Cognitive Systems Research 5 (2004) 119134

www.elsevier.com/locate/cogsys

A symbolic model of human attentional networks

Action editor: Ron Sun

Hongbin Wang a,*, Jin Fan b, Todd R. Johnson a

a

School of Health Information Sciences, University of Texas Health Science Center at Houston,
7000 Fannin, Suite 600 Houston, TX 77030, USA
Sackler Institute for Developmental Psychobiology, Weill Medical College of Cornell University,
1300 York Ave., Box 140, New York, NY 10021, USA
Received 15 August 2003; accepted 8 December 2003

Abstract
An increasing body of evidence has shown that attention is a multi-type and multilevel cognitive faculty. The
dominant computational modeling approaches to attention have often focused on one specic type of attention at one
specic level. In particular, various connectionist modeling techniques at the subsymbolic level have been widely
adopted. In this paper, we report a symbolic computational model of the Attentional Network Test, which simultaneously involves dierent types of attention (alerting, orienting, and executive control), each subserved by distinctive
attentional networks in the brain. The model was developed in ACT-R, a rule-based cognitive architecture. The results
show that the model, by sequentially ring rules at a rate of about one every 40 ms, was able to capture the eect of each
attentional network. The model implies that while the attentional networks can be distinguished at both neuroanatomical and behavioral levels, dierent attentional networks may adopt similar computational operations at least at a
symbolic rule level.
2004 Elsevier B.V. All rights reserved.

1. Introduction
Although everybody knows what attention is
(James, 1890), the nature of attention remains
elusive after more than one hundred years of active
research and hundreds of publications in this
subject. This is hardly surprising given the com-

Corresponding author. Tel.: +1-713-500-3911; fax: +1-713500-3929.

E-mail address: [email protected] (H. Wang).

plexity of the phenomenon itself. A large body of

evidence, in the broad areas of cognitive psychology and cognitive neuroscience, has demonstrated
that attention is not a unitary but a diversied
faculty of the human cognitive system (Parasuraman & Davies, 1984; Posner & Raichle, 1994) it
is subserved by multiple interrelated attentional
networks in the brain and manifests itself in a
variety of types and at dierent levels in almost
every aspect of human behavior, from perception
and motor control to working memory, skill acquisition, response selection, and consciousness

1389-0417/$ - see front matter 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.cogsys.2004.01.001

120

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

(e.g., Luck, Woodman, & Vogel, 2000; Pashler,

1998; Posner, DiGirolamo, & Fernandez-Duque,
1997).
Partly because of this complexity, traditional
computational modeling approaches to human
attention have typically focused on one specic
type of attention at one specic level. In particular,
various connectionist modeling techniques at the
subsymbolic level have been widely adopted
(Churchland & Sejnowski, 1992; Rumelhart &
McClelland, 1986). Examples include Cohen and
colleagues well-known connectionist model of
executive control of attention for the Stroop task
(Cohen, Dunbar, & McClelland, 1990) and Mozers connectionist modeling work of selective
spatial attention in object recognition (Mozer,
1991). For more recent development and reviews
in these lines of research, see (Botvinick, Braver,
Barch, Carter, & Cohen, 2001; Mozer & Sitton,
1998; OReilly & Munakata, 2000).
While it has been very fruitful, this practice is
problematic mainly due to two reasons. First, by
focusing on only a few specic types of attention, it
fails to adequately appreciate the possible interactions among dierent varieties of attention and
dierent attentional networks. Second by emphasizing only the subsymbolic level of analysis, it
encounters diculties in understanding the link
between neural information processing and the
operation of attention at the symbolic/cognitive
level, which is psychologically as real as the underlying neurophysiological mechanisms of attention (e.g., Newell, 1990).
In this article, we attempt to address these issues by reporting a symbolic computational model
of attention that simultaneously simulated the essential operations of multiple attentional networks. The model was developed in ACT-R, a
production rule based hybrid cognitive architecture (Anderson, 1993; Anderson & Lebiere, 1998).
We believe that such a model oers an opportunity
for us to explore the psychologically plausible
symbolic foundations of attention and, more importantly, to compare/contrast it with neurallybased connectionist models of attention.
It should be noted that developing symbolic
models of attention is not new. Byrne and Anderson (1998) augmented ACT-R with a set of

perceptual-motor modules, which implemented

some rudimentary visuospatial attention functions. Altmann and Davidson (2001) and Lovett
(2002) have developed ACT-R models of the
Stroop task, in which executive control of attention plays a major role. Our model diers from
these models in two aspects. First, our model simultaneously simulates multiple attentional networks, thus capturing the operational features of
dierent types of attention, including alerting,
orienting, and executive control. Second, both
Altmann and Davidsons and Lovetts models relied extensively on the subsymbolic mechanisms of
ACT-R to implement parallel processing and executive control. In our model, symbolic operations
at the production rule level are emphasized.
This article is structured as follows. We rst
briey introduce the attentional networks account
of human attention and an experimental paradigm
that was designed to simultaneously measure the
performance of multiple attentional networks. We
then briey introduce ACT-R. Finally, an ACT-R
model of human attentional networks is reported
and discussed.

2. Human attentional networks

Recent advances in cognitive psychology and
cognitive neuroscience have suggested that there
exist multiple attentional networks in the brain,
each of which subserves a dierent type of attention (Fan, Raz, & Posner, 2003c; Posner & Dehaene, 2000; Posner & Petersen, 1990; Posner &
Raichle, 1994). At least three attentional networks,
for alerting, orienting, and executive control, have
been distinguished at both cognitive and neuroanatomical levels. Specically, alerting involves a
change in the internal state to become ready for
any incoming task-related events. Alerting is an
important source of attention in the sense that
maintaining an adequate level of alertness is critical for optimal performance. Neuroimaging evidence has revealed that the alerting network
consists of some frontal and parietal areas particularly of the right hemisphere. Lesions of these
areas reduce alertness. Orienting, closely related to
the conventional selective visuospatial attention,

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

involves selectively focusing on one or a few items

out of many candidate inputs. Apparently, orienting can be voluntary (top-down and controlled)
or involuntary (bottom-up and automatic), overt
(with head/eye movement) or covert (without
head/eye movement), location-based (orienting to
spatial locations) or object-based (orienting to
objects). Evidence has shown that the orienting
network includes parts of the superior and inferior
parietal lobe, frontal eye elds and such subcortical areas as the superior colliculus of the midbrain
and the pulvinar and reticular nucleus of the
thalamus. Finally, executive control of attention is
related to monitoring and resolving conicts in the
presence of competing information. Executive
control is often needed in higher level mental operations including planning, decision making, error detection, novel or not well-learned responses,
and overcoming habitual actions. Converging evidence from neuroimaging and neuropathological
studies has suggested that the executive control
network consists of the midline frontal areas (anterior cingulate cortex) and lateral prefrontal cortex. See Posner and Dehaene (2000) for a more
detailed description of these attentional networks.
An experimental paradigm called the Attentional Network Test (ANT) was recently devel-

