are a number
of serious problems with this learning algorithm. One of the most serious is the
lack of timing sensitivity in Eq. 2. Learning would occur regardless of the order
in which the neurons came to be activated. However, in conditioning we know
that the temporal order of stimuli is important: if the US follows the CS, then
learning occurs, whereas if the US appears before the CS, then no learning
occurs. Hence Eq. 2 must be modified in some way to include this timing sen-

FIGURE 17.2. Model of classical conditioning using a mod-

ified Hebb synapse. The US elicits a response in the postsyn-
aptic cell (R). Coincidence of the response with the CS leads
to strengthening of the synapse between CS and R.
 Building Network Learning Algorithms from Hebbian Synapses 343

sitivity. Another serious problem is a sort of "runaway instability" that occurs

when the CS-R synapse is strengthened to the point where activity in the CS
neuron by itself causes the R neuron to fire. In that case, Eq. 2 would cause the
synapse to be strengthened upon presentation of the CS alone, without being
followed by the US. However, in real animals we know that presentation of CS
alone causes a learned association to be extinguished, that is, the synaptic
strength should decrease, not increase. The basic problem is that algorithm 2 is
capable of generating positive learning only, and has no way to generate zero or
negative learning.
It is clear then that the literal Hebb rule needs to be modified to produce
desired conditioning phenomena (Klopf, 1987; Tesauro, 1986). One of the most
popular ways to overcome the problems of the literal Hebb rule is by using algo-
rithms such as the following:

Here VA represents the stimulus trace of VA, or the weighted average of VA over
previous times, and VBt represents the time derivative of VB. The stimulus trace
provides the required timing sensitivity so that learning occurs only in forward
conditioning and not in backward conditioning. The use of the time derivative
of the postsynaptic firing rate, rather than the postsynaptic firing rate, is a way
of changing the sign of learning and thus avoiding the runaway instability prob-
lem. With this algorithm, extinction would occur because upon onset of the CS,
no positive learning takes place due to the presynaptic trace, and negative learn-
ing takes place upon offset of the CS. Many other variations and elaborations of
Eq. 3 behave differently, taking into account other conditioning behaviors such
as second-order conditioning and blocking. For details we refer the reader to
Gelperin et al. (1985), Gluck and Thompson (1987), Klopf (1987), Sutton
(1987), Sutton and Barto (1981), and Tesauro (1986). However, all of these other
algorithms are built on the same basic notion of modifying the literal Hebb rule
to incorporate a mechanism of timing sensitivity and a mechanism for changing
the sign of learning.

Associative Memory
Memory and learning are behavioral phenomena, but there are correlates at
many structural levels, from the molecular to the systems levels. Hebb (1949)
went beyond the synaptic and cellular levels to speculate about information pro-
cessing in networks of neurons, or assemblies as he called them, which he con-
sidered the fundamental unit of processing in the cerebral cortex. Information
from the environment is encoded in an assembly by changing the synaptic
strengths at many snapses simultaneously. Even though the Hebb rule is local,
in the sense that only local information is needed to make a decision about how
to change the strength of the synapse, the global behavior of the network of neu-
344 REPRESENTATIONS: BEYOND THE SINGLE CELL

rons may be affected. This raises the possibility that new principles might
emerge when assemblies of neurons are studied.
Probably the most important and most thoroughly explored use of the Hebb
rule is in the formation of associations between one stimulus or pattern of activ-
ity and another. The Hebb rule is appealing for this use because it provides a
way of forming global associations between macroscopic patterns of activity in
assemblies of neurons using only the local information available at individual
synapses.
The earliest models of associative memory were based on network models in
which the output of a model neuron was assumed to be proportional to a linear
sum of its inputs, each weighted by a synaptic strength. Thus

where VB are the firing rates of a group of M output cells, VA are the firing rates
of a group of N input cells, and TBA is the synaptic strength between input cell A
and output cell B. Note that A and B are being used here as arbitrary indices to
represent one out of a group of cells.
The transformation between patterns of activity on the input vectors to pat-
terns of activity on the output vectors is determined by the synaptic weight
matrix, TBA. How should this matrix be chosen if the goal of the network is to
associate a particular output vector with a particular input vector? The earliest
suggestions were all based on the Hebb rule in Eq. 2 (Anderson, 1970; Kohonen,
1970; Longuet-Higgins, 1968; Steinbuch, 1961). It is easy to verify by direct sub-
stitution of Eq. 2 into Eq. 4 that the increment in the output is proportional to
the desired vector and the strength of the learning e can be adjusted to scale the
outputs to the desired values.
More than one association can be stored in the same matrix, so long as the
input vectors are not too similar to one another. This is accomplished by using
Eq. 2 for each input-output pair. This model of associative storage is simple and
has several attractive features: (1) the learning occurs in only one trial; (2) the
information is distributed over many synapses, so that recall is relatively
immune to noise or damage; and (3) input patterns similar to stored inputs will
give output similar to the stored outputs, a form of generalization. This model
also has some strong limitations. First, stored items with input vectors that are
similar (i.e., that have a significant overlap) will produce outputs that are
mixtures of the stored outputs. However, discriminations must often be made
among similar inputs, such as the phonetic distinction between the labial stops
in the words "bet" and "pet." Second, the linear model cannot respond contin-
gently to pairs of inputs (i.e., those that have an output that is different from the
sum of the individual outputs). Some deficiencies can be remedied by making
the learning algorithm and the architecture of the network more complex, as
shown in the next section.
 Building Network Learning Algorithms from Hebbian Synapses 345

The Covariance Rule

Numerous variations have been proposed on the conditions for Hebbian plas-
ticity (Levy, Anderson, & Lehmkuhle, 1984). One problem with any synaptic
modification rule that can only increase the strength of a synapse is that the
synaptic strength will eventually saturate at its maximum value. Nonspecific
decay can reduce the sizes of the weights, but the stored information will also
decay and be lost at the same rate. Another approach is to renormalize the total
synaptic weight of the entire terminal field from a single neuron to a constant
value (von der Malsburg, 1973). Sejnowski (1977a, 1977b) emphasized the need
for a learning rule that decreases the strength of a plastic synapse as specifically
as the Hebb rule increases it and proposed a covariance learning rule. According
to this rule, the change in strength of a plastic synapse should be proportional
to the covariance between the presynaptic firing and postsynaptic firing:

where { V B ) is the average firing rates of the output neurons and { V A ) is the aver-
age firing rates of the input neurons (see also Chauvet, 1986). Thus the strength
of the synapse should increase if the firings of the presynaptic and postsynaptic
elements are positively correlated, decrease if they are negatively correlated, and
remain unchanged if they are uncorrelated. Evidence for long-term depression
has been found in the hippocampus (Levy et al., 1983; Stanton, Jester, Chatterji,
& Sejnowski, 1988) and in visual cortex during development (Reiter & Stryker,
1987; Fregnac, Shulz, Thorpe, & Bienenstock, 1988).
The covariance rule is a special case of the general form of the Hebb rule in
Eq. 1. It does go beyond the simple Hebb rule in Eq. 2; however, it is easy to
show that traditional Hebb synapses can be used to implement Eq. 5, which can
be rewritten as

Both terms on the right-hand side have the same form as the simple Hebb syn-
apse in Eq. 2. Thus the covariance learning algorithm can be realized by apply-
ing the Hebb rule relative to a "threshold" that varies with the product of the
time-averaged presynaptic and postsynaptic activity levels. The effect of the
threshold is to ensure that no change in synaptic strength should occur if the
average correlation between the presynaptic and postsynaptic activities is at
chance level.

Error-Correction Learning
One of the consequences of the linear associative matrix model with Hebbian
synapses is that similar input vectors necessarily produce similar output vectors.
Error-correction procedures can be used to reduce this interference, The weights
346 REPRESENTATIONS: BEYOND THE SINGLE CELL

are changed to minimize the difference between the actual and correct output
vectors:

where V^ is the actual output produced by the network by the current set of
weights and V is the correct output vector supplied by the teacher. Unlike the
previous learning algorithms, which learn in one shot, error-correction proce-
dures such as this are incremental and require several presentations of the same
set of input vectors. It can be shown that the weights will evolve to minimize
the average mean square error over the set of input vectors (Kohonen, 1984).
As with the covariance learning algorithm, the error-correction learning algo-
rithm in Eq. 7 can be rewritten in a form that can be implemented with Hebb
synapses:

Both terms on the right-hand side have the same form as the simple Hebb syn-
apse in Eq. 2. Thus the error-correction procedure can be realized by applying
the Hebb rule twice, first to the actual output produced by the network and then
to the correct output supplied by a teacher, but with a negative rather than a
positive increment. Alternatively, the difference can be computed by another
neuron and used to control the Hebbian learning.
Further improvements have also been made to associative matrix models by
introducing feedback connections, so that the networks are autoassociative, and
by making the processing units nonlinear (Anderson & Mozer, 1981; Hopfield,
1984; Kohonen, 1984; Sejnowski, 1981; Toulouse, Dehaene, & Changeux, 1986).
Associative memory models like this have been proposed for the CA3 region of
the hippocampus (Lynch & Baudry, 1988; McNaughton & Morris, 1987; Rolls,
1987).
The learning algorithms for associative memory introduced in this section are
supervised in the sense that information about the desired output vectors must
be supplied along with the input vectors. One way to provide information about
the desired output to a group of neurons is to have a separate "teaching" input
that "clamps" the output firing rates to the desired values while the input cor-
responding to the desired output is simultaneously active. The climbing fibers
in the cerebellar cortex, which make strong excitatory synapses on individual
Purkinje cells, could have such a teaching role, as first suggested by Brindley
(1964) and later developed by Albus (1971) and Marr (1969) in models of the
cerebellar cortex as an adaptive filter. Evidence for plasticity in the cerebellar
cortex has been found by Ito and his co-workers (Ito, 1982). However, evidence
for Hebbian plasticity in the cerebellum does not necessarily imply that its func-
tion is associative storage, for there are many other possible functions. Evidence
 Building Network Learning Algorithms from Hebbian Synapses 347

for plasticity in the deep cerebellar nuclei has been found as well (Miles & Lis-
berger, 1981; Thompson, 1986).

Learning Internal Representations

The class of network models of associative memory discussed in the last section
has a severe computational limitation in that all the processing units in the net-
work are constrained by either the inputs or the outputs, so that there are no
free units that could be used to form new internal features. What features should
be used for the input units and output units if the network is deeply buried in
the brain? New learning algorithms have been devised for multilayer networks
with nonlinear processing units that overcome some of the limitations of single-
layer networks (Hinton & Sejnowski, 1983; Rumelhart & McClelland, 1986). In
particular, these algorithms use interneurons, or "hidden units," which become
sensitive to the features that are appropriate for solving a specified problem and
for performing context-sensitive computation. We will review one of these
learning algorithms, based on the Boltzmann machine architecture, which can
be implemented by Hebb synapses.
The Boltzmann machine is a network of stochastic processing units that solves
optimization problems (Hinton & Sejnowski, 1983, 1986). The processing units
in a Boltzmann machine have outputs that are binary valued and are updated
probabilistically from summed synaptic inputs, which are graded. As a conse-
quence, the state of the units in a Boltzmann machine fluctuate even for a con-
stant input. The amount of fluctuation is controlled by a parameter that is anal-
ogous to the temperature of a thermodynamic system. Fluctuations allow the
system to escape from local traps into which it would get stuck if there were no
noise in the system. All the units in a Boltzmann machine are symmetrically
connected: this allows an "energy" to be defined for the network and ensures
that the network will relax to an equilibrium state which minimizes the energy
(Hopfield, 1982).
The Boltzmann machine has been applied to a number of constraint satisfac-
tion problems in vision, such as figure-ground separation in image analysis
(Kienker, Sejnowski, Hinton, & Schumacher, 1986; Sejnowski & Hinton, 1987),
and generalizations have been applied to image restoration (Geman & Geman,
1984), binocular depth perception (Divko & Schulten, 1986), and optical flow
(Hutchinson, Koch, Luo, & Mead, 1988). These are problems in which many
small pieces of evidence must be combined to arrive at the best overall inter-
pretation of sensory inputs (Ballard, Hinton, & Sejnowski, 1983; Hopfield &
Tank, 1986).
Boltzmann machines have an interesting learning algorithm that allows
"energy landscapes" to be formed within the hidden units between the input and
output layers. Learning in a Boltzmann machine has two phases. In the training
phase a binary input pattern is imposed on the input group as well as the correct
binary output pattern. The system is allowed to relax to equilibrium at a fixed
348 REPRESENTATIONS: BEYOND THE SINGLE CELL

temperature while the inputs and outputs are held fixed. In equilibrium, the
average fraction of time a pair of units is on together, the co-occurrence proba-
bility, is computed for each connection:

where SB is the output value of the -Bth unit, which can take on the values 0 or
1 only. In the test phase the same procedure is followed with only the input units
clamped, and the average co-occurrence probabilities are again computed:

The weights are then updated according to

where the parameter t controls the rate of learning. A co-occurrence probability

is related to the correlation between the firing or activation of the presynaptic
and postsynaptic units and can therefore be implemented by a Hebb synapse.
In the second phase, however, the change in the synaptic strengths is anti-Heb-
bian since it must decrease with increasing correlation. Notice that this proce-
dure is also error-correcting, since no change will be made to the weight if the
two probabilities are the same.
The Boltzmann machine demonstrates that the Hebb learning rule can be
used to mold the response properties of interneurons within a network and
adapt them for the efficient solution of difficult computational problems. The
Boltzmann machine also shows that noise can play an effective role in improv-
ing performance of a network, and that the presence of noise in the nervous
system does not necessarily imply a lack of precision. Other stochastic network
models have also been studied (Barto, 1985). Although the Boltzmann machine
is not meant to be a realistic brain model, it does serve as an existence proof
that difficult computational problems can be solved with relatively simple pro-
cessing units and biologically plausible learning mechanisms.

Development
The Hebb synapse has also been used by Linsker (1986) to model the formation
of receptive fields in the early stages of visual processing. The model is a layered
network having limited connectivity between layers, and it uses the covariance
generalization of the Hebb rule given in Eq. 5. As the learning proceeds, the units
in the lower layers of the network develop on-center and off-center receptive
fields that resemble the receptive fields of ganglion cells in the retina, and elon-
gated receptive fields develop in the upper layers of the network that resemble
simple receptive fields found in visual cortex. This model demonstrates that
some of the properties of sensory neurons could arise spontaneously during
 Building Network Learning Algorithms from Hebhian Synapses 349

development by specifying the general pattern of connectivity and a few param-

eters to control the synaptic plasticity. One surprising aspect of the model is that
regular receptive fields develop even though only spontaneous activity is present
at the sensory receptors. Similar models that require patterned inputs have also
been proposed (Barrow, 1987; Bienenstock, Cooper, & Munro, 1982; von der
Malsburg, 1973).
The visual response properties of neurons in the visual cortex of cats and
monkeys are plastic during the first few months of postnatal life and can be per-
manently modified by visual experience (Hubel & Wiesel, 1962; Sherman &
Spear, 1982). Normally, most cortical neurons respond to visual stimuli from
either eye. Following visual deprivation of one eye by eyelid suture during the
critical period, the ocular preference of neurons in primary visual cortex shifts
toward the nondeprived eye. In another type of experiment, a misalignment of
the two eyes during the critical period produces neurons that respond to only
one eye, and, as a consequence, binocular depth perception is impaired. These
and many other experiments have led to testable hypotheses for the mechanisms
underlying synaptic plasticity during the critical period (Bear, Cooper, & Ebner,
1987).
Stent (1973) suggested that the effects of monocular deprivation could be
explained if the synaptic weight were to decrease when the synapse is inactive
and the postsynaptic cell is active. An alternative mechanism that incorporates
another Hebbian form of plasticity was proposed by Bienenstock et al. (1982).
The Bienenstock-Cooper-Munro algorithm for synaptic modification is a spe-
cial case of the general Hebb rule in Eq. 1:

where the function (VK, (VB)) is shown in Figure 17.3. The synapse is strength-
ened when the average postsynaptic activity exceeds a threshold and is weak-
ened when the activity falls below the threshold level. Furthermore, the thresh-
old varies according to the average postsynaptic activity:

Bienenstock et al. (1982) showed that this choice has desirable stability proper-
ties and allows neurons to become selectively sensitive to common features in
input patterns.
Singer (1987) suggested that the voltage-dependent entry of calcium into
spines and the dendrites of postsynaptic cells may trigger the molecular changes
required for synaptic modification. This hypothesis is being tested at a molecular
level using combined pharmacological and physiological techniques. N-methyl-
o-aspartic acid (NMDA) receptor antagonists infused into visual cortex block
the shift in ocular dominance normally associated with monocular deprivation
(Kleinschmidt, Bear, & Singer, 1986). The NMDA receptor is a candidate mech-
anism for triggering synaptic modification because it allows calcium to enter a
350 REPRESENTATIONS: BEYOND THE SINGLE CELL

FIGURE 17.3. (left) Change in coupling strength &TBA as a function of the correlation <VBVA >
between the presynaptic and postsynaptic activity levels, as indicated in Eq. 5. The threshold 8 is
given by < VB >< VA >. (right) The postsynaptic factor <o>(VB, <VB >) in the Bienenstock-
Cooper-Munro learning algorithm in Eq. 1, where the threshold 6 is given by 0 = < V B > 2 .

cell only if the neurotransmitter binds to the receptor and the postsynaptic mem-
brane is depolarized. In a sense, the NMDA receptor is a "Hebb molecule," since
it is activated only when there is a conjunction of presynaptic and postsynaptic
activity. The NMDA receptor also is involved in LTP in the hippocampus (Col-
lingridge, Kehl, & McLennan, 1983).
The mechanisms for plasticity in the cerebral cortex during development may
be related to the mechanisms that are responsible for synaptic plasticity in the
adult. The evidence so far favors the general form of Hebbian plasticity in Eq.
1. However, the details of how this plasticity is regulated at short and long time
scales may be quite different during development and in the adult.

CONCLUSIONS

The algorithmic level is a fruitful one for pursuing network models at the present
time for two reasons. First, working top-down from functional considerations is
difficult, since our intuitions about the functional level in the brain may be
wrong or misleading. Knowing more about the computational capabilities of
simple neural networks may help us gain a better intuition. Second, working
from the bottom up can be treacherous, since we may not yet know the relevant
signals in the nervous system that support information processing. The study of
learning in model networks can help guide the search for neural mechanisms
underlying learning and memory. Thus network models at the algorithmic level
are a unifying framework within which to explore neural information
processing.
Three principles have emerged from our studies of learning in neural net-
works. The first is the principle of locality. The Hebb algorithm depends only
 Building Network Learning Algorithms from Hebbian Synapses 351

on information that is present or can be extracted from highly localized regions

of space and time. This is of practical importance for any physical system since
nonlocal algorithms have a high overhead for the communications needed to
bring together the relevant information. In spite of the limitation of locality,
networks can nonetheless achieve a global organization during both develop-
ment and long-term information storage. The second principle is gradient
descent. A global energy or cost function can usually be found whenever large-
scale network organization emerges from local interactions. That is, global orga-
nization is the result of local changes which optimize a function of the entire
system (Sejnowski, 1987). This raises the important possibility that such global
functions may be exploited in the nervous system and may be discoverable
(Sejnowski, 1987). The third is the principle of differences. Given the limited
accuracy of signals in neurons, mechanisms that depend on accurate, absolute
values are not feasible. Information storage can be made more compact when
differences between signals—effectively error signals—are used to make changes
at synapses.
Hebb's learning rule has led to a fruitful line of experimental research and a
rich set of network models. The Hebb synapse is a building block for many dif-
ferent neural network algorithms. As experiments refine the parameters for Heb-
bian plasticity in particular brain areas, it should become possible to begin refin-
ing network models for those areas. There is still a formidable gap between the
complexity of real brain circuits and the simplicity of the current generation of
network models. As models and experiments evolve the common bonds linking
them are likely to be postulates like the Hebb synapse that serve as algorithmic
building blocks.

Acknowledgment
This chapter was prepared with the support of the Mathers Foundation.

REFERENCES

Albus, J. S. (1971). A theory of cerebellar function. Mathematical Biosciences, 10, 25-61.