121

oped to simultaneously measure the performance

of the three attentional networks and evaluate
their interrelationships (Fan, MaCandliss, Sommer, Raz, & Posner, 2002). It is essentially a
combination of a spatial cueing task (Posner,
1980) and a anker task (Eriksen & Eriksen, 1974),
as illustrated in Fig. 1. In each ANT trial, the
participants look at a xation point in the center
of a computer screen and wait for the stimulus to
appear. The stimulus consists of a row of ve
horizontal arrows and the participants task is to
report the direction (left or right) that the center
arrow (the target) points to by pressing a corresponding key as quickly as possible. The reaction
time (RT) is recorded. The four arrows surrounding the target, with two on each side, are
called the ankers. These anker arrows point either in the same direction as that of the target (the
congruent condition), or in the opposite direction
(the incongruent condition). An additional condition (the neutral condition) is also included in
which the ankers are four straight lines with no
arrowheads. One manipulation is that the stimulus
row is not presented at the xation location. Instead, it can be presented at two locations, either
above a xation point (top) or below it (bottom).
Therefore, to identify the direction of the target,

Fig. 1. A sketch of the design of ANT.

122

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

presumably the participants have to shift their

attention either upward or downward to the
stimulus row. Yet another manipulation is the cue
condition. That is, the stimulus row may be preceded by a cue (the cued condition) or may not
(the no-cue condition). In particular, when there is
a cue, this cue may be presented at the center xation location (the center-cue condition), at the
top or bottom location where the stimulus row is
to appear (the spatial-cue condition), or at both
top and bottom locations (the double-cue condition). Therefore, while a spatial-cue precisely predicts where the stimulus is to appear, in both the
center-cue condition and the double-cue condition
the participant cannot infer that information from
the cue.
The ANT paradigm adopts the following formula to measure the eciency of each of the three
attentional networks:
Alerting eciency
RT(no-cue) ) RT(double-cue),
Orienting eciency
RT(center-cue) ) RT(spatial-cue),
Executive control eciency
RT(incongruent) ) RT(congruent).
The underlying rationale is as follows. First,
since the appearance of a cue in the double-cue
condition precisely alerts the participants that the
stimulus row is going to occur after 500 ms and
there is no such alerting in the no-cue condition, the
RT dierence between the no-cue condition and the
double-cue condition can be used as a measure of
the alerting eciency. 1 Second, in both the centercue condition and the spatial-cue condition the
participants are alerted when the stimulus row is
going to occur. The dierence is where it is going to
occur: while the spatial-cue precisely tells the participants where the stimulus row is going to appear,

Note that both the double-cue and center-cue conditions

alert the participants that the stimulus row is going to occur but
do not tell where. The reason that the double-cue condition
instead of the center-cue condition is used to dene the alerting
eciency is that the double-cue condition involves diused
attention, which is more similar to the situation in the no-cue
condition. Similar reasoning underlies the denition of the
orienting eect: both the center-cue and spatial-cue conditions
involve non-diused cued attention.

the center-cue does not. Presumably, knowing the

stimulus location before the stimulus appears allows the participants to be better prepared for response in that they can shift attention in advance to
the desired location. As a result, the RT dierence
between the center-cue condition and spatial-cue
condition serves a measure of the orienting eciency. Finally, compared to the congruent condition, the presence of incongruent ankers in the
incongruent condition results in interference in response. Evidence has shown that this type of interference shares the largely identical brain
foundations with that in the Stroop eect and requires similar executive control of attention to resolve (see Fan, Flombaum, McCandliss, Thomas,
& Posner, 2003a). Therefore, the executive control
cost is measured as the RT dierence between the
incongruent and congruent conditions.
Fan et al. (2002) tested 40 normal adult participants using the ANT paradigm. They adopted a
balanced within-subject factorial design and presented 3 blocks of trials to each participant with 96
trials in each block (3 anker conditions
4 cue
conditions
2 possible stimulus locations
2 possible target arrow directions
2 repetitions). Their
reaction time results led to the eciency measures
of 47  18 ms, 51  21 ms, and 84  25 ms, for
alerting, orienting, and executive control, respectively. A further correlation test showed that the
eects of the three attentional networks were statistically independent.
ANT as a general and robust framework for
measuring attentional eciency has been evaluated in a variety of studies. For example, Rueda
et al. (in press) adopted the ANT paradigm to
study human attentional development. They
measured the attentional eciency of children in
dierent age groups (from 6 to 10 years old) and
compared the results with adults. They found that
each network showed a dierent developmental
course. ANT has also been used in various patient
populations, including attention decit hyperactivity disorder (ADHD, Booth, Carlson, & Tucker, 2001), and borderline personality disorders
(Posner et al., 2002). More recently, the attentional
eciency measures from ANT have been linked to
neuroimaging results and genetic variations (Fan,
Fossella, Sommer, Wu, & Posner, 2003b).

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

3. ACT-R
ACT-R is a production rule based cognitive
modeling architecture developed by John Anderson and colleagues over a period of nearly two
decades (Anderson, 1983, 1990, 1993; Anderson &
Bower, 1973; Anderson & Lebiere, 1998). In essence, ACT-R explains human cognition by proposing a model of the knowledge structures and
knowledge deployment that underlie cognition.
Although ACT-R consists of a nontrivial subsymbolic component for computations involving
activation and association, it relies extensively on
various symbolic structures for knowledge representation. For example, ACT-R makes a fundamental distinction between declarative and
procedural knowledge (see also Schacter & Tulving, 1994). Declarative knowledge corresponds to
things people are aware of and can usually describe to others. Procedural knowledge is knowledge that people display in behavior but are not
conscious of. Declarative knowledge in ACT-R is
represented in terms of chunks. Procedural
knowledge is represented in terms of production
rules, which are condition-action pairs. Both
chunks and production rules are fundamental
symbolic structures in ACT-R and are regarded as
the atomic components of thought in the sense that
they are as far down as one can go in the symbolic
decomposition of cognition. In ACT-R, on average every 50 ms, one production rule is chosen to
re, a few declarative chunks are processed, and
cognition advances one step. Therefore, it is
claimed that ACT-R captures the symbolic grain
size of cognition.
Canonical ACT-R is mainly a theory of higherlevel cognition, excelling in modeling phenomena
in human memory, learning, and problem solving.
One recent advance in ACT-R is the release of
ACT-R 5.0, which intends to close the gap between ACT-R and the external environment by
augmenting ACT-R with a set of perceptual-motor
modules (Byrne & Anderson, 1998). To implement
the visual system, ACT-R 5.0 implements some
rudimentary attention functions. For example,
there is a move-attention command that higher
cognition can send to the visual system. This
command allows the visual system to shift atten-

123

tion to a designated location in the environment

and then start visual recognition. The time cost of
this command is specied by a parameter, which is
typically set to be 85 ms, a reasonable setting based
on many previous studies (see Egeth & Yantis,
1997).
Although ACT-R 5.0 represents important
progress in the symbolic modeling of attention,
one limitation is that it only handles the spatial
orienting type of attention. It is not clear how
ACT-R 5.0 can be used to model other types of
attention such as alerting and executive control at
the symbolic rule level.

4. ANT on ACT-R
4.1. The model
We developed a computational model for the
ANT task in the framework of ACT-R 5.0. Our
purpose is two-fold. First, we want to explore how
alerting and executive control of attention can be
modeled symbolically in ACT-R 5.0 and how the
three types of attention work together to produce
the cognitive performance. In particular, we hope
to show how the eects of dierent attentional
networks can be explained by just ring dierent
sets of production rules but without resorting to
various subsymbolic mechanisms. Second, such a
model oers a possibility for us to cross-validate
those models based on various connectionist
modeling results and neuroimaging data, and by
doing so we hope that we can probe the possible
connections about the function of attention among
dierent levels of description.
We started by analyzing the major functional
components in the ANT task. Based on the design
presented in Fig. 1, we distinguished 6 stages in
performing a generic ANT trial (see Fig. 2):
1. Fixation and cue expectation. In this stage, the
participant xates at the xation point and prepares for something to occur. Note that at this
time the expectation of the participant was uncertain either a cue (in the cued conditions)
or a stimulus row (in the no-cue condition)
could appear. As a result, we hypothesize that
the participant has to engage in an uncertain

124

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

Fig. 2. A functional decomposition of the ANT task. The number in parentheses in each stage represents the number of production
rules that are used to implement these functions.

preparatory state that is appropriate for either

possibility.
2. Cue or stimulus. This second stage starts when a
visual object appears, which can be either a cue
(in the cued conditions) or a stimulus row (in
the no-cue condition). A decision has to be made
based on what appears. If it is a cue, go to stage 3
for cue processing. If it is a stimulus row, go directly to stage 5 for stimulus processing.
3. Cue processing. In this stage the cue is processed and attention is shifted to the corre-

sponding cued location. More specically, for

the center-cue, attention remains at the center
xation location. For the spatial-cue, attention
is moved to the cued location (top or bottom)
where a stimulus row is to appear. For the
double-cue condition, instead of assuming that
attention is diused to both the cued locations,
we hypothesize that attention shifts randomly
to one of the cued locations. We will discuss
this issue in more detail in the model discussion section.

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

4. Stimulus expectation. In this stage the participant is expecting a stimulus to appear, with attention remaining at the cued location(s). Note
that in this stage the participant has engaged in
a preparatory state that is particularly suitable
for stimulus processing.
5. Stimulus processing. In this stage the stimulus
has to be processed and the direction of the target arrow identied. Specically, in the spatialcue condition, since attention has already been
allocated to the location where the stimulus appears, presumably a more accurate and rapid
stimulus processing can occur. In the centercue condition, an additional attentional movement needs to be initiated to shift attention to
the stimulus location. Depending upon the result from stage 3, similar operations need to
be performed in the double-cue condition.
Again, we leave the details of this condition to
the model discussion section.
6. Response. In this nal stage, the participant
responds by initiating a key pressing process
based on the result from the previous stage.
We implemented our model by mapping these
functional components into ACT-R production
rules. We used 36 rules to fully implement these
functions and cover all the conditions. The distribution of the rules in each stage is also shown
in Fig. 2. The name and a brief description of
each rule are provided in Appendix A.
With these rules our model could perform the
ANT task and interact with the same experimental environment that human participants
use. In a typical simulation trial, visual objects,
including the xation, the cue(s), and the stimulus, are presented in a computer window in a
timed sequence according to the design as shown
in Fig. 1. The model can see what is currently
in the window through its visual system. At any
time point, based on the goal and what the
model currently sees, one of the best matching
rules is selected to re, which may lead to a
chain of more rule rings. In the current model
each rule ring is set to cost 40 ms instead of
the default 50 ms (see the model discussion
section for more details). The trial ends when
the model makes a response by pressing a
response key. The time from the stimulus pre-

125

sentation to the key-press is then recorded as the

reaction time.
As an example, a running trace generated when
the model performs a center-cue incongruent ANT
trial is listed in Appendix B. It shows that the trial
starts at time 0 and the xation lasts for 1255 ms,
during which two production rules have red. A
center cue appears at time 1255 and disappears at
time 1355 (for a duration of 100 ms). Two production rules have red during this cue processing
period. At time 1755 (after a 400 ms wait) the
stimulus appears. It then takes eight production
rules ring in sequence for the model to make a
response and initiate a key-press at time 2165. The
key-press is not nished until time 2375, resulting
in a reaction time of 620 ms (2375 ) 1755). 2
The ACT-R code of the model can be downloaded online from http://www.sahs.uth.tmc.edu/
hwang/antmodeling.htm.
4.2. Model evaluation
We evaluated the performance of the model by
using the model as a simulated human participant
to perform the ANT experiment. The experimental
design was exactly same as that used by Fan et al.
(2002), with 3 blocks of 96 randomly mixed trials.
The RT results of testing 100 simulated subjects
are summarized in Table 1 and Fig. 3, along with
the experimental results from Fan et al. (2002). A
correlation analysis shows very high correlations
(0.99 for RTs and 0.97 for error rates) between the
simulation and experimental results. From the results in Table 1, we estimated the eects of the
three attentional networks, which are shown in
Table 2, together with Fan et al. (2002)s results. It
shows a close match between the two sets of data,
with a notable exception that the simulated standard deviations are consistently smaller than the
empirical ones. The reason is that we did not add
any between-subject variance in our model. As a

Note that the RT of 620 ms shown in this trace is dierent

from the RT of the same condition in Table 1, which is 580 ms.
The reason is that 580 ms is an average result of many trials. In
some trials, the RT may be lower depending on the outcome of
conict resolution.