Alkon, D. L. (1987). Memory traces in the brain. Oxford: Oxford University Press.
Anderson, J. A. (1970). Two models for memory organization using interacting traces. Math-
ematical Biosciences, 8, 137-160.
Anderson, J. A., & Mozer, M. C. (1981). Categorization and selective neurons. In G. E. Hinton
& J. A. Anderson (Eds.), Parallel models of associative memory. Hillsdale, N.J.:
Erlbaum.
Ballard, D. H., Hinton, G. E., & Sejnowski, T. J. (1983). Parallel visual computation. Nature,
306, 21-26.
Barrow, H. G. (1987). Learning receptive fields. In M. Caudill & C. Butler (Eds.), Proceedings
of the First International Conference on Neural Networks (Vol. 4, pp. 115-121). San
Diego: SOS Press.
Barto, A. G. (1985). Learning by statistical cooperation of self-interested neuron-like computing
elements. Human Neurobiology, 4, 229-256.
Bear, M., Cooper, L. N., & Ebner, F. F. (1987). A physiological basis for a theory of synapse
modification. Science, 237, 42-48.
352 REPRESENTATIONS: BEYOND THE SINGLE CELL

Bienenstock, E. L., Cooper, L. N., & Munro, P. W. (1982). Theory for the development of
neuron selectivity: Orientation specificity and binocular interaction in visual cortex.
Journal of Neuroscience, 2, 32-48.
Brindley, G. S. (1964). The use made by the cerebellum of the information that it receives from
sense organs. International Brain Research Organization Bulletin, 3, 80.
Brown, T. H., Ganong, A. H., Kariss, E. W., Keenan, C. L., & Kelso, S. R. (1989). Long-term
potentiation in two synaptic systems of the hippocampal brain slice. In J. H. Byrne
& W. O. Berry (Eds.), Neural models of plasticity. Orlando, Fla.: Academic Press.
Carew, T. J., Hawkins, R. D.. & Kandel, E. R. (1983). Differential classical conditioning of a
defensive withdrawal reflex in Aplysia californica. Science, 219, 397-400.
Chattarji, S., Stanton, P., & Sejnowski, T. J. (1988). Commissural, but not mossy fiber, synapses
exhibit both associative long-term potentiation (LTP) and depression (LTD) in the
CAS region of the hippocampus. Society for Neuroscience Abstracts, 14, 567.
Chauvet, G. (1986). Habituation rules for a theory of the cerebellar cortex. Biological Cyber-
netics, 55, 201-209.
Churchland, P. S., & Sejnowski, T. J. (1988). Neural representations and neural computations.
In L. Nadel (Ed.), Neural connections and mental computation (pp. 15-48). Cam-
bridge, Mass.: MIT Press.
Collingridge, G. L., Kehl, S. L., & McLennan, H. (1983). Journal of Physiology (London), 334,
33.
DiPrisco, G. V. (1984). Hebb synaptic plasticity. Progress in Neurobiology, 89, 98-102.
Divko, R., & Schulten, K. (1986). Stochastic spin models for pattern recognition. In J. S.
Denker (Ed.), AIP Conference Proceedings 151: Neural Networks for Computing (pp.
129-134). New York: AIP.
Dolphin, A. C., Errington, M. L., & Bliss, T. V. P. (1982). Nature, 297, 496.
Farley, J., & Alkon, D. L. (1985). Cellular mechanisms of learning, memory and information
storage. Annual Review of Psychology, 36, 419-494.
Finkel, L. H., & Edelman, G. M. (1987). Population rules for synapses in networks. In G. M.
Edelman, W. E. Gall, & W. M. Cowan (Eds.), Synaptic function, (pp. 711-757). New
York: Wiley.
Fregnac, Y., & Imbert, M. (1984). Development of neuronal selectivity in the primary visual
cortex of the cat. Physiology Review, 64, 325.
Fregnac, Y., Shulz, D., Thorpe, S., & Bienenstock, E. (1988). A cellular analogue of visual cor-
tical plasticity. Nature, 333, 367-370.
Gelperin, A. (1986). Complex associative learning in small neural networks. Trends in Neuro-
sciences, 9, 323-328.
Gelperin, A., Hopfield, J. J., & Tank, D. W. (1985). The logic of Limax learning. In A. Selver-
ston (Ed.), Model neural networks and behavior (pp. 237-261). New York: Plenum.
Geman, S., & Geman, D. (1984). Stochastic relaxation, Gibbs distribution and the Bayesian
restoration of images. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence, 3, 79-92.
Gluck, M. A., & Thompson, R. F. (1987). Modeling the neural substrates of associative learning
and memory: A computational approach. Psychological Review, 94, 176-191.
Hawkins, R. D., & Kandel, E, R. (1984). Is there a cell-biological alphabet for simple forms of
learning! Psychological Review, 91, 375-391.
Hebb, D. O. (1949). Organization of behavior. New York: Wiley.
Hinton, G. E., & Sejnowski, T. J. (1983). Optimal perceptual inference. In Proceedings of the
IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(pp. 448-453). Silver Spring, Md.: IEEE Computer Society Press.
Hinton, G. E., & Sejnowski, T. J. (1986). Learning and relearning in Boltzmann machines. In
D. Rumelhart & J. McClelland (Eds.), Parallel distributed processing: Explorations
in the microstructure of cognition: Psychological and biological models (pp. 282-317).
Cambridge, Mass.: MIT Press.
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective compu-
 Building Network Learning Algorithms from Hebbian Synapses 353

tational abilities. Proceedings of the National Academy of Sciences USA, 79, 2554-
2558.
Hopfield, J. J. (1984). Neurons with graded response have collective computation abilities. Pro-
ceedings of the National Academy of Sciences USA, 81, 3088-3092.
Hopfield, J. J., & Tank, D. W. (1986). Computing with neural circuits: A model. Science, 233,
625-633.
Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interactions, and functional
architecture in the cat's visual cortex. Journal of Physiology, 160, 106-154.
Hutchinson, J., Koch, C., Luo, J., & Mead, C. (1988). IEEE Computer, 21 the flocculus hypoth-
to, M. (1982). Cerebellar control of the vestibulo-ocular reflex-around the flocculus hypoth-
esis. Annual Review of Neuroscience, 5, 275-296.
Ito, M. (1984). The cerebellum and neural control. New York: Raven.
Kandel, E. R., Klein, M., Hochner, B., Shuster, M., Siegelbaum, S. A., Hawkins, R. D., Glanz-
man, D. L., & Castellucci, V. F. (1987). Synaptic modulation and learning: New
insights into synaptic transmission from the study of behavior. In G. M. Edelman,
W. E. Gall, & W. M. Cowan (Eds.), Synaptic function (pp. 471-518). New York:
Wiley.
Kelso, S. R., & Brown, T. H. (1986). Differential conditioning of associative synaptic enhance-
ment in hippocampal brain slices. Science, 232, 85-87.
Kelso, S. R., Ganong, A. H., & Brown, T. H. (1986). Hebbian synapses in hippocampus. Pro-
ceedings of the National Academy of Sciences USA, 83, 5326-5330.
Kienker, P. K,, Sejnowski, T. J., Hinton, G. E., & Schumacher, L. E. (1986). Separating figure
from ground with a parallel network. Perception, 15, 197-216.
Kleinschmidt, A., Bear, M. F., & Singer, W. (1986). Neuroscience Letters, 26 (Suppl.), S58.
Klopf, A. H. (1987). A neuronal model of classical conditioning. (Technical Report AFWAL-
TR-87-1139). Dayton, Ohio: Wright-Patterson Air Force Base Aeronautical
Laboratories.
Kohonen, T. (1970). Correlation matrix memories. IEEE Transactions on Computers, C-21,
353-359.
Kohonen, T. (1984). Self-organization and associative memory. New York: Springer-Verlag.
Komatsu, Y., Fujii, K., Maeda, J., Sakaguchi, H., &Toyama, K. (1988). Long-term potentiation
of synaptic transmission in kitten visual cortex. Journal of Neurophysiology, 59, 124-
141.
Levy, W. B., Anderson, J. A., & Lehmkuhle, W. (1984). Synaptic change in the nervous system.
Hillsdale, N.J.: Erlbaum.
Levy, W. B., Brassel, S. E., & Moore, S. D. (1983). Partial quantification of the associative
synaptic learning rule of the dentate gyrus. Neuroscience, 8, 799-808.
Linsker, R. (1986). From basic network principles to neural architecture: Emergence of orien-
tation columns. Proceedings of the National Academy of Sciences USA, S3, 8779-
8783.
Longuet-Higgins, H. C. (1968). Holographic model of temporal recall. Nature, 217, 104-107.
Lynch, G. & Baudry, M. (1988). Structure-function relationships in the organization of mem-
ory. In M. S. Gazzaniga (Ed.), Perspectives in memory research and training (pp. 23-
92). Cambridge, Mass.: MIT Press.
Marr, D. (1969). A theory of cerebellar cortex. Journal of Physiology, 202, 437-470.
Marr, D. (1982). Vision. San Francisco: Freeman.
McNaughton, B. L., Douglas, R. M., & Goddard, G. V. (1978). Synaptic enhancement in fascia
dentata: Cooperativity among coactive afferents. Brain Research, 157, 277.
McNaughton, B. L., & Morris, R. G. (1987). Hippocampal synaptic enhancement and infor-
mation storage within a distributed memory system. Trends in Neurosciences, W,
408-415.
Miles, F. A., & Lisberger, S. G. (1981). Plasticity in the vestibulo-ocular reflex: A new hypoth-
esis. Annual Review of Neuroscience, 4, 273-299.
Moore, J. W., Desmond, J. E., Berthier, N. E., Blazis, D. E. J., Sutton, R. S., & Barto, A. G.
354 REPRESENTATIONS: BEYOND THE SINGLE CELL

(1986). Simulation of the classically conditioned nicitating membrane response by a

neuron-like adaptive element: Response topography, neuronal firing, and interstim-
ulus intervals. Behavioral Brain Research, 21, 143-154.
Mpitsos, G. J., & Cohna, C. S. (1986). Discriminative behavior and Pavlovian conditioning in
the mollusc. Journal of Neurobiology, 17, 469-486.
Reiter, H. O., & Stryker, M. P. (1987). A novel expression of plasticity in kitten visual cortex
in the absence of postsynaptic activity. Society for Neuroscience Abstracts, 13, 1241.
Rolls, E. T. (1987). Information representation, processing and storage in the brain: Analysis
at the single neuron level. In J.-P. Changeux & M. Konishi (Eds.), Neural and molec-
ular mechanisms of learning (pp. 503-540). Berlin: Springer-Verlag.
Rumelhart, D. E., & McClelland, J. L. (1986). Parallel distributed processing: Explorations in
the microstructure of cognition: Vol. 1. Foundations. Cambridge, Mass.: MIT Press.
Sahley, C., Rudy, J. W., & Gelperin, A. (1981). An analysis of associative learning in a terres-
trial mollusc: Higher-order conditioning, blocking, and a transient US pre-exposure
effect. Journal of Comparative Physiology, 144, 1-8.
Sahley, C. L., Rudy, J. W., & Gelperin, A. (1984). Associative learning in a mollusc: A com-
parative analysis. In D. Alkon & J. Farley (Eds.), Primary neural substrates of learn-
ing and behavioral change (pp. 243-258). New York: Cambridge University Press.
Sejnowski, T. J. (1977a). Statistical constraints on synaptic plasticity. Journal of Mathematical
Biology, 69, 385-389.
Sejnowski, T. J. (1977b). Storing covariance with nonlinearly interacting neurons. Journal of
Mathematical Biology, 4, 303-321.
Sejnowski, T. J. (1981). Skeleton niters in the brain. In G. E. Hinton & J. A. Anderson (Eds.),
Parallel models of associative memory (pp. 189-212). Hillsdale, N.J.: Erlbaum.
Sejnowski, T. J. (1987). Computational models and the development of topographic projec-
tions. Trends in Neurosciences, 10, 304-305.
Sejnowski, T. J., & Hinton, G. E. (1987). Separating figure from ground with a Boltzmann
machine. In M. A. Arbib & A. R. Hanson (Eds.), Vision, brain and cooperative com-
putation (pp. 703-724). Cambridge, Mass.: MIT Press.
Sejnowski, T. J., & Tesauro, G. (1989). The Hebb rule for synaptic plasticity: Algorithms and
implementations. In J. N. Byrne & W. O. Berry (Eds.), Neural model of plasticity (pp.
94-103). Orlando, Fla.: Academic Press.
Shepherd, G. M., Brayton, R. K., Miller, J. P., Segev, I., Rinzel, J., & Rail, W. (1985). Signal
enhancement in distal cortical dendrites by means of interactions between active
dendritic spines. Proceedings of the National Academy of Sciences USA, 82, 2192-
2195.
Sherman, S. M., & Spear, P. D. (1982). Organization of visual pathways in normal and deprived
cats. Physiological Reviews, 62, 738.
Singer, W. (1987). Activity-dependent self-organization of synaptic connections as a substrate
of learning. In J. P. Changeux & M. Konishi (Eds.), The neural and molecular bases
of 'learning (pp. 301-336). New York: Wiley.
Stanton, P., Jester, J., Chattarji, S., & Sejnowski, T. J. (1988). Associative long-term depression
(LTD) or potentiation (LTP) is produced in the hippocampus dependent upon the
phase of rhythmically active inputs. Society for Neuroscience Abstracts, 14, 19.
Steinbuch, K. (1961). Die Lernmatrix. Kybernetik, I, 36-45.
Stent, G. W. (1973). A physiological mechanism for Hebb's postulate of learning. Proceedings
of the National Academy of Sciences USA, 70, 997-1001.
Sutton, R. S. (1987). A temporal-difference model of classical conditioning. (Technical Report
TR87-509.2). Waltham, Mass.: GTE Labs.
Sutton, R. S., & Barto, A. G. (1981). Toward a modern theory of adaptive networks: Expecta-
tion and prediction. Psychological Review, 88, 135-170.
Tesauro, G. (1986). Simple neural models of classical conditioning. Biological Cybernetics, 55,
187-200.
 Building Network Learning Algorithms from Hebbian Synapses 355

Tesauro, G. (1988). A plausible neural circuit for classical conditioning without synaptic plas-
ticity. Proceedings of the National Academy of Sciences USA, 85, 2830-2833.
Thompson, R. F. (1986). The neurobiology of learning and memory. Science, 233, 941-947.
Toulouse, G., Dehaene, S., & Changeux, J. P. (1986). Spin glass model of learning by selection.
Proceedings of the National Academy of Sciences USA, 83, 1695-1698.
von der Malsburg, C. (1973). Self-organization of orientation sensitive cells in striate cortex.
Kybernetik, 14, 85.
Wigstrom, H., & Gustafsson, B. (1985). On long-lasting potentiation in the hippocampus: A
proposed mechanism for its dependence on coincident pre- and postsynaptic activity.
Acta Physiologica Scandinavica, 123, 519.
Wilshaw, D. (1981). Holography, associative memory, and inductive generalization. In G. E.
Hinton & J. A. Anderson (Eds.), Parallel models of associative memory (pp. 83-104).
Hillsdale, N.J.: Erlbaum.
Woody, C. D. (1982). Memory, learning and higher function. Berlin: Springer-Verlag.
18
A Neural Architecture
for the Representation of Scenes
C H R I S T O P H VON DER M A L S B U R G

Learning and memory deal with the process by which knowledge and skills are
absorbed and stored in the brain. To understand them it is important to have
precise ideas about the stuff that is being stored: What is its nature? How is it
organized? How is it implemented in neural hardware? The neuroscience of
learning and memory is at present conditioned by a set of concepts based on
specific answers to these questions. In this chapter I will raise a number of issues
that are not solved in this framework. These issues have come up during
attempts at model reconstructions of elementary capabilities of the brain, such
as pattern recognition, perceptual segmentation, and scene representation. Back-
engineering is an important window to the brain. It is true that if a given func-
tional problem had a million technical solutions, the construction of any one of
them would say nothing about nature's solution. However, there is good reason
to believe that relevant functional problems have very few solutions, and any
one of them will suggest important constraints to the neuroscience of brain func-
tion. Backengineering is a fruitful exercise because it necessitates whole-system
solutions. This puts it in contrast to many studies that limit their view to small
subtasks and subsystems, thus leaving out essential aspects. This chapter contin-
ues a series of conceptual discussions (von der Malsburg, 1981, 1987b; von der
Malsburg & Bienenstock, 1986). More technical accounts of the approach advo-
cated here are contained in Bienenstock and von der Malsburg (1987), von der
Malsburg (1985, 1988), von der Malsburg and Schneider (1986), von der Mals-
burg and Bienenstock (1987) and Werman (1987).

THE ISSUE OF ARCHITECTURE

In computer science one often speaks of the architecture of a system. The term
refers to a conceptual level of structure which abstracts from details of imple-
mentation and from consideration of particular applications. It has been said

356
 A Neural Architecture for the Representation of Scenes 357

that the brain is just an incoherent collection of simple tricks, each one to be
understood on its own. If that was true, nothing could be discussed on a general
abstract level and there would be no functional architecture to the brain.
This view is conceivable, but is it likely to be realistic? How could such
unprincipled structure have come about in the brain? The obvious answer is that
evolution created it through simple trial and error. If there really were no general
design principles, this would be a very inefficient process. A generous estimate
limits genetic information that can have accumulated over the course of evolu-
tion to a few billion bits (von der Malsburg & Singer, 1988). This is in line with
the amount of information in our genes. Since about 33 bits are required to des-
ignate one out of the 10'° cells in our cerebral cortex, the specification of its 1014
synapses needs slightly more than 33 X 1014 bits (not speaking about the rest of
the brain and nervous system). El,volution therefore must have made use of reg-
ularities and strategies, otherwise it could not have developed the brain (von der
Malsburg, 1987a). These strategies and regularities certainly constitute an
architecture.
The point just made can be illustrated by the analogy to computer programs.
Each little algorithm in a software system could be a realization of an indepen-
dent little trick. There is no inherent necessity to conform to any abstract style
or architecture. But that is not what you find in actual programs. A very impor-
tant aspect of the recent history of computer science is the development of high-
level computer languages, programming styles, verification procedures, and data
structures. This was important to speed up "evolution." Developing or debug-
ging an architecturally unstructured sequence of machine instructions would
simply be prohibitively slow. The same reason must have forced evolution to
develop an architecture.
More specifically, the view has been expressed that the brain had no represen-
tation of its environment; that is, there was no coherent system by which a one-
to-one correspondence could (in principle) be established between stiuations in
the environment and states of the brain. In terms of an analogy, a car is designed
to deal with road situations but it would be very difficult or impossible to find
any kind of "representation" of road situations in the structure or dynamical
state of the car. It is much better understood in terms of a collection of devices
and mechanisms to solve individual problems one at a time. Analogously, the
brain could be just a collection of reflexes and mechanisms designed to deal with
individual problems and situations.
The issue of correspondence between mind and reality is a very important and
a very old subject of philosophy. We can, of course, have no knowledge of reality
in any direct sense (Ding an sich). When we speak of "reality" we refer to a
mental construction that is objective in the sense that it is not subject to indi-
vidual arbitrary choices and that we can communicate and agree about it. It
would therefore be absurd to claim we had no mental representation of reality.
In a dualist point of view this statement about mind still leaves open the issue
of existence of a representation in the brain. To avoid a lengthy philosophical
discussion let me leave this line of argument to the reader.
358 REPRESENTATIONS: BEYOND THE SINGLE CELL

There is a further argument in favor of the existence of a representation of our

environment in the brain, again based on development, this time on the time
scale of individual learning and memory. When we are thrown into a new and
unfamiliar environment we repeatedly run into situations to which we cannot
react in appropriate ways. After a while, however, we acquire experience and
learn to deal flexibly with every new situation. This is possible only on the basis
of a perceptive system able to make all distinctions necessary for appropriate
decisions, although many of these decisions had never been met before. Our
level of success in dealing with new situations in familiar environments indi-
cates that the raster of distinctions that can be made by our perceptive system
is complete in a certain sense. This raster of distinctions is a perception of the
environment and may certainly be called a representation of it.
Proceeding on the assumption that there is indeed an abstract and coherent
architecture to the brain, let me briefly discuss the two aspects of neural archi-
tecture, network topology and dynamics. Network dynamics deals with the
nature of signals and the mechanisms by which they are processed in neurons.
It also deals with the way in which processing mechanisms are plastically
changed and organized. Network topology speaks about the connectivity dia-
gram of brain networks, their subdivision into areas, the fiber projections
between areas, and the internal structure of connections within them. The term
brain architecture is often seen as synonymous with network topology. On the
other hand, neural dynamics is far from being clear in its details and deserves
more attention. My concern here is mainly with neural dynamics, although I am
aware of the importance of more intensive work on network topology.

THE PROBLEM OF SCENE REPRESENTATION

Associative memory (Anderson, 1970; Cooper, 1974; Hopfield, 1982; Kohonen

1977; Palm, 1980; Steinbuch, 1961; Willshaw, Buneman, & Longuet-Higgins,
1969) crystallizes a set of answers to many issues of brain architecture. It sug-
gests a data structure that can be decomposed into elementary symbols or prop-
ositions. Each element corresponds to a neuron or to a small set of neurons.
Dynamics takes the form of an exchange of excitation and inhibition such that
activity stabilizes in certain stationary states. New stationary states can be stored
as memory traces with the help of Hebbian plasticity. Associative memory sug-
gests a network topology in which the brain is cut into subsystems, each of which
has complete connectivity (or something like a sparse random subset of the com-
plete network), with sparse and regular connections between subsystems.
Memory traces can be dynamically reconstructed from partial information.
Thus incomplete sensory information regarding a familiar object can be com-
pleted by the addition of missing elements and by the addition of abstract
descriptors and names; or a sensory stimulus can be completed by the dynamic
reconstruction of a response which was once learned to be associated to the stim-
ulus. With these properties associative memory is a candidate concept for brain
architecture.
 A Neural Architecture for the Representation of Scenes 359

Is this the complete answer to the question of brain function? One big
unsolved problem with associative memory is appropriate generalization from
one situation in which something has been learned to other, different, situations
in which the knowledge is applied. The main problem here is that situations
never recur identically in natural environments, and that generalization is an
absolute necessity for an animal to survive. The only type of generalization
inherent in associative memory is over Hamming distance, that is, between
states which have a large cell-wise overlap with each other.
The issue of generalization cannot be discussed without specifying the nature
of the general task of the nervous system. This is to control an organism in a
temporally changing scene in order to achieve goals and avoid disaster. The cen-
tral entity for the nervous system is thus a generalized scene. Its internal repre-
sentation, its creation from perceptual data and stored knowledge, as well as its
manipulation and transformation are the central issues of brain architecture.
The generalized scene comprises the scene in the more narrow sense which sur-
rounds the organism: descriptions of objects, their parts, their relationships, and
their state of movement, including the organism itself. This will mostly be built
up by visual image interpretation helped by stored knowledge. Beyond that, the
scene describes objects in terms of interpretation, role, significance, material,
and the like; it represents action plans, including those of other agents within
the scene. The generalized scene is not limited to static situations, dealing also
with temporal processes. Finally, motor patterns concerning manipulation, nav-
igation, and communication are to be created as part of the generalized
scene.
The term scene implies static structure. This needs comment. A scene has to
be organized by a temporal process in the nervous system. This process cannot
be regulated by specific procedures for specific scenes. Flexibility calls for a very
general mechanism of organization. Associative memory, and indeed all known
systems displaying self-organization, are based on mechanisms of relaxation
toward stationary states. If the generalized scene is to comprise temporal pro-
cesses such as motor patterns, then there seems to be a problem: How can such
patterns be created as stationary states? The solution very likely lies in the pos-
sibility of representing temporal processes in a nontemporal way. Any graphical
representation with time as abscissa is an example of a nontemporal represen-
tation of a temporal process. A neural implementation of such representation is
represented by the holophone (Longuet-Higgins 1967). After a motor pattern has
been created in a stationary representation it can be "read off" by a standard
mechanism which transforms it into the temporal domain.

THE GENERALIZATION PROBLEM

The usual basis for generalization is abstraction, by which a situation is freed

from its particular context and is thus made available for many contexts. This
can be done by an incomplete description of a situation in a way that reduces it
to a useful principle that can be frequently applied. An abstract situation or
360 REPRESENTATIONS: BEYOND THE SINGLE CELL

scheme is applied by attaching it to a new scene. Thus generalization involves

three difficult steps:
1. Useful abstractions must be discovered. Not any part of one scene is appli-
cable to others. One important step to abstraction is the decomposition of a
scene into subsituations that are coherent in some sense.
2. When a useful abstraction or principle has been formed its applicability to a
new scene must be discovered. This step necessitates a concept of structural
relatedness, possibly based on partial overlap or on some kind of isomorphy.
3. There has to be a way to attach an abstract descriptor to a new situation and
thereby modify its processing.
Opportunities to generalize are based on the existence of regularities and sym-
metries. In a biological context an important source of regularity is the genetical
reproduction of structures. Another important practical example is generaliza-
tion on the basis of perceptual invariance. A single object in the external world
can create a large variety of stimulus patterns on sensory surfaces, depending on
incidental variables. For instance, the retinal image created by light reflected by
an object depends in its position, size, shape, color, and intensity distribution
on eye position, object distance and orientation, and on illumination. It is
important for an organism to have an internal representation of an object which
is independent of incidental variables and on the basis of which it can generalize
from one retinal image to others of the same object.
The ability to generalize is an important criterion in the discussion of neural
architectures. The three steps necessary for generalization are realized in asso-
ciative memory at best in a very rudimentary form. The reason for its limita-
tions is its inadequate structure concept, which is based on overlap, or Euclidean
or Hamming distance, between two states. It has been shown (Cooper, 1974;
Kohonen, 1977) that associative memory can be made to abstract a prototype
from a set of patterns which have large overlap with each other. Thus a noise-
free prototype can be abstracted from a sequence of noisy exemplars. Applica-
bility of a memory trace to an input pattern is discovered on the basis of suffi-
cient overlap between the two. Application of this "abstract descriptor" is done
by replacing the input pattern by the stored one. This mechanism of generaliza-
tion has serious weaknesses. The first is that situations are treated in a mono-
lithic fashion. Segmentation of a situation into subsituations is not possible in a
flexible way (it has to be imposed by a rigid connection pattern between subsys-
tems). Thus the recurrence of a subpattern in different situations cannot be put
to profit. Second, a subpattern b can be applied to an input pattern a only if b
has been previously associated and attached to a by synaptic plasticity. The
appearance of a then leads to the coactivation of b. There is no concept of struc-
tural similarity that could lead to the discovery within the system of a relation-
ship between a and b such that they can be associated spontaneously. Finally,
the principle by which descriptors are attached to other subpatterns, simple
coactivation, is very poorly structured. If, for instance, a descriptor symbol b is
to be attached to symbol a and a descriptor b' to symbol a' at the same time, an
 A Neural Architecture for the Representation of Scenes 361

ambiguity arises if all four symbols are simply coactivated because an attach-
ment b a' and b' a would lead to the same neural state. This is the syntax
problem which is the main subject of this chapter.
The poor generalization properties of associative memory are considerably
improved in layered systems (Ackley, Hinton, & Sejnowksi, 1985; Fukushima,
1980; Marko & Giebel, 1970; Rosenblatt, 1961; more references are found in
Rumelhart & McClelland, 1986). The central concept here is feature detectors
(or association units or hidden units). These are enabled to detect subpatterns
in an input stimulus by appropriate connections restricted to those neurons on
which the subpatterns are defined. A feature dector generalizes over all input
patterns which share the required subpattern. Layered structures have much
more power to generalize than associative memory and are especially not lim-
ited to small Hamming distance between input patterns. The basis of this higher
power to generalize is a restriction of connectivity. Whereas in associative mem-
ory a neuron receives connections from all other neurons with which it is coac-
tive in some memory state (although for reasons of economy only a sample
thereof is realized), connections in layered structures are restricted to those
between layers (which is, of course, the essence of layering) and, further, to small
subsets of all connections between layers.
Steps 2 and 3 of our generalization scheme are easily accomplished in layered
structures: The applicability of abstractions is discovered with the activation of
association units. Activation of an association unit in turn activates the correct
output units to which it has been connected. The first step, however—discovery
of opportunities to abstract and generalize and implementation of appropriate
connections with input and output layers of the sysem—is difficult. The diffi-
culty has to do precisely with the important restriction of interlayer connections.
Let us consider an individual excitatory connection between two neurons a and
j8 as an elementary rule of the kind "If a is on, then /3 also should be on." The
problem of finding the right connections is equivalent to finding the right ele-
mentary rules. For associative memory this is simple. During storage of a state
the activity of all neurons is controlled from outside, and if neurons a and /3 are
simultaneously active in a state to be memorized, then « and /3 are to be con-
nected with each other. In a layered structure in which activity in input and
output layers is controlled from outside an appropriate subset of hidden units is
to be activated and then appropriate connections are to be introduced or
strengthened. The previous idea of strengthening all connections between coac-
tive neurons in different layers introduces too many connections. Instead, all
those connections that fit only a particular situation and do not generalize
should be left out.
A solution to this problem which is implicit in many of the neural systems
that have been proposed (see, for instance, Ackley et al., 1985; Rosenblatt, 1961;
more references in Rumelhart & McClelland, 1986) is based on statistical prop-
erties of large sets of situations. A starting configuration of connections is given.
This may be random or it may already implement some general property of the
problem at hand, such as locality of connections. Then a sequence of situations
362 REPRESENTATIONS: BEYOND THE SINGLE CELL

is presented by the activation of input neurons. In each situation hidden units

and output units settle into activity states on the basis of current connections.
Connections are changed in small increments according to some rule which has
the important property of stabilizing the system in a state of satisfactory reaction
and of modifying it in states of unsatisfactory reaction (where the satisfactory/
unsatisfactory decision is taken by some external teacher who knows correct
responses). The final connectivity pattern is shaped by statistical properties of
the set of applied stiuations. Those subpatterns on input and output layers are
finally represented by hidden units which have consistently appeared in many
successful situations. If, for instance, mirror symmetry about a given axis is to
be detected in an input plane, then pairs of symmetrically lying pixels come to
be represented (Sejnowski, Kienker, & Hinton, 1986).
A problem with this method is that it may necessitate extreme numbers of
training situations (dozens of thousands of input patterns were necessary to train
the system just cited). This will be particularly worrisome in numerically large
systems and especially with problems in which generalization necessitates detec-
tion of large subpatterns. In contrast, there are many types of phenomena for
which the structure of a problem and potent generalizations can be deduced
from very few exemplars. This is possible only if a system can detect significant
subpatterns in individual situations without having recourse to the statistics of
many situations. In the example in which the symmetry of pixel patterns is to
be classified this is possible by the detection of structural relationships between
symmetrically lying parts of a single pattern.
The detection and processing of structural relationships and of significant pat-
terns before they are represented by association units, such that they can help in
the establishment of such units, is not possible in the neural architecture implicit
in associative memory and in the layered structures discussed. A modification
of neural architecture that permits the processing of this type of syntactical infor-
mation is introduced below.

SYMBOL SYSTEMS WITH SYNTAX

According to the Concise Oxford Dictionary, syntax is "the grammatical

arrangement of words in speech or writing to show their connection and rela-
tion; set of rules governing this arrangement." Syntax is an important and inte-
gral aspect of all known symbol systems. This will be illustrated shortly. Neural
networks, as discussed thus far, are a notable example.
Any symbol system consists of elementary symbols, which can be combined
into more complex composite symbols. The elmentary symbols of written lan-
guage are letters and other characters. These are combined into words, phrases,
sentences, paragraphs, chapters, and books. Since text ultimately refers to the
temporal process of spoken language, the syntactical arrangement of subsymbols
into higher symbols is a linear chain. The whole information content of a long
textual communication is contained in the syntactical arrangement of charac-
ters. In mathematics complex formulas are composed by combining elementary
 A Neural Architecture for the Representation of Scenes 363

characters and other symbols. It is possible to replace a whole formula like

(2af^e~a2x2dx)2 by another elementary symbol, say , but it is clear that new
abbreviations cannot replace syntactically structured composite symbols alto-
gether. The processing of syntactical structures is the whole subject of algebra!
The introduction of new abbreviations has to be preceded by the recognition of
significant constellations of symbols. For this reason alone processing of unab-
breviated syntactical arrays of symbols must be possible.
Compare this to a neural network. A complex symbol in a network is an activ-
ity pattern. This can be decomposed into elementary symbols, active neurons.
An unstructured list of the neurons that are active in a state is all the informa-
tion there is. (With analog signals a vector of numbers is required.) The arrange-
ment of entries in the list is immaterial since there are no dynamical variables
that could code different arrangements. The structure of connections between
neurons is important for the establishment of activity states but it cannot be read
off signals. Consequently, neural networks of the kind discussed so far have only
one composition rule by which a single complex symbol can be formed at a time:
by the simultaneous activation of a set of neurons.
Let me illustrate this point by the contrast to another symbol system, data
structures in computer science. An important general data structure consists of
hierarchical lists. Take the three lists, A = (a b c d), B = [(a b) (c d)], and C
= [ ( a d ) (c b)}. Although they comprise the same elementary symbols, these
lists are discriminable objects and the distinction between them is important.
Only A is natural to the neural architecture discussed so far, the distinction
between A, B, and C being impossible. This point was raised a long time ago.
Rosenblatt (1961, p. 478) discussed the following situation. A layered structure,
a "perceptron," has been trained to respond (with four output units) to squares
and triangles in all positions on a retina. Unit rfi responds to a triangle, rD
responds to a square, rtop responds to a stimulus in the top half of the retina, and
rbot responds to a stimulus in the bottom half. What if there is a triangle in the
top half and square in the bottom half of retina? All four units go on, corre-
sponding to the list (rA r,op rn rbol). This state does not discriminate the situation
from another one in which the positions of square and triangle are interchanged.
A list [(rA rtop) (ra rbot)] would describe the situation correctly, the other situation
being described by [(rA rbm) (ru r,op)]. Rosenblatt proposed solving the problem
with the help of selective attention. He added back-coupling to his system, which
introduced an instability such that the system either activated only units (/A rtop)
or only units (ra rbM), depending on chance or on influences from other
modalities.
A very similar problem was experimentally investigated by Treisman and
Gelade (1980). A visual situation is presented tachistoscopically in which there
are colored letters together with distractor digits. After a short presentation time
(typically 120 msec) the display is masked by visual noise. The surprising result
is that subjects make many conjunction errors, mistaking, for instance, a red O
and a green X for a red X and/or a green O. This happens in every two or three
trials. No such errors occur, of course, if the pattern is displayed sufficiently long.
The interpretation seems straightforward. There are four types of neurons,
364 REPRESENTATIONS: BEYOND THE SINGLE CELL

responding to red (rred), green (rgrn), letter X (rx), and letter O (r0). After a short
presentation active neurons are described by the list A = (rx rred r0 rgrn). Syntac-
tical information is missing. The system knows that there has to be syntactical
binding between letters and colors and proceeds by guessing one of the possibil-
ities, say B = [(rx rgm) (r0 rred)]. Theoretically, the problem could be mended by
the introduction of new types of neurons, encoding conjunctions (rx rred), (r0 rgrn),
and so on. The experiment shows that either these units do not exist or they can
be activated only indirectly with great delay. Treisman and Gelade propose
attention to solve the problem. Given enough time, the visual system concen-
trates attention sequentially on one letter after another, thereby coactivating all
feature units that correspond to one letter. The rest of the brain then can read
off syntactical relations from the temporal conjunctions thus created.
Another important type of experiment in which syntactical relationships play
a central role is described by Julesz (1984). This work deals with the perception
of fields formed out of small figural elements, called textons. Segregation of fig-
ure from ground is found to be effortless and rapid if figure and ground differ in
their mix of texton types. Segregation is tedious and slow if figure and ground
differ only in local syntactical bindings of the same texton elements. One may
again conclude that syntax is processed in the temporal domain, with the help
of a sequence of local fixations of attention.

NEURAL REPRESENTATION OF SYNTAX

Is there a coherent and general way in which syntax can be represented in neural
architecture? The problem has always been camouflaged by the fact that any
rigid predetermined set of syntactical relations can be represented by neurons
again, as in our previous example by a neuron rXrea encoding the binding (rx rred)
of neurons rx and rrei. The point made here, however, is that for learning it is
important to process syntactical information which is not yet represented by
neurons, in order to detect syntactical combinations that are worth being rep-
resented by new neurons. Even in a static, nonlearning system it may be advan-
tageous to have the possibility to represent and process syntax which is not rep-
resented by neurons, either because the required number of neurons would be
prohibitive or, more importantly, because those neurons would be too special
to be a basis for generalization. The psychophysical work cited (Julesz, 1984;
Treisman & Gelade, 1980} is an illustration of this.
To pose the syntax problem clearly let us speak of a neural system that has a
fixed connectivity structure and in which, consequently, each neuron can be
interpreted as a symbol with fixed meaning. Let us refer to all of them as ele-
mentary symbols, although their majority actually codes for combinations of
others. We are now interested in an architecture that is able to represent and
process syntactical structure on these elementary symbols in a flexible way.
The mathematical structure of syntax is not clear. Let there be N elementary
symbols. A simple version of syntax assigns a single variable to each pair of
elementary symbols. There would then be N2 such variables, and the complete
structure would be a graph with elementary symbols as nodes and syntax vari-
 A Neural Architecture for the Representation of Scenes 365

ables as links. A graph assigns a syntax variable to each pair of nodes. More
complex systems would assign variables to larger sets of nodes, say tripletts or
quadrupletts. A system assigning variables to sets of not more than n neurons is
said to be "of order n."
A conceptually simple way of representing syntax is the introduction of a two-
class system of neurons. Class 1 neurons act as elementary symbols, class 2 neu-
rons encode syntax variables. In the simplest case, a system of order 2 repre-
senting graphs on N nodes or class 1 neurons could be constructed with the help
of N2 class 2 neurons. Such systems have been described in the literature (Hop-
field & Tank, 1986; Kree & Zippelius, 1988; McClelland, 1985; Phillips, Han-
cock, Willson, & Smith, 1988; Sejnowski, 1981). It is important also to process
higher order relations. The direct method of representing them in a complete
way by neurons is not practical. The number of neurons required would be pro-
hibitive. It is conceivable, however, that a combinatorial code can be found in
which sets of class 2 neurons represent higher order relations.

REPRESENTATION OF SYNTAX BY SIGNAL CORRELATIONS

A much more parsimonious and extremely flexible representation of syntax can

be based on temporal correlations between neural signals (von der Malsburg,
1981, 1985, 1987; von der Malsburg & Bienenstock, 1986, 1987). Only node
neurons and their connections are required for the system. Minor alterations to
conventional neural architecture are necessary. If s, denotes the membrane
potential of neuron i, <r, its axonal signal, and Wtj the synaptic weight connecting
neuron j to neuron /, then neural signal dynamics is usually written as

where « and j3 are parameters, Neurophysiology has in this formulation been

simplified in several respects in order to show essentials. Inputs to the system
may be formally subsumed under the second term. A rudimentary inhibitory
system is modeled by the third term, the idea being that inhibition is a service
system to keep the sum of all signals in the system at about a constant level.
Noise may originate in a number of places—in the statistical process of spike
generation, as a consequence of the complexities of synaptic transmission, or
within the membrane—and it is symbolically represented here by an additive
term p,. The axonal signal is derived from the membrane potential by some
functional

which models a threshold function, refractory time, and may also have stochas-
tic aspects to it. The form of Eqs. 1 and 2 is completely conventional. However,
whereas these equations are normally implemented to tend to stationary signal
366 REPRESENTATIONS: BEYOND THE SINGLE CELL

levels (as long as input signals are constant), I propose here to set noise level
and parameters to create stochastic fluctuations in signal levels. Due to struc-
tured input signals and to excitatory and inhibitory connections between neu-
rons, fluctuations in the signals of different neurons are correlated with each
other. Syntactical binding between neurons is simply expressed by a correlation
between their signals (Fig. 18.1). There is no limit to the syntactical order that
can be represetned. There are, however, limits to the amount of syntactical
information which can be expressed in a given time interval, due to limited tem-
poral resolution of neural signals.
When coming to the issue of dynamical processing of syntax a more radical
deviation from conventional network dynamics is required. To this end the sin-
gle synaptic weight variable Wu is replaced by two variables, T-lS and W,r Accord-
ing to Eq. 1, variable Wtj is the efficiency with which synapse ij transmits signal
<J to cell i. This variable is allowed to change on a fast time scale (see below).
The quantity Tl} > 0, on the other hand, is a constant (although it may slowly
change due to synaptic plasticity) and determines the range over which Wij is
allowed to change:

Synapses that do not exist, T(j = 0, cannot be switched on. Dynamic synaptic
weights are regulated by a functional

FIGURE 18.1. Syntactical binding in the temporal domain. The signals of three cells are shown.
Each vertical line represents an action potential. The upper two cells are syntactically bound to
each other, their spikes being correlated in time with a certain temporal resolution. The lower two
cells avoid firing simultaneously and are not bound to each other.
 A Neural Architecture for the Representation of Scenes 367

FIGURE 18.2. Dynamic control of the synaptic weight Wtj between cells j and i. If there are no
signals in cells i and j, the dynamic weight has the resting value Wij. Under the control of syn-
chronous signals cr, and cr,. (upper pair in Fig. 18.1), W,j can be increased within fractions of a
second to the upper limit, Tij (upper branch in left half of figure). If signals cr, and cr, are asyn-
chronous (lower pair in Figure 18.1), Wv can be lowered to 0 (lower branch). When the signal a,
in the postsynaptic neuron ceases W,-,- decays with the time constant of short-term memory to the
resting value (right part of figure).

which is somewhat analogous to Hebbian plasticity, though on a faster time

scale: coincidences between incoming signal a, and strong depolarization of
membrane potential s, increase Wip at most up to the limit T-tj, whereas a signal
j hitting an undepolarized or hyperpolerized cell / decreases Wtj. Time constants
for these actively controlled changes are fractions of a second. When there is no
incoming signal the synapse slowly decays to a resting value intermediate
between 0 and T,j (Fig. 18.2).
Synaptic plasticity—change in the "permanent" synaptic parameter Tif—is
controlled by the dynamic weight W,f.

The function F( WtJ) is shaped according to the idea that a connection is allowed
to grow only if the two neurons involved have something to do with each other,
that is, are syntactically bound to each other, which in turn is expressed by signal
correlations and a maximally activated dynamic weight Wtj. As far as learning
and memory are concerned, this is the point of the exercise: to limit synaptic
plasticity to those pairs of neurons which are syntactically bound to each other
in a given situation, although many more neurons are active in that situation.