126

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

Table 1
Means RT and error rates under each condition
Cue type
Congruency

No-cue

Center-cue

Double-cue

Spatial-cue

(a) Mean RTs  Standard deviations in ms from the experiment and (the model simulation)
Neutral
529  47 (527  3)
483  46 (487  3)
472  44 (467  5)
Congruent
530  49 (526  4)
490  48 (486  3)
479  45 (466  6)
Incongruent
605  59 (621  14)
585  57 (580  14)
574  57 (562  15)

442  39 (441  4)
446  41 (441  4)
515  58 (522  16)

(b) Error rates in % from the experiment and (the model simulation)
Neutral
1.17 (0.96)
0.93 (0.92)
Congruent
0.73 (0.75)
0.54 (1.00)
Incongruent
3.49 (3.25)
4.88 (3.79)

0.78 (0.79)
0.44 (0.83)
3.51 (2.67)

1.56 (0.71)
0.59 (0.79)
4.27 (3.50)

result, these simulated variances actually reected

those within-subject variations in performing the
ANT task. Although we can add some betweensubject noise to get a better t of the variances, it
leads to a more complicated model.
One of the key results in Fan et al. (2002)
concerns the interrelationship among the three
attentional networks. They found that the eciency measures of the three attentional networks
were independent of each other based upon the
statistics of correlations. Similar independences
were found in the simulation results, which are
shown in Table 3, along with empirical data from
Fan et al. (2002). It is clear that the two sets of
data reveal similar correlation structures among
the measures. An examination of the model suggests that the independence resulted from the difTable 3
Correlations between attentional networksa
Results

Alerting

Orienting

Fig. 3. Mean RTs in various conditions from (a) the experimental study and (b) our model simulation.

Orienting
Experiment
Simulation

Table 2
Eects of attentional networks

Executive control
Experiment
Simulation

0.05
)0.07

)0.12
)0.14

Mean RT
Experiment
Simulation

0.09
)0.14

0.29
)0.15

Attentional networks
Eects (ms)
(mean  SD)

Alerting

Orienting

Executive
control

Experiment
Simulation

47  18
55  7.4

51  21
45  7.0

84  25
86  7.4

Executive
control

0.08
0.04

0.44b
0.80b

Correlations are calculated based on the relevant measures

of all participants (or simulated participants).
b
Correlation is signicant at the 0.01 level.

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

ferent sets of rules underlying dierent testing

conditions. The signicant positive correlation
between the executive control eect and the overall
mean RT, shown in both the experiment and the
simulation, is interesting. It probably reects the
strong inuence of the RTs in the incongruent
condition on the overall mean RT. Overall these
results suggest that the model captured well the
various attentional eects that the ANT task was
designed to measure.
4.3. Model discussion
What are the mechanisms underlying the models attention-related performance? Here we briey
report several major features of the model that we
think are critically responsible for the performance.
First, to perform any one ANT trial, in stage 5
(stimulus processing), one key-press process (for
responding) and at least one move-attention process (for encoding the target) are necessary. Based
on well-established evidence ACT-R presets the
time cost of these processes to be 210 and 85 ms,
respectively. We did not change these settings.
Subtracting these costs from the RT measures
gives us a rough range of time that the production
rules have to explain. Note that although the
model consists of 36 production rules, in any one
trial not all of them will re (or are eligible to re)
because they are programmed to cover all dierent
conditions. When multiple rules are applicable in a
situation, our model relies on ACT-Rs built-in
conict resolution mechanisms to choose one to
re.
Taken together, these settings allow us to predict
how many production rules are recruited in each
trial. More specically, the model uses, depending
upon the conditions, about 48 production rules to
perform a trial, for total costs of about 160320 ms.
The trace listed in Appendix B demonstrates this.
It shows that after the stimulus appears, eight
production rules re in sequence (notice-stimuluswith-centercue-and-shift, attend-to-at-large-target,
attended-item-is-left-to-the-target, refocus-again-ifincongruent, harvest-target, goahead-responding-ifit-is-the-target, decide-left, respond), resulting in
two move-attention and one key-press commands.
The nal reaction time is 620 ms.

127

Second, although the experimental results in

Fan et al. (2002) show a performance with very
high accuracy, participants made errors, in all
conditions. How could the model make errors? We
speculated that error could come in two major
sources: random noise in the response and the
distraction caused by the incongruent ankers.
While the rst source might be responsible for the
baseline (neutral condition) error rate, which is
about 1.11%, the second source might account for
the higher error rate in the incongruent conditions,
which is about 4.04%. We modeled the rst source
of error through two production rules (random-left
and random-right), which say to respond left (or
right) regardless of the direction of the target arrow. They compete with the decide-left or decideright rule to determine the response in each trial,
with their chances of winning being very low (the
odds are set to be 1:20). The second source is
modeled through a production called hurryupresponding-no-matter-whether-target-or-not. This
rule rushes to respond as soon as an arrow, which
may not be the center target arrow, is encoded.
This rule competes, again with relatively low
winning probability, with more cautious rules that
emphasize accuracy (e.g., checking and refocusing)
and causes errors if it wins in the incongruent
condition (the odds are 1:3). As shown in Table 1,
the model error rates are about 0.85 and 3.30, for
the baseline average and the incongruent condition
respectively, indicating a good match to the
experimental data.
Third, how has the alerting eect been explained in the model? The empirical measure of 47
ms is about the time cost of one ACT-R production rule. This is indeed one major mechanism
underlying how our model generated the alerting
eect. There is one critical production rule that can
re in the no-cue condition but not in any other
cued condition. This rule is called not-cue-soswitch-state-and-shift-attention (rule 4 in Appendix
A). As we mentioned earlier, since either a cue or a
stimulus row could occur following stage 1, the
participant had to engage in an uncertain preparatory state. This was dierent from the specic
and certain preparatory state (for stimulus-processing) following stage 4 in the cued conditions.
This rule summarizes the cost associated with the