DYNAMICS OF SYNTACTICAL STRUCTURES

The formulation given here is meant to be sketchy and leaves out many details
that would have to be specified in a simulation. Various concrete versions have
368 REPRESENTATIONS: BEYOND THE SINGLE CELL

been tried, each one optimizing its own set of goals (Bienenstock & von der
Malsburg, 1987; von der Malsburg, 1985, 1988; von der Malsburg & Bienen-
stock, 1987; von der Malsburg & Schneider, 1986). Important aspects, however,
are common to all of them. Very few or no arbitrary choices have been made in
Eq. 1 to 5, yet they favor syntactical patterns of very specific structure. The
momentary set {Wij,} of synaptic weights together with an ensemble of input sig-
nals determines the statistical structure of signals according to Eqs. 1 and 2.
These signals contain certain correlations which are shaped by correlations in
the input signals and by cooperating sets of active connections. These correla-
tions in turn act back on the connectivity structure according to Eq. 4. If input
signals are sufficiently stationary in their statistical properties, this feedback loop
between the connectivity state and internal signal statistics continues to change
the dynamic state of the system until stationarity is reached in terms of connec-
tivity and signal statistics, that is, in terms of syntactical structure.
What is the structure of connectivity patterns which are favored by this
dynamical process? The graph of active connections in a connection patterns is
sparse in the present formulation due to the /3 term in Eq. 1 which limits activit
to fairly small sets of active neurons in each time slice. (A tendency to form
sparse connection patterns is one of the very few "arbitrary" choices to be imple-
mented in the system. Without this tendency the system would tend to the unin-
teresting and featureless total connectivity state.) The sparse graph of a connec-
tivity pattern can locally be characterized by cooperative combinations of active
connections: combinations that are effective in creating strong correlations. A
favorable local constellation of connections is, for instance, a small set of neu-
rons with complete connectivity (the set has to be small due to the sparseness
constraint). Another favorable constellation is a set of short alternative path-
ways of equal delay between two cells a and b: if a fires, a wave of activity is
transported over the several pathways to b where, accordingly, all participating
synapses experience a favorable event.
If the dynamical aspect of brain architecture is described realistically here,
there is an a priori structure imposed on all dynamic states of the brain. This
limits activity to a certain universe of syntactical structures or connectivity-and-
correlation patterns. There is thus a potent structure concept, as was asked for
earlier. Global characterization of these connectivity patterns is an important
unsolved problem. It has already been shown, however, that some very useful
properties are implicit in the system. Two-dimensional topological graphs,
which are ideal for representing and storing visual images, belong to the favored
universe (von der Malsburg & Bienenstock, 1987). If there are two isomorphic
connectivity patterns in different parts of the network, connectivity dynamics is
able to discover this fact and activate a connectivity pattern which maps the two
networks to each other in terms of the isomorphism (Bienenstock & von der
Malsburg, 1987; von der Malsburg, 1988; von der Malsburg & Bienenstock,
1986). This can be conveniently used for position-invariant image recognition
(Bienenstock & von der Malsburg, 1986; von der Malsburg, 1981, 1988). Many
different connectivity patterns on the same set of neurons can be stored and
 A Neural Architecture for the Representation of Scenes 369

retrieved (von der Malsburg, 1985; von der Malsburg & Bienenstock, 1987).
Finally, separation of figure from ground has been demonstrated (von der Mals-
burg & Schneider, 1986).

THE GENERALIZED SCENE IN NEURAL ARCHITECTURE

The dynamical architecture described here is not an easy blueprint for the solu-
tion of all brain problems. The other aspect to architecture, permanent connec-
tivity structure {Tij} is a much more voluminous issue. However, many prob-
lems that have proved to be difficult in conventional architecture may be much
more easily solved in an architecture able to process syntactical structure.
One of the great problems awaiting solution is representation of the general-
ized scene in neural architecture. The difficulty arises from a number of aspects.
One is the necessity to represent many components coexisting in a scene. A rigid
representation of a visual scene, like a photograph, is not a basis for generaliza-
tion and thus is not acceptable. The main reason is that accidental aspects, like
the relative positions of objects, and essential aspects, like geometrical relation-
ships within an object, are mixed with each other. In addition, a pictorial rep-
resentation cannot accommodate abstract aspects of generalized scenes. The sit-
uation cries for a symbol system with syntax. Let me try to indicate the style of
representation that would be possible.
Let us first talk about the representation of a single object. In a syntactical
system designators can be attached to it which specify various aspects of func-
tional importance. There may be a great variety of designators. Some may spec-
ify the position of the object in various frames of reference, for example, in
world coordinates, in body coordinates, or in visual coordinates. Another kind
of designator may specify geometrical relations of the object to others. Desig-
nators may attach to an object type information, names, or material. Also
important is the attachment of designators of functional role: Is it edible? Is it
dangerous? Is it agreeable or disagreeable? Is it the target of my intended move-
ment? Does it play a causal role in some process? For a linguistic object gram-
matical role has to be attached. The alternative to the attachment of a designator
in a system without syntax would be the creation of a new symbol whenever
some specification is to be added.
If there was only one object in the whole scene, attachment of designators
would be possible by simply activating the corresponding symbols. In realistic
scenes, however, there have to be many objects at the same time. In a gram-
matical sentence with two nouns it is, for instance, necessary to attach the role
of subject to one and the role of object to the other. Simple simultaneous activity
in the symbols involved would lead to conjunction ambiguity, as illustrated
above. If there is a flexible syntax system, however, complicated networks of
objects and designators can be created. The designators themselves can be fur-
ther specified by more designators. Also what is perceived as a single object may
be represented by a network of descriptors and relationships (Bienenstock & von
370 REPRESENTATIONS: BEYOND THE SINGLE CELL

der Malsburg, 1987), so that the object cannot be identified with any elementary
symbol. A single object is then represented as a scene.
The great distinction between cultural communication symbols and symbols
in the brain lies in the way they are used. Communication signals are passive
until they are interpreted by a receiver. Brain symbols form a dynamical system
in themselves. The whole significance of one symbol lies in its ability to create
another symbol in the next moment, in another part of the brain or in the same.
An object represented by a network of cross-referencing designators acts as
boundary condition for other pattern formation processes in which more con-
nection patterns are formed. It is claimed here that a fixed neural architecture
constructed along lines similar to those discussed around Eq. 1 to 4 sufficies to
act as the "set of rules governing this arrangement," as required by our dictio-
nary definition. These general metagrammatical rules are to be specified to the
needs of particular types of scenes and objects with the help of specific restric-
tions to connectivity imposed by the "permanent" parameters Tv, which in turn
can be modified by plasticity of the type in Eq. 5. A constellation like "there is
an apple; it is an edible object; it is within the reach of my hand; it is my prop-
erty," together with the symbol "I am hungry," is able to induce another com-
plex symbol, "I have the intention to grasp; the object of the grasp is an edible
object; this object is the apple; the apple has such and such body-frame coordi-
nates," together with motor descriptions of appropriate movements of my arm
and hand.
The claim I am making here will have to be validated in the course of a long
reasearch program. Just now it is not much more than a vision and a scientific
program. At best I may be able to convince a few specialists who are aware of
the difficulties and problems involved in modeling even the most elementary
capabilites of our brain. Only the very first simple steps of the program (von der
Malsburg, 1981) have been demonstrated so far, as cited above. Experimental
demonstration of some aspects of the dynamical architecture formulated in Eqs.
1 to 5 would be useful. Possible experiments have been described (von der Mals-
burg, 1987). The crucial test will, however, be technical demonstrations. If these
are successful, their authority will influence our neurobiological thinking.

REFERENCES

Ackley, D. H., Hinton, G. E., & Sejnowski, T. J. (1985). A learning algorithm for Boltzmann
machines. Cognitive Science, 9, 147-169.
Anderson, J. A. (1970). Two models for memory organization using interacting traces. Math-
ematical Biosciences, 8, 137-160.
Bienenstock, E., & von der Malsburg, C. (1987). A neural network for invariant pattern recog-
nition. Europhysics Letters, 4(1), 121-126.
Cooper, L. N. (1974). A possible organization of animal memory and learning. In B. Lundquist
& S. Lundquist (Eds.), Proceedings of the Nobel Symposium on Collective Properties
of Physical Systems (pp. 252-264). New York: Academic Press.
 A Neural Architecture for the Representation of Scenes 371

Fukushima, K. (1980). Neocognitron: A self-organizing neural network model for a mechanism

of pattern recognition unaffected by shift in position. Biological Cybernetics, 36, 193-
202.
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective compu-
tational abilities. Proceedings of the National Academy of Sciences USA, 79, 2554-
2558.
Hopfield, J. J., & Tank, D. W. (1986). Collective computation with continous variables. In E.
Bienenstock, F. Fogelman Soulie, & G. Weisbuch (Eds.), Disordered systems and bio-
logical organization (NATO ASI Series, Vol F20, pp. 155-170). Berlin: Springer-
Verlag.
Julesz, B. (1984, February). A brief outline of the texton theory of human vision. Trends in
Neurosciences, 41-45.
Kohonen, T. (1977). Associative memory. Berlin: Springer-Verlag.
Kree, R., & Zippelius, A (1988). Recognition of topological features of graphs and images in
neural networks. Journal of Physics, A21, L813.
Longuet-Higgins, H.-C. (1967). Holographic model of temporal recall. Nature, 217, 104.
Marko, H., & Giebel, H. (1970). Recognition of handwritten characters with a system of homo-
geneous layers. Nachrichtentechnische Zeitschrifi, 9, 455-459.
McClelland, J. L. (1985). Putting knowledge in its place: A scheme for programming parallel
processing structures on the fly. Cognitive Science, 9, 113-146.
Palm, G. (1980). On associative memory. Biological Cybernetics, 36, 19-31.
Phillips, W. A., Hancock, P. J. B., Willson, N. J., & Smith, L. S. (1988). On the acquisition of
object concepts from sensory data. In R. Eckmiller & C. von der Malsburg (Eds.),
Neural computers (NATO ASI Series: Computer Systems Sciences, Vol. 41). Berlin:
Springer-Verlag.
Rosenblatt, F. (1961). Principles of neurodynamics: Perceptrons and the theory of brain mech-
anisms. Washington, D.C.: Spartan Books.
Rumelhart, D. E., & McClelland, J. L. (Eds.). (1986). Parallel distributed processing (Vol. 1).
Cambridge, Mass.: MIT Press.
Sejnowski, T. J. (1981). Skeleton filters in the brian. In G. E. Hinton & J. A. Anderson (Eds.),
Parallel models of associative memory (pp. 190-212). Hillsdale, N.J.: Erlbaum.
Sejnowski, T. J., Kienker, P. K., & Hinton, G. E. (1986). Learning symmetry groups with hid-
den units: Beyond the perceptron. Physica, 22D, 260-275.
Steinbuch, K. (1961). Die Lernmatrix. Kybernetik, 1, 1-36.
Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of attention. Cognitive Psy-
chology, 12, 97-136.
von der Malsburg, C. (1981). The correlation theory of brain function (Internal Report 81-2).
MPI Biophysical Chemistry, P.O. Box 2841, 3400 Gottingen, West Germany.
von der Malsburg, C. (1985). Nervous structures with dynamical links. Berichte Bunsengesell-
schaft fuer Physikalische, Chemie, 89, 703-710.
von der Malsburg, C. (1987a). 1st die Evolution blind? In B.-O. Kuppers (Ed.), Ordnung aus
dem Chaos. Munich: Piper.
von der Malsburg, C. (1987b). Synaptic plasticity as basis of brain organization. In J.-P. Chan-
geux & M. Konishi (Eds.), The neural and molecular bases of learning, Dahlem Kon-
ferenzen (pp. 411-431). Chichester: Wiley.
von der Malsburg, C. (1988). Pattern recognition by labeled graph matching. Neural Networks,
1, 141-148.
von der Malsburg, C., & Bienenstock, E. (1986). Statistical coding and short-term synaptic plas-
ticity: A scheme for knowledge representation in the brain. In E. Bienenstock, F.
Fogelman, & G. Weisbuch (Eds.), Disordered systems and biological organization
(NATO ASI Series, Vol. F20, pp. 247-272). Berlin: Springer-Verlag.
von der Malsburg, C. & Bienenstock, E. (1987). A neural network for the retrieval of super-
imposed connection patterns. Europhysics Letters, 3(11), 1243-1249.
372 REPRESENTATIONS: BEYOND THE SINGLE CELL

von der Malsburg, C., & Schneider, W. (1986). A neural cocktail-party processor. Biological
Cybernetics, 54, 29-40.
von der Malsburg, C., & Singer, W. (1988). Principles of cortical network organization. In P.
Rakic & W. Singer (Eds.), Neurobiology of neocortex (pp. 69-99). Chichester: Wiley.
Werman, M. (1987). The capacity of k-gridgraphs as associative memory (Reports in Pattern
Analysis No. 146). Providence, R.I.: Brown Unversity, Division of Applied
Mathematics.
Willshaw, D. J., Buneman, O. P., & Longuet-Higgins, H.-C. (1969). Non-holographic associa-
tive memory. Nature, 222, 960-962.
COMMENTARIES AND
ALTERNATIVE PERSPECTIVES
This page intentionally left blank
19
Representations: Who Needs Them?
WALTER J. F R E E M A N
C H R I S T I N E A. S K A R D A

Biologists by tradition have seldom used the term representation to describe

their findings. Instead they have relied on phrases such as "receptor field" on
the sensory side and "command" or "corollary discharge" on the motor side
when discussing neural control of sensation and motion in goal-directed behav-
ior. Such words connote dynamic process rather than symbolic content. One
might suppose that this neglect of a now common word reflects diffidence about
discussing so-called higher functions of the brain, owing to a humbling lack of
understanding of the brain's complexity. Inspection of biology textbooks belies
this view. Biologists have shown no lack of hubris in pontificating about the
properties of the brain supporting mental functions. On the contrary, they have
always taken pride in being uniquely qualified to explain brain function to any-
one willing to listen.
A turning point came in the 1940s with the popularization of digital and ana-
log computers as "giant brains" and with the adoption of the Turing machine
as a model for explanation. In this conception, which is lucidly illustrated by
von der Malsburg (Chapter 18, this volume), the human ability to understand
the world is likened to the procedure of incorporating information on a tape into
a machine by means of symbols. Cognitive operations are interpreted as the
manipulation of these symbols according to certain semantic rules. At present
we are all so accustomed to this metaphor that it seems self-evident (Goldman-
Rakic, 1987). The brain's job is to incorporate features of the outside world and
make internal syntactical representations of these data, which together consti-
tute a world model that serves to control motor output. Any other account
appears to be "noncognitive" (Earle, 1987) and counterintuitive. In short, to
question this commonsense notion seems quixotic, sophistic, and arbitrary.
We propose, however, that physiologists avoid this way of thinking for two
reasons. One is that no one now understands how brains work, but the use of
the term representation and its attached concepts tends to obscure this fact. The
term gives us the illusion that we understand something that we do not. We
suggest that the idea of representation is seductive and ennervating, promising

375
376 REPRESENTATIONS: BEYOND THE SINGLE CELL

good deals but delivering nothing new. When researchers refrain from using the
term, knowledge of brain function is not significantly affected. We conclude that
use of the term is unnecessary to describe brain dynamics.
The second reason is that the use of the metaphor points us in a direction that
carries physiological research away from more profitable lines of inquiry. We
have found that thinking of brain function in terms of representation seriously
impedes progress toward genuine understanding.
An example is taken from our studies of the behavioral correlates of the elec-
troencephalograms (EEGs) of the olfactory system under conditioning (Freeman
& Skarda, 1985). The EEGs of the olfactory bulb and cortex show a brief oscil-
latory burst of potential that accompanies each inhalation. This can be likened
to a burst of energy carried by a wave of neural activity at a common frequency.
Each burst exists over the entire bulb or cortex with a spatial pattern of ampli-
tude that varies from one burst to the next. We have shown that a stereotypical
pattern recurs whenever a. particular odorant is presented that the animal has
been trained to respond to.
For more than 10 years; we tried to say that each spatial pattern was like a
snapshot, that each burst served to represent the odorant with which we corre-
lated it, and that the pattern was like a search image that served to symbolize
the presence or absence of the odorant that the system was looking for. But such
interpretations were misleading. They encouraged us to view neural activity as
a function of the features and causal impact of stimuli on the organism and to
look for a reflection of the environment within by correlating features of the
stimuli with neural activity. This was a mistake. After years of sifting through
our data, we identified the problem: it was the concept of representation.
Our research has now revealed the flaws in such interpretations of brain func-
tion. Neural activity patterns in the olfactory bulb cannot be equated with inter-
nal representations of particular odorants to the brain for several reasons. First,
simply presenting an odorant to the system does not lead to any odor-specific
activity patterns being formed. Only in motivated animals, that is, only when
the odorant is reinforced leading to a behavioral change, do these stereotypical
patterns of neural activity take shape. Second, odor-specific activity patterns are
dependent on the behavioral response; when we change the reinforcement con-
tingency of a CS we change the patterned activity. Third, patterned neural activ-
ity is context dependent: the introduction of a new reinforced odorant to the
animal's repertoire leads to changes in the patterns associated with all previously
learned odorants. Taken together these facts teach us that we who have looked
at activity patterns as internal representations of events have misinterpreted the
data. Our findings indicate that patterned neural activity correlates best with
reliable forms of interaction in a context that is behaviorally and environmen-
tally co-defined by what Steven Rose (1976) calls a dialectic. There is nothing
intrinsically representaticnal about this dynamic process until the observer
intrudes. It is the experimenter who infers what the observed activity patterns
represent to or in a subject, in order to explain his results to himself (Werner,
1988a, 1988b).
 Representations: Who Needs Them? 377

The impact of this insight on our research has been significant. Once we
stopped looking at neural activity patterns as representations of odorants, we
began to ask a new set of questions. Instead of focusing on pattern invariance
and storage capacity, we began to ask how these patterns could be generated in
the first place from less ordered initial conditions. What are the temporal
dynamics of their development and evolution? What are their effects on the neu-
rons to which they transmit? What kinds of structural changes in brains do they
require and do they lead to? What neuromodulators do these pattern changes
require? What principles of neural operations do these dynamical processes
exemplify and instantiate? In short, we began to focus less on the outside world
that is being put into the brain and more on what brains are doing.
Our efforts to answer these questions led us to develop mathematical, statis-
tical, and electronic models that describe and explain the neural dynamics of
pattern generation. These models have, in turn, caused radical changes in our
views of how brains operate. In particular, we now see brains as physicochemical
systems that largely organize themselves, rather than reacting to and determined
by input. As Carew and co-workers (Chapter 2, this volume) showed, each brain
has a history that begins with simple structures and that evolves through innu-
merable stages and phases of growth and development to increasing order and
complexity. The patterns are formed from within and not imposed from outside,
as is commonly supposed to occur in brains under sensory stimulation. We have
found that an essential condition for these patterns to appear is the prior exis-
tence of unpatterned energy distributions which appear to be noise, but which
in reality are chaos. New forms of order require that old forms of order collapse
back into this chaotic state before they can appear. Therefore, in the EEG we see
each burst appearing from chaotic basal activity and collapsing back into chaos,
thereby clearing the way for the next burst of patterned activity (Skarda & Free-
man, 1987).
These findings challenge two widely held assumptions concerning brain
dynamics. First, conventional theory holds that full information is delivered
into the system and that thereafter it is degraded by noise. This property is anal-
ogized to entropy. However, chaotic systems like the brain are open and, by
virtue of energy throughput, operate far from equilibrium. They internally create
new information and can be described as negentropic (Tsuda & Shimizu, 1985).
The brain has immunity from the first and second laws of thermodynamics
because its assured blood supply brings it more energy than it can use and carries
off waste heat and entropy. As a result the formalisms of information theory that
underlie the representation-based computational metaphor of brain dynamics
do not apply to the neural networks of biological systems, because these for-
malisms make no sense in systems with positive information flow.
Second, the conventional description of signals embedded in noise is inappro-
priate. The same neural system that generates bursts (signals) also generates the
background state of chaotic activity (often thought to be noise). When the sys-
tem switches (bifurcates) from chaos to burst activity, the chaotic activity stops
and the signal starts. Chaos operates up to the moment of bifurcation. It plays
378 REPRESENTATIONS: BEYOND THE SINGLE CELL