128

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

state change (from expecting either a cue or a

stimulus to specically expecting a stimulus) and is
responsible for a major part of the alerting eect. 3
It is important to note that this alerting eect did
not come from the time saving due to an anticipatory shift of attention but from an expectancy
state switching. We believe that this account for
the alerting eect captures the essential characteristics of alerting as a change of preparatory state
(from a uncertain state to a better prepared state),
which may often facilitate later processes.
Fourth, the orienting eect reects the dierence between the center-cue condition and the
spatial-cue condition, with an empirical measure
of about 51 ms. Presumably this is due to the time
saved from knowing in advance where the stimulus
is going to appear. In our model this dierence is,
again, modeled by a production rule named noticestimulus-with-centercue-and-shift (rule 17 in Appendix A), which basically says that if a stimulus
appears while attention is on the center location,
try to shift attention to the stimulus. We assume
that in the spatial-cue condition attention has already been allocated to the correct spatial location
before the stimulus appears, whereas in the centercue condition the ring of this additional production rule is necessary to bring the system to a
comparable level of stimulus processing. This additional step costs 40 ms and is the major source of
the orienting eect.
Fifth, the eect of executive control reects the
eect of the ankers on the task performance. The
empirical measure is about 84 ms. In our model this
eect is modeled using two mechanisms. First, the
result of move-attention is not perfect. When attention is directed to one location, an object nearby
may be selected, especially when the scene is
crowded or the objects are similar (Pashler, 1998).
This kind of imprecision is one fundamental reason

Note that the cost of the production rule (40 ms) does not
match the simulated alerting eect (55 ms, see Table 2). The
reason is that the simulated eect is an average number that
incorporates multiple sources although the cost of the production rules accounts for the major part of the eect. Similar
mechanisms apply to the orienting and executive control eects
discussed next.

for the anker eect. Second, there are three production rules that specically distinguish between
the incongruent and congruent condition in the
cases when a distracter arrow, but not the center
target, is focused on. When this situation arises, we
hypothesize that instead of performing an automatic but expensive (costly) re-focusing, a simpler
congruency detection is conducted. Specically, if a
congruent condition is detected, the system moves
on to respond since a re-focusing is not necessary.
This is implemented by the rule named goaheadresponding-if-congruent (rule 28 in Appendix A). If
an incongruent condition is detected, the system
can either move on to do a direct re-encoding
(harvest-target-directly-if-incongruent, rule 30 in
Appendix A, equivalent to a simple direction reversal) or perform a re-focusing again (refocusagain-if-incongruent, rule 29 in Appendix A).
Compared to the congruent condition, the rst
strategy costs an additional 85 ms (a direct moveattention) and the second strategy costs an
additional 125 ms (one additional rule ring for
refocusing rst and then a move-attention). With
the second strategy a little more likely to be selected
than the rst one (the odds are set to be about 2:1),
this leads to an additional cost of about 111 ms on
average for the incongruent conditions. It is this
cost, accrued quite often since there are four times
more ankers than the target, that is mainly responsible for the executive control eect.
One interesting empirical result from Fan et al.
(2002)s study is the small but reliable dierence
(11 ms) between the center and double-cue conditions. This can be conveniently explained by a
diused attention assumption. More specically,
the center-cue induces the participant to focus attention on the xation location while the doublecue makes the participant diuse attention at both
the top and bottom locations. As a result, compared to the center-cue condition where the stimulus location receives no priming, the stimulus
location in the double-cue condition is primed a
small amount, which may speed up later stimulus
processing. This explanation, however, provides a
challenge to a symbolic model of attention. How
can attention be diused symbolically when we
only have in hand a move-attention command,
which presumably shifts the focus of attention to a

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

pre-specied spatial location? While a parallel

global spatial priming process is plausible from a
neural computing perspective, we argue that it is
equally plausible to speculate a series of symbolic
move-attention operations occurring quickly and
sequentially. This speculation also captures psychologically how we focus our spatial attention in
our everyday life (keep your attention here!). Our
model uses this mechanism to account for what
occurs in the double-cue condition. Instead of using a diused attention mechanism, we assume
that attention is moved twice, each time to one of
the two cues. The nal location is where attention
resides to wait for the stimulus to appear. This is
like a betting strategy. If the betting is correct (i.e.,
the stimulus appears in the nal waiting location),
a situation similar to that in the spatial-cue condition arises. If the betting is wrong (i.e., the
stimulus actually appears in the location opposite
to where one is waiting), a situation similar to the
center-cue arises an additional attentional
shifting (notice-stimulus-with-doublecue-and-shift)
has to be performed when the stimulus appears.
Since the betting is random, on average this gives
rise to half of the cost of a production rule ring,
which is about 20 ms. The simulation result is
19  8 ms.
Finally, we would like to point out that our
model manipulates ve free parameters to obtain
the goodness of t. The rst four ne-tune ACTRs standard conict resolution mechanisms and
are mainly responsible for tting the error rates,
Specically, two parameters the odds of winning
the conict resolution for rule random-left randomright (1:20) and rule hurryup-responding-no-matterwhether-target-or-not (1:3) are implemented in
the model by setting their :p values, which in ACTR designate the probabilities that their rings will
lead to eventual task success. The other two parameters, which are expected gain noise (egs 3)
and utility threshold (ut )100), are set to add
some random noise in ACT-Rs conict resolution
so that the model can sometimes re wrong rules
and make mistakes in response. The fth parameter, the ring time of each rule, is more critically
responsible for tting the RTs and attentional efciencies. We have shortened the cost of each rule
ring from ACT-Rs default 50 to 40 ms. One

129

criticism of our model is that this parameter is

widely accepted as one of a few of ACT-Rs fundamental architectural primitives and changing it
is indicative of a misuse of the architecture. Here,
we oer two main justications for this parameter
change. First, the ANT task is extremely simple
and has tight temporal constraints, which may
speed up the rate of rule ring. Second and more
importantly, this parameter should not be taken
for granted. In particular, as long as the change is
within a limited range and consistent across all the
productions rules in a model, it may not be a
misuse of the architecture. One underlying justication for the default setting, oered by EPIC,
another rule-based cognitive architecture that excels in modeling human perceptual-motor performance (Meyer & Kieras, 1997a, 1997b), is that 50
ms conforms well with the mean period between
zero crossings in the brains 10 Hz a rhythm (e.g.,
Callaway & Yeager, 1960). Note, however, that
the a rhythm is typically not xed at 10 Hz but
varies within a range (e.g., 713 Hz). From this
perspective, shortening the rule ring cost to 40 ms
conforms well with a slightly faster 12.5 Hz brain
rhythm, still within the a waves range but consistent with the tight temporal constraints of the
task. Clearly, further investigations are necessary
to evaluate our modication and clarify the implication of the parameter.

5. General discussion
Though attention is a complex multilevel cognitive function, the dominant approaches to computational modeling of attention have emphasized
its close association with lower level processes such
as perception and thus often adopted some connectionist and neural modeling techniques. In addition, typically only one type of attention is
focused on. In this article, we reported a symbolic
computational model of the ANT task, which incorporates the work of multiple attentional networks including alerting, orienting, and executive
control. The model was developed in ACT-R.
Using basic symbolic knowledge modeling structures such as rules and strategies, all of which often
have clear and straightforward psychological

130

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

meaning, we were able to show that the operation

of attentional networks could be understood at a
symbolic level. In this section we would like to
briey discuss some implications of developing
such a model.
First, our model demonstrates that as a multilevel construct attention can be modeled at both
the subsymbolic and symbolic levels, and more
importantly, that there might exist dierent
mechanisms underlying the operation of attention
at each level. For example, one distinct feature of
our model is that it is a strict serial processing
model: rules re in sequence with one ring leading
to another. This is dierent from various connectionist models, which extensively adopt parallel
and distributed processing. While parallel distributed processing is certainly plausible at the neural
and neural networks levels, serial processing is
psychologically justiable due to the limited capacity of human working memory (Baddeley,
1986; Baddeley & Logie, 1999). Being able to
model the same behavioral regularities at both
levels using dierent mechanisms oers another
strong support for the multilevel nature of
attention.
The concept of the production rule is fundamental to our model of attention. Rules re in
sequence and operate at a rate of about 50 ms (40
ms in the current model) per production rule. As
argued by ACT-R, production rules dene the
atomic components of thought at the symbolic
level. With this claim in mind, when we examine
the eciency measures of attentional networks
reported in Fan et al. (2002) it seems that they
make good sense at a rule level: these measures
(51, 47, and 84 ms, for alerting, orienting, and
executive control, respectively. See Table 2.) are
just in the range of a few rule rings time period.
Perhaps all we need is about one (for alerting and
orienting) or two (for executive control) additional
rule rings to explain symbolically the work of
attentional networks. This is indeed what our
model demonstrated.
While our model simulates the ANT results
reasonably well, one question is whether there is
any independent justication for the particular set
of production rules used in our model. We address
this question in two ways. First, as we reported

earlier, a functional analysis of the ANT task was

performed before we developed the model. Such
an analysis identied a list of psychologically
meaningful functional components underlying the
ANT task, which motivated and constrained the
selection of production rules in the model. Second,
our model can make nontrivial predictions and be
applied to account for phenomena beyond the
normal ANT task reported in this article. An example is attentional deciency. The model predicts
that if the rule not-cue-so-switch-state-and-shiftattention (rule 4 in Appendix A) is somehow
nonfunctional (e.g., there is no such a rule or it
takes a long time to re), the alerting operation
will be aected. A recent study (Booth et al., 2001)
using ANT on ADHD children shows that these
children revealed a damaged alerting function (i.e.,
a larger alerting eect), which may suggest that
these children possessed a dysfunctional rule 4. By
altering the operation of rule 4, we may quantitatively simulate these ADHD childrens attentional performance. Another interesting prediction
our model can make is about the eect of executive
control. If the strength of the rule hurryup-responding-no-matter-whether-target-or-not (rule 25
in Appendix A) is increased due to certain psychological and/or pathological conditions, for example in schizophrenic patients, then the model
would predict a decit executive control due to a
lower neural activities in the executive control attentional networks. It seems that these rules provide an interesting symbolic summary of the
function of the underlying networks.
While a large body of evidence from neuroimaging and neuropathology has suggested that
these distinct types of attention are subserved by
independent attentional networks (see Fan et al.,
2003c, for a review), our model provides some
interesting implications about the operations of
the attentional networks and the interrelationship
among the dierent types of attention. The current model suggests that alerting can be explained
by a rule that captures and performs a preparatory state change, orienting can be explained by a
rule that moves attention before the appearance
of the stimulus, and executive control can be
explained by a set of rules that detect and resolve
a conict. Therefore, the neuroanatomical inde-

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

pendence of dierent attentional networks is

simulated at a symbolic rule level by the interaction of dierent sets of rules red in dierent
task conditions. In this sense, it seems that at
least at a symbolic rule level dierent attentional
networks may adopt similar computational
mechanisms.
Finally, modeling attention at a symbolic rule
level through serial processing mechanisms raises
an interesting question about the role of attention
in cognitive architectures. As ACT-R moves towards a modular architecture (Anderson, Bothel,
Byrne, & Lebiere, submitted), it seems that attention may be more properly modeled as an add-on
module like other perceptual-motor modules (e.g.,
Byrne & Anderson, 1998). However, while a peripheral module may be appropriate for modeling
the visuospatial type of attention (orienting), it
appears unsuitable for alerting and executive
control types of attention due to their intriguing
relationship with central cognition. To us, production rules summarize nicely the conscious efforts and psychological reality associated with
being alerted and resolving conicts, as we demonstrated in our model. However, the need to reduce the cost of each rule ring from the ACT-R
default of 50 ms to 40 ms in this model raises
questions about the adequacy of adopting a purely
serial symbolic approach to modeling attention. In
general we envision that to model the full range of
human attentional functions in ACT-R we may
have to combine modular, subsymbolic, and rulelevel approaches.

Acknowledgements
This research was supported in part by the
Grant N00014-01-1-0074 from the Oce of Naval
Research Cognitive Science Program. We would
like to thank Dr. Michael I. Posner for providing
constructive comments that have improved the
quality of the paper. Correspondence and request
for reprints should be sent to Hongbin Wang,
School of Health Information Sciences, University
of Texas Health Science Center at Houston, 7000
Fannin Suite 600, Houston, TX 77030, USA.
E-mail: [email protected].