no role of "annealing" thereafter because response selection has already taken

place, convergence is assured, and there is no role for "noise." The metaphor of
the "signal to noise ratio" is inappropriate for brain function, yet it is essential
for representation in man-made systems.
These considerations are well illustrated by the preceding four chapters. Of
these the report by Cooper, Bear, Ebner, and Scofield (Chapter 15) deals most
directly with neurophysiological data, and they alone make no reference to rep-
resentations. Their model describes the dynamics of modifiable synapses over
the time scale of learning. It does not address the dynamic of stimulus-induced
neural activity on the time scale of responding. It attempts to explain the
dynamics in terms of membrane conductances and calcium fluxes and not the
semantic content of the input. The strength of their model lies in the identifi-
cation of global variables as important for consideration; the main weakness is
their decision to "delay consideration of the global variables by assuming that
they act to render cortical synapses modifiable or nonmodifiable by experience."
Their chapter emphasizes the great value of the "mean field" state variable
for the analysis and understanding of physical systems composed of ensembles.
They elect not to consider the basis for defining and using such variables in their
work with neural networks, thereby depriving themselves and their readers of
access to the existing literature, including well-reasoned approaches to brain sys-
tems from the classical standpoints of statistical mechanics (Amari, 1974; Wil-
son & Cowan, 1972) and of nonequilibrium thermodynamics from the school
of Prigogine (Babloyantz & Kaczmarek, 1981; Freeman, 1975). Both approaches
emphasize the importance of mutually excitatory (positive) feedback within
laminar distributions of nerve cells by their recurrent collateral axons and the
renewal process in cell firing, leading to the emergence of macroscopic state vari-
ables to represent the activity of local neighborhoods, that is, axonal pulse den-
sity and dendritic current density functions that are continuous in both time and
the spatial dimensions of cerebral cortex. These activity states are readily
observed in many parts of brains by use of electrodes, magnetic probes, and
optical dyes and by computer-implemented spatial filtering, summing, and
enhancement of the raw data for visual display (Freeman, 1987). By means of
these and related well-documented procedures these investigators and others
can test their models directly with respect to brain dynamics.
Chapter 18 by von der Malsburg and part of Chapter 17 by Sejnowski and
Tesauro are at the opposite pole and take representation for granted. They are
also explicitly about machines and not about brains. "Representations" that are
selected and defined by the observer serve as the goals or end points of the evo-
lution of the machines. Sejnowski and Tesauro present elements on both sides.
The description of Gerald Westheimer's dynamics of spatial attractors, which in
some ways recalls Wolfgang Kohler's field theories, is accounted for by the
dynamics of mutually excitatory feedback. This is physiology (Freeman, 1975).
When he asks "Who reads the population code?" he gives engineering answers:
"find all possible depths and find which matches the closest with minimal
error." The algorithms of backpropagation and error correction by the observer-
 Representations: Who Needs Them? 379

instilled "teacher" and "correct answer" are machine processes that do not exist
in biological brains. "NETtalk" can transduce optical characters to sounds that
are recognizable by human observers, so it is a machine that can be shaped to
read to the blind, but one cannot say that the machine has learned to read in the
sense that a schoolchild has.
Kohonen (Chapter 16) most clearly addresses the nature of internal represen-
tations as they are needed and used by engineers and machines. Each represen-
tation has characteristics and attributes that are to be stored, matched, and
retrieved by processes ultimately deriving from mappings. Our physiological
data show that episodic storage of odor trials does not happen, that "retrieval"
is not recovery but re-creation, always with differences, and that stimulus-bound
patterns cannot coexist with re-created patterns to support matching procedures.
We agree with Kohonen's statement that we are faced with semantic difficulties,
and we conclude that they stem from deep incompatibilities between the dynam-
ics respectively of biological and present-day artificial intelligence. The key
words to look for are "best matching" and "error detection," because these refer
to machine cognition and not neural cognition.
These considerations give an answer to our question about representations.
Who needs them? Functionalist philosophers, computer scientists, and cognitive
psychologists need them, often desperately, but physiologists do not, and those
who wish to find and use biological brain algorithms should also avoid them.
They are unnecessary for describing and understanding brain dynamics. They
mislead by contributing the illusion that they add anything significant to our
understanding of the brain. They impede further advances toward our goal of
understanding brain function, because they deflect us from the hard problems
of determining what neurons do and seduce us into concentrating instead on the
relatively easy problems of determining what our computers can or might do.
In a word, representations are better left outside the laboratory when physiolo-
gists attempt to study the brain. Physiologists should welcome the ideas, con-
cepts, and technologies brought to them by brain theorists and connectionists,
but they should be aware that representation is like a dose of lithium chloride;
it tastes good going down but it doesn't digest very well (Bures, Chapter 1, this
volume).

REFERENCES

Amari, S. (1974). A method of statistical neurodynamics. Kybernetik, 14, 201-215.

Amari, S. (1987). Statistical neurodynamics of associative memory. (Tech. Rep. METR87-8).
Tokyo: University of Tokyo, Department of Mathematics, Engineering, and Instru-
ment Physics.
Babloyantz, A., & Kaczmarek, L. (1981). Self-organization in biological systems with multiple
cellular contacts. Bulletin of Mathematical Biology, 41, 193-201.
Earle, D. C. (1987). On the differences between cognitive and noncognitive systems. Brain and
Behavioral Science, 10, 177-178.
Freeman, W. J. (1975). Mass action in the nervous system. New York: Academic Press.
380 REPRESENTATIONS: BEYOND THE SINGLE CELL

Freeman, W. J. (1987). Analytic techniques used in the search for the physiological basis of the
EEG. In A. S. Gevins & A. Remond (Eds.), EEC Handbook (pp. 583-664). Amster-
dam: Elsevier.
Freeman, W. J., & Skarda, C. A. (1985). Spatial EEG patterns, nonlinear dynamics and percep-
tion: The neo-Sherringtonian view. Brain Research, 10, 147-175.
Goldman-Rakic, P. (1987). Circuitry of primate prefrontal cortex and the regulation of behav-
ior by representational memory. In F. Plum (Ed.), Handbook of physiology: Sec. 1.
The nervous system (pp. 373-417). Bethesda Md.: American Physiological Society.
Gray, C. M., Freeman, W. J., & Skinner, J. E. (1986). Chemical dependencies of learning in the
rabbit olfactory bulb: Acquisition of the transient spatial pattern change depends on
norepinephrine. Behavioral Neurosdence, 100, 585-596.
Rose, S. P. R. (1976). The conscious brain. New York: Vintage Books.
Skarda, C. A., & Freeman, W. J. (1987). Brain makes chaos to make sense of the world. Brain
and Behavioral Science, 10, 161-195.
Tsuda, I., & Shimizu, H. (1985). Self-organization of the dynamical channel. In H. Haken (Ed.),
Complex systems: Operational approaches in neurobiology, physics and computers
(pp. 240-251). Berlin: Springer-Verlag.
Werner, G. (1988a). Five decades on the path to naturalizing epistemology. In J. S. Lund (Ed.),
Sensory processing in the mammalian brain (pp. 345-359). New York: Oxford Uni-
versity Press.
Werner, G. (1988b). The many faces of neuroreductionism. In E. Basar (Ed.), Dynamics of
sensory and cognitive processing by the brain (pp. 241-257). Berlin: Springer-Verlag.
Wilson, H. R., & Cowan, J. D. (1972). Excitatory and inhibitory interactions in biological pop-
ulations of model neurons. Biophysics Journal, 12, 1-24.
20
Interactions Within Neuronal
Assemblies: Theory and Experiment
G E O R G E L. G E R S T E I N

Models of neuronal networks have been with us since the mid-fifties, starting
with the work of Rochester (Rochester, Holland, Haibt, & Duda, 1956) and of
Farley (1964; Farley & Clarke, 1961) and their associates. After years of steady
growth, the field has become extremely active, as exemplified by the chapters by
Cooper and co-workers, Kohonen, von der Malsburg, and Sejnowski and
Tesauro in this volume. In contrast, direct experimental access to real neuronal
networks in the brain has been possible only since the early eighties. It is the
purpose of this chapter to juxtapose some of the ideas and assumptions made
in modeling networks to recent experimental results on real neurons.
Neuronal network models typically consist of many nodal elements that have
simplified properties intended to mimic some aspects of real neurons. The min-
imum set of properties for a nodal element specifies the (linear or nonlinear)
way its inputs will be combined and defines the conditions (possibly nonlinear)
under which an output will occur. The extent to which further properties are
incorporated into the nodal elements depends on the purpose of the modeling.
It is possible to have nodal elements that mimic large dendritic trees with the
attendant spatiotemporal summation of inputs. However, for many purposes it
is appropriate to simplify to nodal elements which are, in effect, like "point neu-
rons" without spatial extent.
Connections between nodes in a network model mimic axons and synapses.
The important parameters are delay (conduction) time and connection strength
or influence. For a network model with TV nodes, connections between node
can be specified by an TV X N matrix. An element of this matrix, say WtJ, woul
represent the influence of node j on node /. Excitatory or inhibitory influence
can be represented by the sign of the Wv; its magnitude can reflect the strength
of the influence. A value of zero indicates no connection (influence) between the
specified nodes.
Many of the studies involving neural network models have examined mech-
anisms that could be responsible for learning and memory in biological systems.

381
382 REPRESENTATIONS: BEYOND THE SINGLE CELL

In principle the necessary modulations of a network model could be accom-

plished by an evolution of nodal characteristics. It has been more usual, how-
ever, to locate the necessary modulations in the connection strengths between
the nodal elements. The rules used for such modulations are often based on
Hebb's (1949) proposal that the strength (or efficacy) of a synapse will be
increased by a small amount every time that the presynaptic and postsynaptic
neurons fire in close temporal sequence. Other rules for synaptic modification
have been studied and are reviewed by Palm (in press).
An underlying assumption of most of the rules for change of synaptic weight
is that such changes take place very slowly, on a time scale appropriate to our
conception of learning as a slow process that requires many repetitions of the
appropriate input. Thus modelers have taken the Wij matrix of connectivity to
be almost a constant, with only a slow evolution of its values during the "train-
ing" phases of an experiment. It is precisely this assumption that is challenged
by recent experimental results on real neuronal assemblies. It turns out, as I shall
show, that the "effective connectivity" between neurons, which corresponds to
the modeler's Wij is dynamic on several time scales and is hardly to be described
as a slowly evolving constant.

EXPERIMENTAL AND ANALYTIC TOOLS

The central requirement for experimental observation of neuronal assembly

properties is simultaneous recording of the activity of as many individual neu-
rons as technically possible. This class of experiment allows a much greater
degree of inference than sequential traditional single neuron observations:
instead of being able to examine only the relations between activity of a single
neuron and the occurrence of external events (stimulus, movement, etc.), we
may now examine the interrelationship of the activities of two (or more) neu-
rons. The technical requirements of this type of multineuron experiment are
very demanding, so that relatively few laboratories have built up the necessary
instrumentation. For a review of some of the ways that multineuron experi-
ments involve a nonlinear increase indifficulty over the single-neuron experi-
ment, see Gerstein, Bloom, Espinosa, Evanczuk, and Turner (1983) and Kruger
(1983).
Current technology allows the simultaneous extracellular observation of up to
30 neurons, although half that amount is a more frequently attained goal. Such
numbers are preposterously small on the scale of functional neuron groupings
like visual cortical columns, which may involve many thousand neurons. Thus
current technology for multineuron experiments leaves us with a difficult sam-
pling problem.
The usual way to assess interrelationships among the observed neurons is to
measure the temporal correlation of their firing activity. The use of cross-cor-
relation for two or three neurons is clearly reviewed by Glaser and Ruchkin
(1976), who also give many useful references to the original literature. Basically,
 Interactions Within Neuronal Assemblies: Theory and Experiment 383

for two neurons, the cross-correlogram estimates the amount of near-coincident

firing and determines whether there is a favored order and delay between spikes
in the two trains. Such favored compound events can in principle serve directly
as carriers of information, since convergent, near-coincident activity of two syn-
apses is much more effective than activation of either synapse alone.
In addition, however, we may estimate the amount of near-coincident firing
that would be expected for two neurons exhibiting the observed individual firing
rates (and their possible modulations by stimulus). By subtracting this estimate
from the original directly observed correlation, we may estimate the amount of
"excess" firing correlation in comparison to the model of two independent neu-
rons. This excess can be interpreted as an "effective connectivity" between the
two neurons. Such effective connectivity is not directly related to the actual ana-
tomical connectivity that may exist between the two neurons, although it may
be its currently active subset in some cases. Generally the effective connectivity
is simply a shorthand for the interconnections of the simplest two neuron model
that would replicate the experimentally observed cross-correlograms. In this
respect it is the analogue of the W,j in the connectivity matrix in neuronal net-
work models.
Ultimately we want to examine the dependence of the correlation structure
and the effective connectivity among the observed group of neurons on the
experimental conditions (current or recent stimulus, context, response, etc.) dur-
ing which the recording was taken. If we analyze the neuronal spike trains in
pairs, as outlined above, then even with the modest number of neurons that can
presently be simultaneously observed in a multineuron experiment, combina-
torial arithmetic will force the experimenter to examine and sort out literally
thousands of cross-correlograms.
A recently developed analytical tool, at least in a screening strategy, allows
avoidance of the combinatorial proliferation of cross-correlograms by treating
the entire group of recorded neurons and spike trains as a single entity rather
than as a large set of pairs (Gerstein & Aertsen, 1985; Gerstein, Perkel, & Day-
hoff, 1985). The so-called gravity analysis transforms the TV observed spike trains
into an N-dimensional mathematical abstraction in which each neuron is rep-
resented by a charged particle. The charge is a low-pass filtered version of the
spike train from the corresponding neuron. The charged particles exert force on
each other, with like signs of charge attracting; as a consequence the particles
move about. The largest force occurs when two particles have large charge at the
same time, which, in turn, corresponds to near-coincident firing of the corre-
sponding neurons. The end result is that those particles representing neurons
that tend to have "excess" near-coincident firing will tend to aggregate. Thus the
temporal analysis of a set of TV observed spike trains has been transformed into
a problem of cluster analysis in the abstract N-dimensional space.
The spatial clustering of points in the gravity representation is best examined
in the full TV space, where the shape of the aggregations can be interpreted i
useful ways. However, for purposes of publication, since we can hardly make a
picture of particles in an TV space (with N greater than 3), we will use a projection
384 REPRESENTATIONS: BEYOND THE SINGLE CELL

from the N space to a two-dimensional plane. This, of course, is a procedure

that can be deceptive, since points that cluster in the two-dimensional plane
need not have done so in the full N-dimensional space. However, we will sup-
plement the projection with a graph of the distances between all pairs of particles
in the N space as the aggregation develops, so that validity of the information
in the two-dimensional representation can be checked. The dynamics of the
aggregation process may itself be of interest, and it is then appropriate to make
a movie of the projection. For our purposes here, it will suffice to simply exam-
ine the aggregations attained as the available data reach an end. We assume, of
course, a standard starting configuration of the particles for all calculations.

TIME SCALES

Almost all the measurements used on spike trains involve averaging over some
time period, although this is not always explicitly stated. Thus cross-correlo-
grams, for example, represent a measure of near-coincident firing as averaged
over the piece of recording under study. In the same sense, the aggregations cal-
culated in the gravity representations are the result of the temporal correlation
structure of the observed spike trains, as averaged over the analyzed recording
period. In other words, these analysis methods are not instantaneous measure-
ments and imply a "reasonable" observation time.
If we want to compare the correlation structure among some observed spike
trains in two different stimulus conditions, we must present the two stimuli
alternately until a sufficient amount of recording for analysis in each condition
has accumulated. This approach eliminates misinterpretation of slow physiolog-
ical drifts, since these will apply equally to both stimulus conditions over the
recording time. Before analysis the data are "cut and pasted" to produce mate-
rial corresponding to a single stimulus. However, note that the stimulus alter-
nation cycle is much shorter than the total analysis time: typical runs require 40
to 100 stimulus repetitions during the analysis time. Thus we are comparing the
average correlation structure of the observed set of spike trains in each of two
states that typically alternate on a time scale of 0.5 to 1 sec. Any observed dif-
ferences between the correlation structures corresponding to the two stimuli will
represent changes in the observed neuronal network that are on this time scale.

RESULTS: ASSEMBLY CORRELATION STRUCTURE

With multineuron recording, the analysis tools defined above, and an experi-
mental design that presents several interleaved stimuli repeatedly, we now may
examine the correlation structure exhibited by small groups of real neurons. The
material shown in the figures is drawn from studies on auditory cortex of cat.
Details of the experiments have previously been described (Bloom & Gerstein,
1986; Espinosa & Gerstein, 1988) and can be omitted for present purposes. The
essential point is that we can compare the (average) correlation structure among
 Interactions Within Neuronal Assemblies: Theory and Experiment 385

the observed neurons under two (or more) different stimulus conditions by com-
paring the gravity condensations.
Figure 20.1 shows the distances between particle pairs for data containing
eight cortical neurons under two different stimulus conditions. These gravity
aggregations were run with appropriate compensation for the individual neuron

FIGURE 20.1. Interparticle distances during gravity condensations for eight cortical neurons
under two different stimulus conditions. Individual stimuli within each condition were presented
every 550 msec. Each panel corresponds to one stimulus condition and shows all particle pair
distances for the same eight observed cortical neurons.
386 REPRESENTATIONS: BEYOND THE SINGLE CELL

firing rates, so that we are here examining the "excess" correlation that we earlier
associated with "effective connectivity" and the Wit. (For more detail on the
compensation for individual neuron firing rates see Aertsen, Bonhoeffer, & Kru-
ger, 1987; Gerstein & Aertsen, 1985.) It is apparent that the development of
particle pair distances is different under the two stimulus conditions, and hence
that the "excess" correlation structure of the spike trains was different in the two
conditions.
Figure 20.2 shows the same data in the two-dimensional projection plane of
the gravity representation. The starting arrangement is the same for each stim-
ulus condition, as shown on the left; the final configurations on the right show
that certain particles change their "allegiance." Again, this shows that the
"excess" correlation structure was different in the two stimulus conditions.
Now it turns out that it is easy to find multineuron data that behave as shown
in the figures. During the past year, we have obtained multineuron recordings
from other laboratories which were recorded from a variety of creatures and
brain structures. Many of these recordings demonstrate the same type of phe-
nomenon seen in Figures 20.1 and 20.2: the correlation structure to some degree
depends on the stimulus conditions.
As mentioned, the alternating stimulus cycle means that we force the change
in correlation structure on a time scale of 0.5 to 1 sec, even though many repe-
titions of the cycle are needed to make the necessary measurements. Other mea-
surements, which we have not described here, show that modulations of the cor-
relation structure can take place on an even faster time scale of milliseconds
(Aertsen et al., 1987; Aertsen et al, 1989).
Thus the results of experimental observation of small neuronal assemblies
show that the correlation structure of firing is a dynamic, rapidly fluctuating
quantity. This means that among real neurons, the effective connectivity (which
corresponds to the W/j of the modelers) behaves quite differently from the usual
assumption of a slowly evolving constant.

INTERPRETATION AND POSSIBLE MECHANISMS

A dynamic correlation structure among a group of observed neurons can be

explained by various known mechanisms that either operate at the level of syn-
aptic properties or involve properties of the neuronal network. For example, the
effective connectivity between two neurons could be rapidly modulated by a pre-
synaptic input from another, unobserved neuron. Alternatively, the synaptic

FIGURE 20.2. For the same data as in Figure 20.1, projections from the 8-space to a convenient
plane. Top and bottom row represent, respectively, condensations during the two stimulus con-
ditions. Left column are the (identical) projected initial particle configurations; right column are
the projected final configurations. The straight lines in the right column join the initial and final
projected positions of the particles; the actual projected trajectory is not shown.
388 REPRESENTATIONS: BEYOND THE SINGLE CELL

path between two neurons could go through an unobserved interneuron which

itself has other input and hence modulates the effective connectivity between
the two observed neurons.
At the network level, we may invoke the following type of explanation. Sup-
pose we consider the observed neurons to be embedded in a large mass of inter-
connected neural ussue. Presumably stimuli that have different parametric val-
ues enter this large neural mass in different ways, probably at somewhat different
locations. (This would correspond to the mapping of the sensory periphery onto
the structure under study.) Thus each different stimulus sets up a disturbance
that takes a somewhat different path through the neural mass and hence crosses
the observed neurons in a somewhat different way, apparently causing the cor-
relation structure of firing to vary.
Although all such explanations are plausible in a qualitative way, their quan-
titative applicability remains to be investigated, most simply by computer sim-
ulation methods. Such work is under way in several laboratories. An important
(and largely unknown) parameter that is involved in such work is the amount
of "spontaneous" (other) activity relative to the stimulus-related activity. In the
limiting case of extremely small stimulus-related activity in a massive back-
ground of other network activity we would expect minimal modulation of the
correlation structure by the stimulus. On the other hand, even in such a situa-
tion, modulation maybe imposed by other sources of input to the system. Since
the time structure of such unknown input is not available, there is a serious
measurement problem in that the usual averaging methods cannot be used.

SIGNIFICANCE FOR MODELING

As noted previously, most current models for networks of neurons do not explic-
itly contain Wti with dynamic properties on the time scale demonstrated in the
physiological measurements. The one exception is found in the work of von der
Malsburg (1987), where Hebbian rules modulate synaptic efficacies on very short
time scales and produce some useful consequences.
Is it important to adjust other network modeling so as to mimic the observed
rapid modulation of effective connectivity? One possibility is that the rapid
modulation is a stimulus-associated statistical variation that has no real physi-
ological significance and is essentially a noiselike phenomenon. Another possi-
bility is that the network explanation of the rapid modulation is adequate, and
that models with essentially fixed W(j nevertheless exhibit rapid stimulus-asso-
ciated modulation of correlation structure. In both these cases, dealing with
average Wij which vary only slowly, would seem to be justified, and develop-
ment of network models can proceed along its current path. On the other hand,
since real neuronal networks in the brain do show rapid modulation of corre-
lation structure, it would perhaps be both safer and quite appropriate to see what
changes occur in a given model if such properties are explicitly installed. The
process of modeling always involves selection of the subset of real-world prop-
 Interactions Within Neuronal Assemblies: Theory and Experiment 389

erties that are to be mimicked; it is important to make sure that such selection
does not drastically affect the modeling results.