131

Appendix A. The production rules used in the model

Stage 1. Fixation and cue expectation
1. P notice-xation
An ANT trial starts when a xation appears
2. P encode-xation-and-waiting
Encode the xation and start expecting
for a cue
Stage 2. Cue or stimulus?
3. P notice-something-appeared
Some visual object appears on the
screen
4. P not-cue-so-switch-state-and-shiftattention
Resist the surprise and shift attention to
the stimulus
Stage 3. Cue Processing
5. P notice-a-center-cue
notice a center cue
6. P notice-a-top-cue
notice a top cue
7. P notice-a-bottom-cue
notice a bottom cue
8. P given-a-top-cue-nd-a-bottom-cue
check if it is in fact a double cue
9. P given-a-bottom-nd-a-top-cue
check if it is in fact a double cue
10. P nd-no-more-cue-so-spatialcue
No more cue so it is a spatial cue
11. P nd-more-cue-so-doublecue
It is a double cue indeed
Stage 4. Stimulus expectation
12. P anticipating-the-stimulus
Wait and anticipate the stimulus
Stage 5. Stimulus processing
13. P notice-stimulus-at-cued-top-location-andattend
Notice an arrow at the cued top
location so encode it
14. P notice-stimulus-at-cued-top-location-buta-neutral-item-is-selected
Notice a neutral item at the cued top
location so refocus on the center item
15. P notice-stimulus-at-cued-bottom-locationand-attend
Notice an arrow at the cued top
location so encode it

132

16.

17.

18.

19.

20.

21.

22.

23.
24.

25.

26.

27.

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

P notice-stimulus-at-cued-bottom-locationbut-a-neutral-item-is-selected
Notice a neutral item at the cued top
location so refocus on the center item
P notice-stimulus-with-centercue-and-shift
The stimulus row appears while attention
is on the center location so shift attention
to it
P notice-stimulus-with-doublecue-and-shift
The stimulus row appears in a location
that is dierent from the current attended
location in the doublecue condition so shift
attention to it
P notice-stimulus-with-doublecue-and-an
-arrow-is-focused-on-so-attend
The stimulus row appears in a location
that is same as the current attended
location in the doublecue condition so encode the arrow
P notice-stimulus-with-doublecue-but-aneutral-item-is-focused-on-so-shift
The stimulus row appears in a location
that is same as the current attended
location in the doublecue condition but it
seems a neutral distracter is selected so
refocus
P attend-to-at-large-target
Encode the arrow that is currently
focused on
P shift-to-at-large-target-from-a-neutralitem
If the currently focused on item is a
neutral distract, refocus
P harvest-target
Encode the target after refocusing
P goahead-responding-if-it-is-the-target
If the encoded item is the center arrow,
goahead to respond
P hurryup-responding-no-matter-whethertarget-or-not
As long as the encoded item is an arrow,
hurry up to respond
P attended-item-is-right-to-the-target
Hold on, the currently attended arrow is to
the right of the target
P attended-item-is-left-to-the-target
Hold on, the currently attended arrow is to
the left of the target

28.

P goahead-responding-if-congruent
But move on to respond since they are
congruent
29. P refocus-again-if-incongruent
Refocus on the center location
30. P harvest-target-directly-if-incongruent
Encode the center arrow directly since its
location is available
Stage 6. Response
31. P decide-left
Decide to make a left response since the
target is a left arrow
32. P decide-right
Decide to make a right response since the
target is a right arrow
33. P random-left
Randomly decide to make a left response
34. P random-right
Randomly decide to make a right
response
35. P respond
Respond by pressing the decided key
36. P rexating-and-wait-for-next-trial
refocusing on the xation and be ready
for the next trial

Appendix B. The trace in a center-cued incongruent

trial
? (do-trial :cue centercue :anker incongruent)
;; show xation at 0 and for 1255 ms
Time 0.000: Notice-Fixation Selected
Time 0.040: Notice-Fixation Fired
Time 0.040: Module :vision running command
move-attention
Time 0.125: Module :vision running command
focus-on
Time 0.125: Encode-Fixation-And-Waiting Selected
Time 0.165: Encode-Fixation-And-Waiting
Fired
Time 1.255:  Running stopped because time
limit reached.
;; show cue (or nocue) at 1255
Time 1.255: Notice-A-Center-Cue Selected
Time 1.295: Notice-A-Center-Cue Fired

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

Time 1.295: Module :vision running command

move-attention
Time 1.295: Anticipating-The-Stimulus Selected
Time 1.335: Anticipating-The-Stimulus Fired
Time 1.355:  Running stopped because time
limit reached.
;; the cue disappears and continue xating at 1355
Time 1.380: Module :vision running command
focus-on
Time 1.755:  Running stopped because time
limit reached.
;; the stimulus appears at 1755
Time 1.755: Notice-Stimulus-With-CentercueAnd-Shift Selected
Time 1.795: Notice-Stimulus-With-CentercueAnd-Shift Fired
Time 1.795: Module :vision running command
nd-location
Time 1.795: Attend-To-At-Large-Target Selected
Time 1.835: Attend-To-At-Large-Target Fired
Time 1.835: Module :vision running command
move-attention
Time 1.835: Attended-Item-Is-Left-To-The-Target Selected
Time 1.875: Attended-Item-Is-Left-To-The-Target Fired
Time 1.875: Module :vision running command
nd-location
Time 1.920: Module :vision running command
focus-on
Time 1.920: Refocus-Again-If-Incongruent Selected
Time 1.960: Refocus-Again-If-Incongruent Fired
Time 1.960: Module :vision running command
nd-location
Time 1.960: Harvest-Target Selected
Time 2.000: Harvest-Target Fired
Time 2.000: Module :vision running command
move-attention
Time 2.000: Goahead-Responding-If-It-Is-TheTarget Selected
Time 2.040: Goahead-Responding-If-It-Is-TheTarget Fired
Time 2.085: Module :vision running command
focus-on
Time 2.085: Decide-Left Selected
Time 2.125: Decide-Left Fired
Time 2.125: Respond Selected

133

Time 2.165: Respond Fired

Time 2.165: Module :motor running command
press-key
Time 2.165: Module :vision running command
nd-location
Time 2.165: Rexating-And-Wait-For-NextTrial Selected
Time 2.205: Rexating-And-Wait-For-NextTrial Fired
Time 2.205: Module :vision running command
move-attention
Time 2.290: Module :vision running command
focus-on
Time 2.315: Module :motor running command
preparation-complete
Time 2.365: Module :motor running command
initiation-complete
Time 2.375: Device running command outputkey
Time 2.465: Module :motor running command
nish-movement
Time 2.465: Checking for silent events.
Time 2.465:  Nothing to run: No productions, no
events.
;; response at 2375, RT 620 ms
Time 4.090: * Running stopped because time limit
reached.
;; trial ends at 4090

References
Altmann, E. M., & Davidson, D. J. (2001). An integrative
approach to stroop: Combining a language model and a
unied cognitive theory. In Proceedings of the 23rd annual
meeting of the Cognitive Science Society (pp. 2126).
Hillsdale, NJ: Erlbaum.
Anderson, J. R. (1983). The architecture of cognition. Cambridge, MA: Harvard University Press.
Anderson, J. R. (1990). The adaptive character of thought.
Hillsdale, NJ: Lawrence Erlbaum Associates.
Anderson, J. R. (1993). Rules of the mind. Hillsdale, NJ:
Lawrence Erlbaum Associates.
Anderson, J. R., Bothel, D., Byrne, M., & Lebiere, C.
(submitted). An integrated theory of the mind. Psychological Review.
Anderson, J. R., & Bower, G. H. (1973). Human
associative memory. Mahwah, NJ: Lawrence Erlbaum
Associates.

134

H. Wang et al. / Cognitive Systems Research 5 (2004) 119134

Anderson, J. R., & Lebiere, C. (1998). The atomic components

of thought. Hillsdale, NJ: Lawrence Erlbaum Press.
Baddeley, A. D. (1986). Working memory. Oxford: Oxford
University Press.
Baddeley, A. D., & Logie, R. H. (1999). Working memory: The
multiple-component model. In A. Miyake & P. Shah (Eds.),
Models of working memory: Mechanisms of active maintenance and executive control. New York: Cambridge University Press.
Booth, J., Carlson, C. L., & Tucker, D. (2001). Cognitive
inattention in the ADHD subtypes. In Paper presented at
the tenth scientic meeting of the International Society for
Research in Child and Adolescent Psychopathology (ISRCAP), Vancouver.
Botvinick, M. M., Braver, T. S., Barch, D. M., Carter, C. S., &
Cohen, J. D. (2001). Conict monitoring and cognitive
control. Psychological Review, 108, 624652.
Byrne, M. D., & Anderson, J. R. (1998). Perception and action.
In The atomic components of thought. Mahwah, NJ:
Lawrence Erlbaum Associates.
Callaway, E., & Yeager, C. L. (1960). Relationship between
reaction time and electroencephalographic alpha base.
Science, 132, 17651766.
Churchland, P. S., & Sejnowski, T. J. (1992). The computational
brain. Cambridge, MA: MIT Press.
Cohen, J. D., Dunbar, K., & McClelland, J. L. (1990). On the
control of automatic processes: A parallel distributed
processing account of the stroop eect. Psychological
Review, 97(3), 332361.
Egeth, H. E., & Yantis, S. (1997). Visual attention: Control,
representation, and time course. Annual Review of Psychology, 48, 269297.
Eriksen, B. A., & Eriksen, C. W. (1974). Eects of noise letters
upon the identication of a target letter in a nonsearch task.
Perception and Psychophysics, 16, 143149.
Fan, J., Flombaum, J. I., McCandliss, B. D., Thomas, K. M., &
Posner, M. I. (2003a). Cognitive and brain consequences of
conict. Neuroimaging, 18(1), 4257.
Fan, J., Fossella, J. A., Sommer, T., Wu, Y., & Posner, M. I.
(2003b). Mapping the genetic variation of executive attention onto brain activity. Proceedings of the National
Academy of Sciences USA, 100(12), 74067411.
Fan, J., MaCandliss, B. D., Sommer, T., Raz, A., & Posner, M.
I. (2002). Testing the eciency and independence of
attentional networks. Journal of Cognitive Neuroscience,
14(3), 340347.
Fan, J., Raz, A., & Posner, M. I. (2003c). Attentional
mechanisms. In Encyclopedia of Neurological Sciences. San
Diego, CA: Academic Press.
James, W. (1890). Principles of psychology. New York: Holt.
Lovett, M. C. (2002). Modeling selective attention: Not just
another model of stroop (NJAMOS). Cognitive Systems
Research, 3(1), 6776.

Luck, S. J., Woodman, G. F., & Vogel, E. K. (2000). Eventrelated potential studies of attention. Trends in Cognitive
Sciences, 4(11), 432440.
Meyer, D. E., & Kieras, D. E. (1997a). A computational theory
of executive cognitive processes and mutliple-task performance: Part 1. Basic mechanism. Psychological Review,
104(1), 365.
Meyer, D. E., & Kieras, D. E. (1997b). A computational theory
of executive cognitive processes and mutliple-task performance: Part 2. Accounts of psychological refractory-period
phenomena. Psychological Review, 104(4), 749791.
Mozer, M. C. (1991). The perception of multiple objects: A
connectionist approach. Cambridge, MA: MIT Press.
Mozer, M. C., & Sitton, M. (1998). Computational modeling of
spatial attention. In H. E. Pashler (Ed.), Attention. East
Sussex, UK: Psychology Press.
Newell, A. (1990). Unied theories of cognition. Cambridge,
MA: Harvard University Press.
OReilly, R. C., & Munakata, Y. (2000). Computational
explorations in cognitive neuroscience. Cambridge, MA:
MIT Press.
Parasuraman, R., & Davies, D. R. (1984). Varieties of attention.
San Diego, CA: Academic Press.
Pashler, H. E. (1998). The psychology of attention. Cambridge,
MA: MIT Press.
Posner, M. I. (1980). Orienting of attention. Quarterly Journal
of Experimental Psychology, 32, 325.
Posner, M. I., & Dehaene, S. (2000). Attentional networks. In
M. S. Gazzaniga (Ed.), Cognitive neuroscience: A reader.
Malden, MA: Blackwell Publishers.
Posner, M. I., DiGirolamo, G. J., & Fernandez-Duque, D.
(1997). Brain mechanisms of cognitive skills. Consciousness
and Cognition, 6, 267290.
Posner, M. I., & Petersen, S. E. (1990). The attention systems of
the human brain. Annual Review of Neuroscience, 13,
2542.
Posner, M. I., & Raichle, M. E. (1994). Images of mind. New
York: Scientic American Library.
Posner, M. I., Rothbart, M. K., Vizueta, N., Levy, K. N.,
Evans, D. E., Thomas, K. M., & Clarkin, J. F. (2002).
Attentional mechanisms of borderline personality disorder.
Proceedings of the National Academy of Sciences USA,
99(25), 1636616370.
Rueda, M. R., Fan, J., McCandliss, B. D., Halparin, J. D.,
Gruber, D. B., Lercari, L. P., & Posner, M. I. (in press).
Development of attentional networks in childhood. Neuropsychologia.
Rumelhart, D. E., & McClelland, J. L. (1986). Parallel
distributed processing: Explorations in the microstructure
of cognition. Foundations (Vol. 1). Cambridge, MA: MIT
Press.
Schacter, D. L., & Tulving, E. (1994). Memory systems 1994.
Cambridge, MA: MIT Press.