Acknowledgment
This work was supported by the System Development Foundation, SDF 0013,
and by the Office of Naval Research, N00014-83-K-0387 and N00014-87-K-
0766.

REFERENCES

Aertsen, A. M. H. J., Bonhoeffer, T., & Kruger, J. (1987). Coherent activity in neuronal popu-
lations: Analysis and interpretation. In E. R. Caianiello (Ed.), Physics of cognitive
processes (pp. 1-34). Singapore: World Scientific Publishing.
Aertsen, A. M. H. J., Gerstein, G. L., Habib, M. K., & Palm, G. (1989). Dynamics of neuronal
firing correlation: Modulation of "Effective Connectivity." Journal of Neurophysiol-
ogy, 61, 900-917.
Bloom, M. J., & Gerstein, G. L. (1986). Dynamic changes in cortical connectivity during sound
localization. Society for Neuroscience Abstracts, 12, 1274.
Espinosa, I. E., & Gerstein, G. L. (1985). Correlation among spike trains in cat's auditory cortex
during presentation of three-tone sequences. Society for Neuroscience Abstracts, 11,
249.
Espinosa, I. E., & Gerstein, G. L. (1988). Cortical auditory neuron interactions during presen-
tation of 3-tone sequences: Effective connectivity. Brain Research, 450, 39-50.
Farley, B. G. (1964). The use of computer technics in neural research. In R. F. Reiss (Ed.),
Neural theory and modeling (pp. 43-72). Stanford, Calif.: Stanford University Press.
Farley, B. G., & Clarke, W. A. (1961). Activity in networks of neuronlike elements. In C. Cherry
(Ed.), Informaton theory (Fourth London Symposium). London: Butterworth.
Gerstein, G. L., & Aertsen, A. M. H. J. (1985). Representation of cooperative firing activity
among simultaneously recorded neurons. Journal of Neurophysiology, 54, 1513-
1528.
Gerstein, G. L., Bloom, M., Espinosa, I., Evanczuk, S., Turner, M. (1983). Design of a labora-
tory for multi-neuron studies. IEEE Transactions on Systems, Man and Cybernetics,
SMC-13, 668-676.
Gerstein, G. L., Perkel, D. H., & Dayhoff, J. E. (1985). Cooperative firing activity in simulta-
neously recorded populations of neurons: Detection and measurement. Journal of
Neuroscience, 5, 881-889.
Glaser, E. M., & Ruchkin, D. S. (1976). Principles of microbiological signal analysis. New York:
Academic Press.
Hebb, D. O. (1949). The organization of behavior. New York: Wiley.
Kruger, J. (1983). Simultaneous individual recordings from many cerebral neurons: Techniques
and results. Review of Physiology, Biochemistry, and Pharmacology, 98, 177-233.
Palm, G. (in press). Local rules for synaptic modification in neural networks. In E. Schwartz
(Ed.), Computational neuroscience. Cambridge, Mass.: MIT Press.
Rochester, N., Holland, J. H., Haibt, L. H., & Duda, W. L. (1956). Tests on a cell assembly
theory of the action of the brain, using a large digital computer. IRE Transactions on
Information Theory, IT-2, 80-93.
von der Malsburg, C. (1987). Synaptic plasticity as basis of brain organization. In J. P. Chan-
geux & M. Konishi (Eds.), The neural and molecular bases of learning (pp. 411-432).
New York: Wiley.
21
Neural Networks and Networks
of Neurons
G A R Y LYNCH
J O H N LARSON
DOMINIQUE MULLER
RICHARD GRANGER

Neural network models hold the promise of serving as a bridge, or translator,

between the more familiar worlds of physiology and behavior. Behavior emerges
from the collective activities of brain cells linked together in spatiotemporal pat-
terns, and those activities are the result of the anatomical and physiological
properties of individual neurons. Unfortunately, even a reasonably complete
description of cellular neurobiology does not necessarily lead to precise predic-
tions about the types of interactions that might occur between neurons in a par-
ticular brain region—let alone predictions about the coherent spatiotemporal
patterns that presumably underlie visible behavior. It is easy to understand why
this should be the case. Temporal interactions between any large group of
objects are inherently difficult subjects for analysis; one need go no further than
classical physics (e.g., celestial and statistical mechanics) to see this. Moreover,
the influences that brain cells have over each other are by no means simple (i.e.,
synaptic responses are complex) and are transmitted via complicated architec-
tures with any number of specialized and often subtle design features. In all, the
objects of interest (neurons) are of several types, possess a host of relevant char-
acteristics, and sense and exert a variety of influences over convoluted net-
works.
Obviously, a great deal of simplification is required if models of brain net-
works are to be developed. One of the successes of neural network modeling has
been the demonstration that much can be learned about the aggregate activity
of cells and its relationship to behavior using only very simple physiological
rules and circuit designs. Moreover, in some cases the resulting models are trac-
table subjects for mathematical analysis, a point that is sometimes not appreci-

390
 Neural Networks and Networks of Neurons 391

ated by behavioral neuroscientists but that is of fundamental importance in

efforts to develop theories of network behavior. By using mathematical tech-
niques, it is possible to show that a given model will always achieve certain end
points and thus that its behavior is not unique to a particular computer simu-
lation (see Cohen & Grossberg, 1983; Hopfield, 1982; Kohonen, 1984; Rumel-
hart, Hinton, & Williams, 1986, for examples). Equally important, mathe-
matical analyses can provide explanations and make predictions that are not
likely to arise from experimental (in the present case, simulation) work (Kuhn,
1977).
Network models vary considerably, depending on the simplifying assump-
tions made and the goals of the enterprise. In some cases they are used to explore
the consequences of hypotheses about the general operating principles of brain
circuitries. Hopfield (1982), for example, investigates the possibility that the
interactions between neurons converge to stable "states" (i.e., spatial or spatio-
temporal patterns of cell activity) in much the way certain physical systems (e.g.,
gases, magnetic domains) do; these states would then form the representations
of external objects. Gelperin and colleagues (Cook, Delaney, & Gelperin, 1985)
found that a network of this type when used as a model of an invertebrate ner-
vous system produces behaviors similar to those observed in experiments.
Kohonen (1984; Chapter 16, this volume) notes that brain circuitry often resem-
bles an input-output matrix and investigates the idea that such networks use a
kind of matrix algebra to produce coherent activity and representations.
Another increasingly popular approach uses multilayered networks and a par-
ticular learning algorithm (backpropagation of error) to address a remarkable
variety of behavioral problems (Rumelhart et al., 1986; Sejnowski & Tesauro,
Chapter 17, this volume). Much of the appeal of this approach lies in the learn-
ing rule, since it can be shown mathematically to minimize errors (across trials)
between the input cues to be encoded and the representations assigned to them
(Widrow & Hoff, 1960). Of evident utility, this became more intriguing theoret-
ically when it was understood (Sutton & Barto, 1981) that the algorithm corre-
sponds to rules for classical conditioning defined by behavioral researchers (Res-
corla & Wagner, 1972).
The success of neural networks in producing behaviors, and even complex
behaviors, that resemble those exhibited by humans and other animals suggests
that the models will ultimately provide some understanding of how aspects of
psychology emerge from aspects of neurobiology. Clearly what is needed are
experimentally testable predictions about neurons and about behavior. The
models are concerned with spatiotemporal patterns in arrays of cells; unfortu-
nately, using current technologies, these are extremely difficult to measure in
brains. Indeed, our ignorance on just this point provides some of the impetus
for modeling. In general the models do not yield detailed predictions about syn-
aptic events, membrane properties, and the like, on the one hand, and acquisi-
tion rates, memory interference effects, and so forth, on the other. A bridge is
being built but we must wait to see if and to what degree it connects the two
worlds of behavioral neuroscience.
392 REPRESENTATIONS: BEYOND THE SINGLE CELL

NEURONS IN NETWORKS

Another kind of modeling is concerned with the activities of individual (real)

neurons and the question of how these arise from the brain networks within
which the cells are found. Rather than beginning with hypotheses about the rules
governing aggregate behavior of cells, this approach is concerned with finding
local rules that produce effects observed in physiological experiments on single
cells. Having identified such rules, one would hope that these will serve as the
basis of more general theories of how cortex encodes experience. The emphasis
in most models thus far has been on development, perhaps because many
striking properties of cortical cells appear to be shaped in part by early expe-
rience.
Physiologists studying the visual system have provided a wealth of data on
how experimental manipulations carried out during critical periods of devel-
opment produce striking and detailed changes in the responses of cortical neu-
rons to peripheral cues (see Singer, Chapter 10, this volume). Several groups are
now building models with which to investigate various neuronal rules that could
produce such effects (Cooper, Bear, Ebner, & Scofield, Chapter 15, this volume;
Linsker, 1986, 1988; Singer, Chapter 10, this volume). Cooper and co-workers
(Bear, Cooper, & Ebner, 1987) have had remarkable success in reproducing in
simulations (and with mathematical treatments) some of the more prominent
experimental effects observed in cat visual cortex; equally important, they used
the model to make testable predictions, which in at least one case were con-
firmed. Their work also leads to unexpected but testable ideas about plasticity
in neurons (e.g., the phi factor). In some senses then this work has achieved the
two-way outlook of behavioral neuroscience mentioned earlier: it speaks to
behavior (albeit at the level of behaviorally relevant activities of single cells) and
to neurobiological properties. Linsker (1986) has also produced a model in
which salient aspects of visual cortical physiology emerge using a development
rule that is similar to those employed in certain of the neural network
models mentioned earlier. The possibility that the hypothetical rules (back-
propagation) used in neural networks might explain specific developmental
effects observed at the single-cell level was explicitly tested by Zipser and
Andersen (1988) with some success (see Andersen & Zipser, Chapter 13, this
volume).
Yet another approach to exploring how various physiological rules in net-
works might produce effects observed in individual neurons and in the aggregate
behavior of the network would be to start with physiologically defined rules.
That is, rather than using hypothetical learning algorithms or deducing them
from behavior or cellular end points, one could begin with experimental obser-
vations on synaptic plasticity, identify ways in which these are relevant to in
situ events, and then ask what types of effects they produce when implemented
in a network model. We have been carrying out a program of this type and the
following sections briefly review some of our results.
 Neural Networks and Networks of Neurons 393

LONG-TERM POTENTIATION AS A LEARNING RULE

Long-term potentiation (LTP) is in many ways an obvious candidate for the

substrates of memory: it is an extremely persistent (Barnes, 1979; Bliss & Gard-
ner-Medwin, 1973; Staubli & Lynch, 1987) synapse-specific (Andersen, Sund-
berg, Sveen, & Wigstrom 1977; Dunwiddie & Lynch, 1978; Lynch, Dunwiddie,
& Gribkoff, 1977) increase in the size of excitatory postsynaptic responses elic-
ited by brief episodes of high-frequency stimulation. Moreover, recent experi-
ments demonstrate that drugs that block LTP also block certain forms of learn-
ing (Morris, Anderson, Lynch, & Baudry, 1986; Staubli, Baudry, & Lynch, 1984;
Staubli, Faraday, & Lynch, 1985; Staubli, Larson, Baudry, Thibault, & Lynch,
1988). To use LTP as a learning rule we need to specify in some detail the rela-
tionship between physiological patterns occurring in situ and the induction of
potentiation. Since LTP has been most fully characterized in hippocampus, this
is a reasonable place to begin such an effort.
The hippocampal EEG assumes a characteristic 4- to 7-Hz rhythm known as
theta whenever rats are engaged in exploratory behavior (Vanderwolf, 1969).
The pyramidal cells in the structure fire in very short bursts of 3 to 5 spikes with
the bursts often occurring during a particular phase of theta (Bland, Andersen,
Ganes, & Sveen, 1980; Fox & Ranck, 1981; Ranck, 1973). The question then
becomes one of whether a relationship exists between these characteristics (theta
and bursting) and the induction of LTP. Our work indicates that this is the case.
Short bursts (4 pulses at 100 Hz) delivered to the inputs to field CA1 elicit robust
LTP when the bursts are spaced apart by 200 msec; intervals that are much
shorter (< 100 msec) or longer than 200 msec produce progressively less LTP
(Larson, Wong, & Lynch, 1986). It should be noted here that the potentiation
induced in this fashion is extremely stable and can persist without evident dec-
rement for weeks (Staubli & Lynch, 1987).
Convergence of different inputs on target cells is an essential feature of net-
works; therefore it is essential to know how convergence interacts with theta
bursting and LTP. To answer this, we stimulated one collection of inputs to a
target cell, waited 200 msec, and then activated a completely separate set of
inputs to the same neuron. Five to ten such pairs were given with a delay of
several seconds between each pair. The second, delayed input exhibited robust
LTP, whereas the first was not detectably changed (Larson & Lynch, 1986). As
a rule, then, we assume that activation of one afferent "primes" its target cells
so that other inputs can change their synapses if stimulated about 200 msec later.
What about convergence within one phase of theta as opposed to convergence
between waves? The experiments just described used synchronous stimulation
of two large groups of axons because previous work (McNaughton, Douglas, &
Goddard, 1978) indicated that LTP induction requires the cooperative action of
many synapses. However, in situ the firing of different cells will not be so per-
fectly synchronized. To explore this issue, we carried out experiments in which
three sets of afferents terminating in the same dendritic region were stimulated
394 REPRESENTATIONS: BEYOND THE SINGLE CELL

sequentially with their bursts only partially overlapping. Each input was first
given a priming pulse; the three inputs then received a four-pulse burst (100 Hz)
at a delay of 180 msec (first input), 200 msec (second input), and 220 msec (third
input). Thus the burst to the second input overlapped that to both the first and
third, while the first and third bursts did not overlap at all. The degree of LTP
induced was greatest at synapses activated by the first burst, intermediate for the
second, and least for the third (Larson & Lynch, 1989). It is interesting to note
that this within-phase convergence rule is the opposite of the between-phase
rule: with convergence during one phase the later arrival facilitates LTP in the
early arrival, while across two phases the early arrival facilitates LTP in the later
arriving input.
Having identified these LTP rules we can ask why they exist. This is a ques-
tion of immediate importance to our understanding of the neurobiological fac-
tors that initiate synaptic change, but it is also of considerable significance for
attempts to develop rules for networks. That is, a description of the physiological
processes underlying the between- and within-phase effects is necessary for the
development of rules that generalize beyond these particular activity patterns.
Here we will only summarize the work on this issue; more complete descriptions
can be found elsewhere (Larson & Lynch, 1986, 1988; Lynch & Larson, 1988;
Lynch, Larson, Staubli, & Baudry, 1987).
The central feature of the "priming" effect that drives the between-phase rule
appears to be a transient refractory period for feedforward IPSPs (Larson &
Lynch, 1986; see also McCarren & Alger, 1985). Stimulated afferents in hippo-
campus initiate EPSPs in pyramidal cells but also trigger interneurons that pro-
duce IPSPs in the same cells (Alger & Nicoll, 1982). This IPSP lasts about 50 to
100 msec and then appears to become refractory (or suppressed) for about 200
to 500 msec. The theta interval (200 msec) defines a time after which the initial
IPSP has dissipated and is difficult to reintroduce. EPSPs are considerably pro-
longed in the absence of the IPSP and summate within a burst to a much greater
degree than is true for control conditions (Larson & Lynch, 1988). In essence,
priming considerably amplifies the depolarization produced by a bursting input.
Under some circumstances the amplified response is adequate to trigger an
NMDA receptor-mediated response component as shown in experiments using
selective antagonists of the receptor (Larson & Lynch, 1988). If this NMDA-
related potential is blocked, LTP does not occur, a result that accords with
previous findings (Collingridge, Kehl, & McLennan, 1983; Harris, Ganong, &
Cotman, 1984; Morris et al., 1986). The NMDA receptor channel in its high-
conductance state (i.e., under extreme depolarization) admits calcium into the
target cell (MacDermott, Mayer, Westbrook, Smith, & Barker, 1986; Mayer &
Westbrook, 1987) and calcium appears to be the trigger for LTP (Lynch, Larson,
Kelso, Barrionuevo, & Schottler, 1983). The mechanisms whereby calcium pro-
duces LTP are controversial but are probably linked to the anatomical changes
that correlate with LTP (see Lynch, Bodsch, & Baudry, 1988, for a discussion of
intermediary mechanisms for LTP).
The refractoriness of the IPSP, temporal summation of responses during
 Neural Networks and Networks of Neurons 395

bursts, and the voltage threshold of the NMDA-mediated response are features
that can be simulated and thereby provide some generality for a network rule.
Work on the causes of the within-phase rule is still in progress, but the causes
appear to involve yet another peculiar feature of the NMDA receptor: its chan-
nel is slow to respond but once activated tends to remain open (Jahr & Stevens,
1987). Later depolarizing inputs then may serve to sustain NMDA-dependent
events at synapses that were stimulated tens of milliseconds earlier (Larson &
Lynch, 1989). Additional experiments should serve to specify the relevant
events to a degree that is sufficient for incorporation into a plasticity rule for
network simulations.

A NOTE ON THE SUBSTRATES OF LONG-TERM POTENTIATION

LTP can be modeled as a simple increase in synaptic strength; that is, activation
of a given synapse produces a larger than normal voltage change in the target
cell. Although this seems appropriate, it would be useful to know precisely what
type of semipermanent change is in fact responsible for physiological
potentiation.
Electron microscopic studies indicate that LTP is correlated with changes in
the morphology of dendritic spines and a small increase in the numbers of cer-
tain types of contacts (Chang & Greenough, 1984; Lee, Oliver, Schottler, &
Lynch, 1981; Lee, Schottler, Oliver, & Lynch, 1980; Wenzel & Matthies, 1985).
The possibility exists that these effects are different degrees of a single process,
that is, different degrees of spine transformation. It is reasonable to assume that
the observed morphological correlates of LTP are responsible for synaptic facil-
itation, but there is no direct evidence on this point and secondary effects of
structural reorganization may be involved. The recent discovery that LTP
increases one component of postsynaptic response (that mediated by non-
NMDA receptors) without affecting a second component (the NMDA-mediated
component) considerably limits the possible locales for potentiation. This is all
the more likely since paired-pulse facilitation (a very transient form of potenti-
ation due to increased transmitter release) increases both components of the
postsynaptic potential (Muller & Lynch, 1988). It is thus very unlikely that LTP
is caused by a simple increase in transmitter release or a simple increase in syn-
aptic contacts. More likely explanations are changes in the biophysics of the
spine-dendrite coupling or an increased number of synapses that are impover-
ished in NMDA receptors.

IMPLEMENTING LONG-TERM POTENTIATION RULES

INTO A SIMPLE CORTICAL NETWORK

The hippocampus has proven extremely useful in developing an LTP-based

learning rule, but it is not appropriate as a model for a simulation to test the
396 REPRESENTATIONS: BEYOND THE SINGLE CELL

consequences of that rule (see Lynch, Granger, Larson, & Baudry, 1988, for a
discussion). Instead, we chose to use the superficial layers of the olfactory (piri-
form-entorhinal) cortex for the purpose. This region is disynaptically connected
to peripheral receptors for odors and thus can be assumed to play a primary role
in virtually any olfactory-related behavior (see Lynch, 1986). Accordingly, it is
not unreasonable to assume that events in the cortex will have detectable effects
on behavior; this would greatly simplify the task of testing predictions from the
simulation. Additionally, layers I and II of the cortex are comparatively simple
and their anatomy has been defined in great detail, two essential features for a
simulation (Lynch, 1986). The simulations we have carried out used 100 to 200
cells, local inhibitory interneurons, and sparse connectivity between the lateral
olfactory tract (the extrinsic input to the cortex) and associational feedback sys-
tems with the target cells (Granger, Ambros-Ingerson, & Lynch, 1989; Lynch &
Granger, in press). Physiological rules for the model included many of the LTP-
related features described above as well as three forms of hyperpolarization: the
feedforward IPSP, the late-hyperpolarizing potential (Alger, 1984), and a cell-
specific after-hyperpolarization (Hotson & Prince, 1980). These are all promi-
nent aspects of hippocampal physiology and have been identified in the piriform
cortex as well (Haberly, 1985). Later versions of the simulation incorporated
probabilistic transmission.
A particularly difficult problem for models of cortical networks is how to sim-
ulate inputs that might occur during behavior. In the present case we would
particularly like to know if the theta rhythm patterns used in the studies on hip-
pocampus would constitute a "meaningful" signal to the olfactory cortex. There
is reason to think that this might be the case since rats sniff at rates close to theta
and indeed activity in hippocampus is correlated with the sniffs. Accordingly, a
series of "electrical odor" experiments were conducted in which stimulation of
the lateral olfactory tract (LOT) was used in place of natural odors in an olfac-
tory discrimination task (Roman, Staubli, & Lynch, 1987). The animals were
first trained on a series of discriminations using novel pairs of odors over several
days. Then tests were made to determine if they could learn to discriminate an
electrical odor from a natural odor or between two electrical odors (i.e., between
stimulation at two sites in the LOT). Under these conditions, theta burst stim-
ulation to the LOT was treated by the animals in a manner not obviously dif-
ferent from real odors. Interestingly enough, stimulation used in the learning
situation produced LTP in the LOT-piriform synapses but did not do so when
used with naive rats (Roman et al., 1987). It might be noted that others have
shown that induction of LTP in the perforant path-hippocampal system is
dependent on the arousal state of rats (Jones Leonard, McNaughton, & Barnes,
1987).
In any event, LTP does appear in the piriform cortex during learning (at least
of electrical odors) and theta bursting in the LOT is "interpretable" by the cortex
as a meaningful input signal. Accordingly, the simulation receives as inputs
short bursts of activity with the bursts separated by (simulated) 200 msec; each
burst might be thought of as the input resulting from one of a series of "sniffs."
Interesting results emerged when the simulation was trained using LTP rules
 Neural Networks and Networks of Neurons 397

on a series of "odors" (an odor was represented as a specific 20% subgroup of

the LOT lines) that overlapped or shared components among themselves (i.e., a
given odor might share some percentage of its chemistry, and thus its input lines,
with other members of a set of odors). After the model had learned several indi-
vidual cues with such overlap, the spatial pattern (of activity) it generated on
the first sniff became differentiated from the response pattern it emitted on later
sniffs. Specifically, the first response was the same or nearly the same for all
members of the odor group, whereas its later responses were specific to the cue
being presented. In one experiment we presented the simulation with 22 sepa-
rate cues; it produced 5 category (first sniff) patterns and 22 distinct third sniff
responses. The network thus extracts information about similarity while at the
same time emphasizing differences and then "reads out" this stored information
by using successive sampling. Note also that the representations are stored hier-
archically in that the category response always appears prior to the individuated
response pattern. It is also important to emphasize that the LTP-based rules
promote acuity; even cues that overlap by 90% generate network outputs after
learning which have fewer than 50% of their cells in common.
These results indicate that empirically derived local (i.e., synaptic) rules for
modifying contacts and global rhythmic rules do yield coherent aggregate behav-
ior when implemented in a greatly simplified model of a cortical layer. The role
that learning plays is somewhat different from that typically assigned to it but
the phenomena we have observed are prominent in human recognition mem-
ory. People do partition their cue world into categories and do exhibit a hierar-
chical encoding (i.e., group-individual) not unlike that found in the simulation.
A more critical evaluation of the simulation is being made by recording the
activity of individual neurons in the olfactory cortex while rats are learning odor
discriminations and then comparing the results with the predictions of the
model. Several important features of the simulation have been confirmed. Cod-
ing appears to be very sparse in that only a small percentage of neurons react to
any given odor and activity does seem to be time locked, probably due to the
inhalation cycle and theta. We have also observed effects of learning on cell
activity. A far more rigorous testing of the model will require large numbers of
cues that can be subdivided into natural categories. Cue partitioning or catego-
rization appears not to have been studied as such in rats. It is perhaps a measure
of success of the simulation work that it emphasizes a form of learning that has
received so little attention in animal psychology. Tests of the very specific pre-
dictions of the model (e.g., different patterns of activity arise with different
sniffs) under these conditions will provide a rigorous evaluation of its validity
as a description of how specific physiological characteristics produce aggregate
effects in networks of neurons.

REFERENCES

Alger, B. E. (1984). Characteristics of a slow hyperpolarizing synaptic potential in rat hippo-

campal pyramidal cells in vitro. Journal of Neurophysiology, 52, 892-910.
398 REPRESENTATIONS: BEYOND THE SINGLE CELL

Alger, B. E., & Nicoll, R. A. (1982). Feed-forward dendritic inhibition in rat hippocampal
pyramidal neurones studied in vitro. Journal of Physiology (London), 328, 105-123.
Andersen, P., Sundberg, S. H., Sveen, O., & Wigstrom, H. (1977). Specific long-lasting poten-
tiation of synaptic transmission in hippocampal slices. Nature, 226, 736-737.
Barnes, C. A. (1979). Memory deficits associated with senescence: A neurophysiological and
behavioral study in the rat. Journal of Comparative Physiology and Psychology, 93,
74-104.
Bear, M., Cooper, L., & Ebner, F. (1987). The physiological basis for a theory of synapse mod-
ification. Science, 237, 42-48.
Bland, B. H., Andersen, P., Ganes, T., & Sveen, O. (1980). Automated analysis of rhythmicity
of physiologically identified hippocampal formation neurons. Experimental Brain
Research, 38, 205-219.
Bliss, T. V. P., & Gardner-Medwin, A. R. (1973). Long-lasting potentiation of synaptic trans-
mission in the dentate area of the unanesthetized rabbit following stimulation of the
perforant path. Journal of Physiology (London), 232, 357-374.
Chang, F. L. F., & Greenough, W. T. (1984). Transient and enduring morphological correlates
of synaptic activity and efficacy change in the rat hippocampal slice. Brain Research,
309, 35-46.
Cohen, M. A., & Grossberg, S. (1983). Absolute stability of global pattern formation and par-
allel memory storage by competitive neural networks. IEEE Transactions on Sys-
tems, Man, and Cybernetics, 13, 815-825.
Collingridge, G. L., Kehl, S. J., & McLennan, H. (1983). Excitatory amino acids in synaptic
transmission in the Schaifer collateral-commissural pathway of the rat hippocampus.
Journal of Physiology (London), 334, 33-46.
Cooke, I., Delaney, K., & Gelperin, A. (1985). Complex computation in a small neural network.
In N. M. Weinberger, J. L. McGaugh, & G. Lynch (Eds.), Memory systems of the
brain (pp. 173-192) New York: Guilford.
Dunwiddie, T., & Lynch, G. (1978). Long-term potentiation and depression of synaptic
responses in the rat hippocampus: Localization and frequency dependency. Journal
of Physiology (London), 276, 353-367.
Fox, S. E., & Ranck, J. B., Jr. (1981). Electrophysiological characteristics of hippocampal com-
plex-spike cells and theta cells. Experimental Brain Research, 41, 399-410.
Granger, R., Ambros-Ingerson, J., & Lynch, G. (1989). Derivation of encoding characteristics
of layer II cerebral cortex. Journal of Cognitive Neuroscience, 1, 61-87.
Haberly, L. B. (1985). Neuronal circuitry in olfactory cortex: Anatomy and functional impli-
cations. Chemical Senses, 10, 219-238.
Harris, E. W., Ganong, A. H., & Cotman, C. W. (1984). Long-term potentiation in the hippo-
campus involves activation of Ar-methyl-D-aspartate receptors. Brain Research, 323,
132-137.
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective compu-
tational abilities. Proceedings of the National Academy of Sciences USA, 79, 2554-
2558.
Hotson, J. R., & Prince, D. A. (1980). A Ca++-activated hyperpolarization follows repetitive
firing in hippocampal neurons. Journal of Neurophysiology, 43, 409-419.
Jahr, C. E., & Stevens, C. F. (1987). Glutamate activates multiple single channel conductances
in hippocampal neurons. Nature, 325, 522-525.
Jones Leonard, B., McNaughton, B. L., & Barnes, C. A. (1987). Suppression of hippocampal
synaptic plasticity during slow-wave sleep. Brain Research, 425, 174-177.
Kohonen, T. (1984). Self-organization and associative memory. Berlin: Springer-Verlag.
Kuhn, T. (1977). The essential tension. Chicago: University of Chicago Press.
Larson, J., & Lynch, G. (1986). Induction of synaptic potentiation in hippocampus by pat-
terned stimulation involves two events. Science, 232, 985-988.
Larson, J., & Lynch, G. (1988). Role of N-methyl-D-aspartate receptors in the induction of syn-
aptic potentiation by burst stimulation patterned after the hippocampal theta
rhythm. Brain Research, 441, 111-118.
 Neural Networks and Networks of Neurons 399

Larson, J., & Lynch, G. (1989). Theta pattern stimulation and the induction of LTP: The
sequence in which synapses are stimulated determines the degree to which they
potentiate. Brain Research, 489, 49-58.
Larson, J., Wong, D., & Lynch, G. (1986). Patterned stimulation at the theta frequency is opti-
mal for induction of hippocampal long-term potentiation. Brain Research, 368, 347-
350.
Lee, K., Oliver, M., Schottler, F., & Lynch. G. (1981). Electron microscopic studies of brain
slices: The effects of high frequency stimulation on dendritic ultrastructure. In G.
Kerkut & H. V. Wheal (Eds.), Electrical activity in isolated mammalian C.N.S.
preparations, (pp. 189-212). New York: Academic Press.
Lee, K., Schottler, F., Oliver, M., & Lynch, G. (1980). Brief bursts of high-frequency stimula-
tion produce two types of structural change in rat hippocampus. Journal of Neuro-
physiology, 44, 247-258.
Linsker, R. (1986). From basic network principles to neural architecture: Emergence of spatial-
opponent cells. Proceedings of the National Academy of Sciences USA, 83, 7508-
7512.
Linsker, R. (1988, March). Self-organization in a perceptual network. IEEE Computer, pp. 105-
117.
Lynch, G. (1986). Synapses, circuits, and the beginnings of memory. Cambridge, Mass.: MIT
Press.
Lynch, G., Bodsch, W., & Baudry, M. (1988). Cytoskeletal proteins and the regulation of syn-
aptic structure. In R. J. Lasek & M. M. Black (Eds.), Intrinsic determinants of neu-
ronalform and function (pp. 217-242). New York: Alan R. Liss.
Lynch, G. S., Dunwiddie, T. V., & Gribkoff, V. (1977). Heterosynaptic depression: A postsyn-
aptic correlate of long-term potentiation. Nature, 266,737-739.
Lynch, G., & Granger, R. (in press). Simulation and analysis of a simple cortical network. Psy-
chology of Learning and Motivation.
Lynch, G., Granger, R., Larson, J., & Baudry, M. (1988). Cortical encoding of memory:
Hypotheses derived from analysis and simulation of physiological learning rules in
anatomical structures. In L. Nadel, L. A. Cooper, P. Culicover, & R.M. Harnish
(Eds.), Neural connections and mental computation (pp. 247-289). Cambridge,
Mass.: Bradford Books/MIT Press.
Lynch, G., & Larson, J. (1988). Rhythmic activity, synaptic changes, and the "how and what"
of memory storage in simple cortical networks. In J. L. Davis, R. W. Newburgh, &
R. E. J. Wegman (Eds.), Brain structure, learning, and memory (pp.33-68). Boulder,
Colo.: Westview Press.
Lynch, G., Larson, J., Kelso, S., Barrioneuvo, G., & Schottler, F. (1983). Intracellular injection
of EGTA block the induction of hippocampal long-term potentiation. Nature, 305,
719-721.
Lynch, G., Larson, J., Staubli, U., & Baudry, M. (1987). New perspectives on the physiology,
pharmacology, and chemistry of memory. Drug Development Research, 10, 295-315.
MacDermott, A. B., Mayer, M. L., Westbrook, G. L., Smith, S. J., & Barker, J. L. (1986).
NMDA-receptor activation increases cytoplasmic calcium concentration in cultured
spinal cord neurones. Nature, 321, 519-522.
Mayer, M. L., & Westbrook, G. L. (1987). Permeation and block of Ar-rnethyl-n-aspartic acid
receptors channels by divalent cations in mouse cultured central neurones. Journal
of Physiology (London), 394, 501-527.
McCarren, M., & Alger, B. E. (1985). Use-dependent depression of IPSPs in rat hippocampal
pyramidal cells in vitro. Journal of Neurophysiology, 53, 557-571.
McNaughton, B. L., Douglas, R. M., & Goddard, G. V. (1978). Synaptic enhancement in fascia
dentata: Cooperativity among coactive afferents. Brain Research, 157, 277-293.
Morris, R. G. M., Anderson, E., Lynch, G., & Baudry, M. (1986). Selective impairment of learn-
ing and blockade of long-term potentiation by an N-methyl-D-aspartate receptor
antagonist, AP-5. Nature, 319, 774-776.
Muller, D., & Lynch, G. (1988). Long-term potentiation differentially affects two components
400 REPRESENTATIONS: BEYOND THE SINGLE CELL

of synaptic responses in hippocampus. Proceedings of the National Academy of Sci-

ences USA, 85, 9346-9350.
Ranck, J. B., Jr. (1973). Studies on single neurons in dorsal hippocampal formation and septum
in unrestrained rats. Experimental Neurology, 41, 462-531.
Rescorla, R., & Wagner, A. R. (1972). A Theory of Pavlovian conditioning: Variations in the
effectiveness of reinforcement and nonreinforcement. In A. H. Black & W. F. Prokasy
(Eds.), Classical conditioning (pp. 64-99). New York: Appleton-Century-Crofts.
Roman, F., Staubli, U., & Lynch, G. (1987). Evidence for synaptic potentiation in a cortical
network during learning. Brain Research, 418, 221-226.
Rumelhart, D., Hinton, G., & Williams, R. (1986). Learning internal representations by error
propagation. In D. Rumelhart & J. McClelland (Eds.), Parallel distributed processing.
Cambridge, Mass.: MIT Press.
Staubli, U., Baudry, M., & Lynch, G. (1984). Leupeptin, a thiol proteinase inhibitor, causes a
selective impairment of spatial maze performance in rats. Behavioral and Neural
Biology, 40, 58-69.
Staubli, U., Faraday, R., & Lynch, G. (1985). Pharmacological dissociation of memory: Ani-
somycin, a protein synthesis inhibitor, and leupeptin, a protease inhibitor, block dif-
ferent learning tasks. Behavioral and Neural Biology, 43, 287-297.
Staubli, U., Larson, J., Baudry, M., Thibault, O., & Lynch, G. (1988). Chronic administration
of a thiol-proteinase inhibitor blocks long-term potentiation of synaptic responses.
Brain Research, 444, 153-158.
Staubli, U., & Lynch, G. (1987). Stable hippocampal long-term potentiation elecited by "theta"
pattern stimulation. Brain Research, 435, 227-234.
Sutton, R. S., & Barto, A. G. (1981). Toward a modern theory of adaptive networks. Expecta-
tion and prediction. Psychological Review, 38, 135-171.
Vanderwolf, C. H. (1969). Hippocampal electrical activity and voluntary movement in the rat.
Electroencephalography and Clinical Neurophysiology, 26, 407-418.
Wenzel, J. & Matthies, H. (1985). Morphological changes in the hippocampal formation
accompanying memory formation and long-term potentiation. In N. Weinberger, J.
McGaugh, & G. Lynch (Eds.), Memory systems of the brain (pp. 150-170). New
York: Guilford.
Widrow, G. & Hoff, M. E. (1960). Adaptive switching circuits. Institute of Radio Engineers,
Western Electronic Show Convention Record (Part 4), 96-104.
Zipser, D. & Andersen, R. A. (1988). A back-propagation programmed network that simulates
response properties of a subset of posterior parietal neurons. Nature, 331, 679-684.
Index

Ablation studies, 7-8 versus hippocampus, learning, 205

Abstract data vectors, 333 H.M. case, 149-50
Abstract designs, 294 neocortical backprojections, 200, 202-
Abstraction, generalization basis, 359-62 6
Acetylcholine working memory, metabolism, 292
Alzheimer's disease, 250-51 Anterior commissure, 153-54
learning role, 229 Anterograde amnesia
ocular dominance plasticity, 220-21 hippocampus, time factors, 150-51
Acoustic frequency, 141 in memory subsystem model, 109-14
Activation Aplysia
experience-dependent plasticity, 225 developmental analysis, 27-51
memory changes, 108 learning principles, 55-58
"Active forgetting," 330 neural architecture, learning, 67-73
Activity pattern, 363 synaptic plasticity, 52-77
Adenylate cyclase, Aplysia, 57 time factors, conditioning, 146-47
Affect, 114-17 Arousal, and neuronal plasticity, 224
Age factors, and plasticity, 226-27 Associability effects, 89
Agnosia, 235 Association cortices
Algorithmic explanations Alzheimer's disease, 243
Hebbian synapse, 338-55 limbic system connections, 236-39
in neurosciences, 54-55 Associative matrix, 64-67
Allocortex versus competitive matrix, 191-93
Alzheimer's disease, 240, 246 conditioning explanation, 64-67
association cortices linkage, 237-39 error-correction procedures, 345-46
Alpha-conditioning, 58-61 Associative memory. See also Classical
Alzheimer's disease, 234-55 conditioning
anatomical alterations, 245-55 anatomy, 295
laminar specificity, 245-48 brain metabolism, 291-92
neural systems dissection, 234-55 functional equation, 326
D-amino-5-phosphonovalerate, 160 Hebb rule, algorithms, 342-47
Amnesia network model, 334-35
affective reactions, 114-17 neural architecture, 358-59
dissociation, 161-69 and neural computing, 323-37
hippocampus, time factors, 150-51 scene representation, 358-61
in memory subsystem model, 109-29 and visual self-organization, 227-28
partial dissociation, 166-69 versus working memory, 291-92, 294-
and recall, 117-26 95
recognition versus recall, 110-11 Asynchrony, 218-19
reflection role, 120-28 Attention
sentence comprehension, context, 121-26 conditioning potentiation, 94
types of, 126-27 layered network, syntax, 363-64
Amygdala neocortical backprojections, hippocampus,
Alzheimer's disease, 251-52 206
backprojection tutors, 202-4 neuronal development, vision, 220, 224
in conditioned taste aversion, 8 Auditory cortex
cortical connections, 237-39 Alzheimer's disease, 241
emotional "tag" role, 152 plasticity, 140-41

401
402 Index

Autoassociation matrix Cingulate cortex, 164-65

hippocampus, 194-95, 346 Cingulate gyrus, 241
model, 334-35, 346 Classical conditioning. See also Associative
Autoshaping, 63 memory; Pavlovian conditioning
Awareness, learning conditions, 3-6 cell biology, Aplysia, 57
Axospinous synapses, 315, 318 general principles, Aplysia, 55-58
Hebb rule, 339-40, 342-43
Back-coupling, 363, 378 matrix memory explanation, 68-73
Backprojections, 199-207 neural architecture theory, 67-73
Backpropagation algorithm, 276-84, 391 prewiring hypothesis, 58-67
Basal ganglia, 252 timing importance, 6-8
Bees, 148-49 Classification
Beta-conditioning, 58-61 matrix memories, hippocampus, 189-91
Between-phase convergence, 394-95 working memory role, 293
"Bidirectional" amnesia, 150 Clustering effect, 330-33
Bienenstock-Cooper-Munro algorithm, 309- Codon formation, 193
10, 349-50 Cognitive theories. See also Event
Binocularity representation
asynchrony effect, 218 critique, 375-79
developmental aspects, 213-21 Coherence, 198
gating mechanisms, 219-21 Columnar organizing principle, 302
neural network model, 307-21 Compensatory responses, 90
Blocking Competitive learning, 283
circuitry, 56-57 Competitive matrix memory, 189-91
conditioning pathway specificity, 95 versus associative matrix, 189-93
matrix memory weakness, 71 hippocampus, 189-207
Boltzmann machine architecture, 347-48 neocortical backprojections, hippocampus,
Boltzmann probability distribution, 303 199-207
Brain metabolism, 291-92 and "tutors," 202
Brain waves, 335-36 Completion, autoassociation matrix, 195
"Bubbles," 328-31 Computational theory, 54-55
Computer analogy, 17, 357
CA1 area Computer simulations, 330-33
Alzheimer's disease, 240, 246 Conditioned inhibition, 95
competitive learning matrix, 195-96 Conditioned response
R.B. case, 164 conditioned stimulus influence, 90-94
CA3 area mediated performance, 82-89
associative memory models, 346 in Pavlovian conditioning, 79-94
matrix memories, 193-95 plasticity locus, 80-89
Cajal's theory, 62 slow-motion video study, 84
Calcium and stimulus substitution, 58-62, 80-
age-dependent plasticity, 226-27 89
long-term potentiation, 318-19 Conditioned stimulus
NMDA receptor, synapse, 349-50 form of, influence, 90-94
use-dependent plasticity role, 221-23, in Pavlovian conditioning, 79-80
228 plasticity locus, 80-89
Caudate nucleus, 252 representations, 81
Cerebellum, 346-47 substitution view, 58-62, 79-89
Cerebral cortex versus unconditioned stimulus, behavior,
Alzheimer's disease, 234-55 90-96
backprojecting tutors, 202-4 Conditioned taste aversion, 3-8
Cerebral hemispheres Conditioning. See also Conditioned response
differential roles, 180-91 Hebb rule, algorithms, 342-43
macaques versus humans, 151 -54 preexisting reflex modulation, 58-67
reaction time study, 152-54 time factors, Aplysia, 146
Chaotic systems, 377 Configurations, 99-100
Cholinergic system Connectivity
age-dependent plasticity, 226 dynamic system, syntax, 367-69
Alzheimer's disease, 250-51 generalization basis, 361-62
ocular dominance plasticity, 220-21 time-scale assumption, 382-89
 Index 403

Consolidation Differences principle, 351

asynchrony effects, 218 Differential experience, 179-80
binocularity, development, 213-18 Differential facilitation
cortical backprojections, 205 and associative matrix, 64-67
Context effects critique, 62-67
memory content, 140-41 versus stimulus substitution, 58-62
occasion setting similarity, 142 Dihydropyridin, 227
sentence comprehension, amnesics, 121-26 Dishabituation
Contextual memory, 234-35 cellular analogues, development, 40-44
Contour perception, 213 definition, 28
"Core amnesia," 126-27 developmental aspects, Aplysia, 28-44
Corpus callosum, 153 general principles, Aplysia, 55-58
Cortical plasticity sensitization dissociation, adults, 44-46
calcium fluxes, 221-23 versus sensitization, emergence, 31-37, 46-
NMDA receptor role, 221-23 47
Corticocortical neural systems Dissociation
Alzheimer's disease, 239-55 in amnesia, 161-69
anatomy, 236-39, 243 criteria, 138-39, 160-61
visual cortex, cat, model, 313, 3)5-18 frontal lobe role, 168-69
Cost-benefit analysis, 10 medial temporal area, 164
Counterconditioning, 63 Distractibility, 295-96
Covariance rule, 345, 348 Distributed associative matrix
Crossbar switch, 325-26 and associative matrix, 64
Cross-correlation, 382-84 and blocking, 71-72
Cues, working memory, 296 and cell-biological explanation, 68-73
Cyclic AMP, 229 strengths, 67-69
weaknesses, 69-73
Dark-reared kittens, 214 Distributed feedback network, 328-29
Data gain fields, 280-81 Dorsal raphe nucleus, 253
Data structures, 363 Dorsal thalamus, 269
Dead-dressing problem, 206 Dorsolateral prefrontal cortex, 287
Declarative memory Double dissociations, 168-69
dissociation, amnesics, 163-70 Dual-process view, 49
hippocampal systems, 198-99
medial temporal region, 164 Eccles's theory, 180
partial dissociation, 166-69 Edelman's principle, 302-3
Decortication, 295 "Effective connectivity," 382-89
Decoupled models, 324 Electrical activity, 327
Delayed-response tasks "Electrical odor" experiments, 396
representational knowledge, 292-94 Electroconvulsive seizures, 147
working memory tests, 286-96 Electroencephalogram, 376
Dendrites, Hebb rule, 341-42 Emotion, 114-17
Dendritic spines Enriched experience, 179-80
long-term potentiation, 395 Entorhinal cortex
visual cortex, 318 Alzheimer's disease, 240, 245-47
Dentate granule cells, 193 association cortices, linkage, 237-38
Dentate gyrus, 291 neocortical projections, 164-65, 200
2-Deoxyglucose method, 291-92 systems theory, 196-99
Designators, 369-70 Entropy, 377
Developmental studies Enzyme synthesis, 147
activity-dependence, vision, 216-20 Episodic memory
algorithms, Hebb synapse, 348-50 Alzheimer's disease, 234-35
Aplysia, 27-51 in amnesia, dissociation, 163
gene-environment interaction, 211-12 Error backprojections, 207
self-organization, 211-29 Error-correction learning, 345-47, 378
Developmental trigger, 47 Escape locomotion, Aplysia, 37-40
Dialectic, 376 Event representation
Diencephalic region competitive learning system, 205
Alzheimer's disease, 252-53 critique, 375-79
declarative memory, 164 delayed-response tests, 292-93
404 Index

Event representation (continued) Geniculocortical synapses, 318

hippocampal systems, 198-99, 205-6 Glial barrier, 147
limits, 88-89, 375-79 "Graceful degradation," 69
neural architecture, 356-70 Gradient descent, 351
Pavlovian conditioning, 67, 79-89, 148 "Grandmother" cell code, 302
"perceptual" nature, 83-85 Gravity analysis, 383-86
plastic changes, 81 Grether's learning theory, 60-61
Snarsky's theory, 148 Gustatory trace-poisoning association, 3-8
working memory role, 292-94
Excitation, inhibition balance, 303-4 Habituation
Excitatory postsynaptic potentials, 40-44, 48 cellular analogues, development, 40-44
"Exclusive-or" problem, 72 definition, 28
Expectance, 96 developmental aspects, Aplysia, 28-31, 40-
Experience-dependent learning 49
constraints, 223-26 general principles, Aplysia, 55-58
vision, 219-26 Hamming distance, 359, 361
Explicit memory, 163 "Hebb molecule," 350
External cues, 296 Hebb rule, 338-55
Eye position cat visual cortex model, 311
network model, 276-84 historical context, 52-53
spatial representation, 272-84 implementation, 339-42
learning algorithms, 338-55
Fact memory tests, 163 neuronal development, binocularity, 218-
Familiarity, 113-14 19
"Fast learning," 332 and neuronal networks, 302, 311, 334
Feature analyzers, 198 time-scale assumption, 382-89
Feature detectors, 361 Hemispheric specialization. See Cerebral
Feature-negative discriminations, 97 hemispheres
Feature-positive discriminations, 96-99 Hermissenda, 341
Feed backward systems, 269 Heschl's gyrus, 241
Feed forward systems, 269 Hierarchical clustering, 333
Feedback networks Hippocampus
associative matrix models, 346 Alzheimer's disease, 240, 245-47
dynamic localization, 328-29 versus amygdala, learning, 205
Figure-ground segregation, 364, 369 autoassociation matrix, 194-95
Filter function, 325 backprojecting tutors, 202-5
Free recall, 168-69 connections of, 187
Frontal lobe syndrome, 296 cortical connections, 236-39
Frontal lobes cortical disconnections, 246
Alzheimer's disease, 241-43 declarative memory, 164
dissociation studies, 168-70 dissociation studies, 164, 169-70
versus medial temporal area, 168-69 functions, primates, 184-88
Functional ablation, 8 learning rule experiments, 394-96
Functional independence systems, 101 matrix memories, 188-96
Functional laws, 326-28 memory role, time, 148-51
neocortex backprojections, 199-207
GABAergic synapses neuronal networks, 184-209
dark-reared kittens, 315 occasion setting role, 97-99
monocular deprivation effect, 318 spatial memories, 185-88
Gating mechanism systems theory, anatomy, 196-99
experience-dependent plasticity, 224-25, trace "dating" role, 152
229 working memory, metabolism, 291-92
ocular dominance, 219-21 H.M. case
Gaussian interpolation algorithm, 278, 281- dissociation, 164
82 hippocampal studies, 149-50, 164
Generalization Holophone, 359
neural architecture, 358-59, 367-70 Hull's learning theory, 58-62
problem of, 359-62 "Hyperspecific" learning, 129
scene representation, 358-59, 369-79 Hypothalamus, 252-53
 Index 405

Image recognition, 368 Mapping array, 330-33

Information theory, 377 Mathematical techniques, 391
Inhibition, excitation balance, 303-4 Matrix memories, 69-73. See also Associative
Inhibitory process, Aplysia, 44-49 matrix; Distributed associative matrix
Interneurons, 347-48 associative versus competitive, 189-93
Interstimulus intervals and blocking, 71-72
cell biology, 57 and cell biology explanations, 68-73
habituation, development, 28-29 hippocampus, 188-96
Introspection, 152 in neural network models, 381-82
positive feature, 69-71
Korsakoffs amnesia properties, 66-67
affective responses, 114-17 weakness, 71-73
dissociation, 161-65, 168-69 Mean field approximation, 314-17, 378
frontal lobes, 168-69 Medial temporal region
memory subsystem model, 111-17 declarative memory, 164
recognition memory, 111-14 dissociation studies, 164, 169
versus frontal lobes, 169
Labyrinthine system, 149 Mediated learning, 85-89
Language, 152 Mediated performance experiment, 82-85
Lashley's theory, 159, 266, 268 Melody preference study, 114-17
Lateral feedback, 325-26 Membrane-triggering equations, 326-27
Lateral geniculate nucleus, 309-21 Memory systems, 101-2
"Lateral" inhibition, 303 Mental operation-level models, 325
Lateral olfactory tract, 396-97 Mesencephalon, 253
Laterality. See Cerebral hemispheres Metamemory, 168-69
"Law of effect" "Mexican hat function," 328-29
motor learning, 9-13 Mirror reading, 162
reinterpretation, 10-13 "Modification threshold," 308-10, 318-19,340
Layered systems, 361 Monocular occlusion
Learning algorithms cortical network effects, 315-18
definition, 338-39 and Hebb synapse, 349-50
Hebbian synapses, 338-55, 391 neural network effects, 315-18
Learning rule, 339 neuronal effects, 215-17
and long-term potentiation, 393-97 Monotonic output representations, 278, 282
Licking behavior, 10-13 Mossy fiber system, 191-96
Limbic system Motivation, 224
Alzheimer's disease, 234-55 Motor learning, 9-13
cholinergic factors, 250-51 Mountcastle's principle, 302
cortical connections, 236-39 Multilayer networks, 196, 347-48, 361
Linear associative matrix model, 345-46 Multineuron experiments, 382-89
Little's model, 303 Multiple-entry modular memory system,
Locality principle, 350-51 107-29
Locomotion Myelencephalon, 253
definition, 23
developmental studies, Aplysia, 37-40 Negentropy, 377
Locus coeruleus, 253 Neocortex
Long-delay learning, 57-58 hippocampal backprojections, 199-207
Long-term memory, 336 memory storage, 265-69
Long-term potentiation organization, 267-68
calcium role, 228, 318-19, 394 phytogeny, 268
as learning rule, 393-97 Nervous system-level models, 325
modification threshold model, 318-19 "NETtalk," 379
substrates, 395 Network models. See Neuronal networks
"Look-ahead" puzzles, 294 Neural architecture, 67-77
generalization, 369-70
Machine analogy, critique, 378-79 issue of, 356-58
Macromolecular processes, 146-47 learning forms, Aplysia, 67-77
Magnetism, 314 scene representation, 356-70
Mammillary bodies, 292 and syntax, 364-65
406 Index

Neural models, categorization, 324-25 Olfactory bulb

Neural networks. See Neuronal networks activity pattern, neurons, 376
Neurocomputers, 17 regeneration, learning, 14-16
Neuroribrillary tangles, 245-48 Olfactory cortex
Neuromnemonics, 137-42 Alzheimer's disease, 240, 243
Neuron proliferation, Aplysia, 47 learning rule experiments, 396-97
Neuronal networks Olfactory discrimination, 396-97
cat visual cortex model, 306-21 Ontogeny recapitulates phytogeny, 56
critique, 375-79 Operant conditioning, 9-13
definition, 324-25 Optic chiasm, 153
functional laws, 326-28 Optic nerve regeneration, 13-14
hippocampus, 184-209 Organizational models, 325
long-term potentiation studies, 390-97 Orientation, hippocampal role, 148-49
multineuron experiments, 382-89 Orienting responses, 91 -94
spatial representation, 271-84 Orthogonal backprojections, 201-7
structures and functions, 325-28 Overshadowing, 86-87
syntax, 362-69
theory, 381-89 Paired-pulse facilitation, 395
Neuronal plasticity. See also Synaptic Parahippocampal cortex, 241
plasticity Parahippocampal gyrus
calcium fluxes, 221-23 neocortical backprojections, 200, 206
concept of, 16-17 recall role, 206
visual system development, 211 -29 systems theory, 196-99
Neuronal sprouting, 57 Parallel network models, 276-84
Nictitating membrane, 138 Parietal lobe
NMDA receptor Alzheimer's disease, 243
age-dependence, 227 spatial representation, 271 -84
as "Hebb molecule," 349-50 Parietoprefrontal circuit, 291
long-term potentiation, calcium, 318-19, Partial dissociation, 166-69
394-95 Parvalbumin, 227
network learning rule, 394-95 Pattern vector, 202, 205-6
use-dependent plasticity role, 221-23, 349- Pavlovian conditioning. See also Classical
50 conditioning
Nodal element, 381 compound cues, 96-100
Noise conditioned stimulus form, 90-94
Boltzmann machine, 348 event representation, 79-89
brain dynamics assumptions, 377-78 memory forms, 78-105
Nonequilibrium thermodynamics, 378 plasticity locus, 81-89
Norepinephrine stimulus substitution, 79-89
learning role, 229 Perception
ocular dominance plasticity, 220-21 and configurations, 100
Noting function, reflection, 118-26 in memory subsystem model, 107-17
Nucleus basalis of Meynert, 249-51 "Perceptron," 363
Perforant path
Object Permanence Test, 293-94 Alzheimer's disease, 245-47
Object-place memory task, 185, 198 cortical connectivity, 188-94,237
Occasion-setting function, 87-88, 96-99 Perseveration, 296
context similarity, 142 Persistence, 218-19
hippocampal role, 97-99 Photographic images, 335-36
Occipital lobes, 241-43 Piaget's Object Permanence Test, 293-94
Ocular dominance Picture recognition, 111-14
chemical modulators, 220-21, 319 Piriform cortex, 396
dark-reared kittens, 214-16 Planar gain fields, 281-82
gating mechanisms, 219-21 "Planfulness," 119-20
Hebb synapse, 349-50 Positive feedback
NMDA receptor, calcium, 319, 349-50 brain dynamics, models, 378
plasticity, 214-29 neocortical backprojections, 206
Oculomotor delayed-response paradigm, Positive patterning discriminations, 96-
289 100
 Index 407

Posterior parietal cortex in memory subsystem model, 107-29

competitive learning, 283 and recall, 117-26
spatial representation model, 271-84 subprocesses, 117-26
Postsynaptic activation "Reflex prewiring," 59-67
functional laws, 328 plausibility, 62-67
Hebb rule, 340-41 Relaxation equation, 326
neural network model, 308-9 Repetition priming effect, 24
neuronal development, 216-18 Resonance properties, 225
Potentiated conditioning Response-reinforcement ratio, 10-13
conditioned stimulus influence, 92-94 Retinal position, 271-84
and event representation, 87-88 Retinal transplants, 13-14
Predictability, 71-73 Retrieval processes
Prefrontal cortex amnesia, systems, 125-26
representational knowledge, 292-94 hippocampus, time factors, 150-51
single-unit activity, 288-90 Retrograde amnesia, 150
working memory, 285-98 Runaway instability problem, 343
Presynaptic facilitation
Hebb rule, 340-41 Salience, 89
and learning forms, Aplysia, 68-69 Seizures, 147
Primary sensory areas, 241 Selective attention, 363
Priming Self-organization, 211-29
and anterograde amnesia, 109 learning similarity, 227-29
dissociation, amnesics, 161-62, 170 principles, 211-12
long-term potentiation, 394 sensory maps, computing, 330-33
neocortical backprojections, 206 visual system, model, 212-29
Principle of differences, 351 Semantic memory, 128, 163
Principle of locality, 350-51 Semantic priming, 206
Procedural memory, 128, 163-64 Sensitization
Proisocortices, 240-41 behavioral and cellular parallels, 43-44
Protein synthesis cellular analogues, development, 40-44
differential experience effect, 179 definition, 28
time factors, 146-47 developmental aspects, Aplysia, 31-49
Pseudoconditioning, 60-61, 68-69 versus dishabituation, emergence, 31-37,
Pursuit rotor learning, 162 46-47
Putamen, 252 dishabituation dissociation, adults, 44-46
Pyramidal cells, 268 escape locomotion, Aplysia, 37-40
general principles, 55-58
R2 motor neuron, 40-44 neural architecture theory, 67-73
Ramon y Cajal's theory, 62 Sensory preconditioning, 63-64
Raphe nuclei, 253 Sensory system
R.B. case, 164 information storage, 139
Reaction time, 152-54 in memory subsystem model, 107-14
Reactivation function, 119-26 paradox, 139-40
Recall reconstitution studies, 13-16
dissociation, 168-70 Sentence comprehension study, 121 -26
hippocampal-cortical system, 205-6 Serial discrimination, 96-100
hippocampus, time factors, 150-51 Sherrington's theory, 180
versus recognition, amnesia, 110-14 Shifting activation, 118-26
Recognition Short-term memory, 336
autoassociation matrix, 195 Sigmoidal function
dissociation, frontal lobe, 168-70 neural network laws, 327
hippocampus, time factors, 150-51 spatial representation, 278
memory subsystem model, 110-14 "Signal to noise ratio," 378
pictures versus words, 111-14 Silent synapses, 62
versus recall, amnesia, 110-11 Simultaneous pattern discrimination, 99-100
Recurrent collaterals, 194-95 Single-layer networks, 196
Reflective processes Siphon-withdrawal reflex
and amnesia, 120-27 developmental studies, 28-37
hippocampal role, 152 escape locomotion comparison, 37-40
408 Index

Skill learning neural network model, 308-21

in amnesics, dissociation, 162-63 in Pavlovian conditioning, locus, 80-89
and anterograde amnesia, 109 time-scale assumption, 382-89
Slow-motion video monitoring, 84 types of, 55-58
Snarsky's problem, 148 visual cortex development, 211-29, 348-50
Source amnesia, 168 Synaptic weights, 189
Spatial delayed alternation, 286-90 associative memory, model, 334-35, 342-
Spatial gain fields, 274, 280-81 47
Spatial representation cat visual cortex, 308-21
and eye position, 275-76 computational theory, 189, 331
hippocampus, 149, 185, 197-99,205 covariance rule, 345
neuronal network model, 276-84 dynamic system, syntax, 365-68
posterior parietal cortex, 271-84 functional laws, 327-28
topographical organization, 282-83 Hebb rule, algorithms, 342-47
working memory mechanisms, 288-89 neural network model, 308-9, 327-28,
Speech recognition, 332 331-35,381-82
Spike trains, 383-86 time-scale assumption, 382-89
Spin system analogues, 303 Synchronous stimulation, 393-94
Spiny stellate cells, 201 Syntax, 362-70
"Split-brain" macaques, 154 Systems research, 17
Sprouting, 57
"Start" signal, 149-50 Tanzi's concept, 181
Statistical fluctuations, 303 Taste-poisoning association, 5-8
Sternberg's paradigm, 152-54 Taxonomy, abstract data, 333
Stimulus devaluation, 142 "Teaching" role, 346, 379
Stimulus intensity, 46 Telencephalon, 249-52
Stimulus substitution Temporal factors
deviation from, 90-94 cerebral hemispheres, 151-54
versus differential facilitation, 58-62 classical conditioning, synapse, 340, 343
in Pavlovian conditioning, 79-89 conditioned taste aversion, 6-8
perceptual nature, 83-85 connectivity assumption, 382-89
plausibility, 59, 62 Hebb rule, 342-43
Stochastic independence, systems, 101 hippocampus, 148-51
Stochastic network models, 347-48 modeling difficulty, 336
Strabismus, 214 multineuron experiments, 382-89
"Strobe," 205 neural signals, syntax, 365-67
Stroop Test, 294 neuronal activation, development, 218
Subiculum neuronal plasticity factor, 224-25, 229
Alzheimer's disease, 240, 246 Temporal memory, 168
connectivity, 238 "Temporal paradox," 145
systems theory, 196-99 Textons, 364
Substitution view. See Stimulus substitution Thalamus, 252-53
Symbol systems, 362-64 Theta rhythm
Synaptic modification. See Synaptic plasticity autoassociation matrix, 195
Synaptic number, 179-80 hippocampus, 393-94
Synaptic plasticity, 52-77 temporal factors, 149-50
Aplysia studies, 52-77 Thought disorders, 295-96
Bienenstock's algorithm, 309-10 Thresholds
calcium flux, 221-23 calcium role, 318-20
cat visual cortex, 308-21 cat visual cortex, 308-21, 349
computational simplification, 331 covariance rule, 345
concept of, 16-17 functional laws, 327
dynamic systems, syntax, 365-67 Hebb rule, 341,345, 349
Hebb rule, 339-51 neural network model, 308-21, 327
hippocampus, matrix, 189 neuronal adaptations, 224-25, 229
learning algorithm, 338-55 ocular dominance, 219-21, 308
long-term potentiation studies, 390-97 physiological mechanism, 318-19
multineuron experiments, 382-89 Time factors. See Temporal factors
neural architecture, 67-73 Tower of London test, 294
 Index 409

Transnuclear mechanism, 146-47 Visual system

"Tutors," 202-4 activity-dependent development, 214-18
Two-class system, 365 critical period, 213
Two-dimensional typology, 325, 327 information processing, neocortex, 267-68
plasticity mechanism, 221-23
Unconditioned stimulus self-organization model, 212-29
versus conditioned stimulus behavior, 90- Vitamin D-dependent calcium-binding
96 protein, 227
in Pavlovian conditioning, 79-80
representation of, plasticity, 81-89 Wasps, 148-49
"US center," 80 Wisconsin Card Sort Test, 293
"US representation," 81-82 Within-phase convergence, 394-95
Word association priming, 162-63
V4 neurons, 283-84 Word list generation, 294
Vertical processing units, 268 Word recognition, 111-14
Vestibular system, 149 Working memory
Visual cortex anatomy, 295
activity-dependent modification, 216- versus associative memory, 291-92, 294-95
18 concepts, 285-86
Hebb synapse, 349-50 cortical localization, 285-98
learning algorithm, synapses, 348-50 evolutionary significance, 294-95
monocular deprivation, 215-16 hippocampal systems, 198-99
neural network model, cat, 306-21, 348- metabolic studies, 291 -92
50 neurophysiology, 288-90
neuronal development, 214-18, 348-50 spatial delayed alternation, 286-90
Visual representation. See Spatial
representation Zero mean field theory, 317