Scribd
Upload
27 views

Lecture Notes For ECE 695-09/08/03

The document summarizes key concepts in label prediction problems. It discusses: 1) The problem setup where the goal is to learn a hypothesis h that maps inputs x to labels y based on a training dataset, while minimizing prediction errors. 2) Basic solutions involve restricting the search for h to a hypothesis class H to avoid overfitting, and providing guarantees that H is sufficiently constrained. 3) Theoretical bounds on the difference between a hypothesis' expected prediction error and its error on the training data, which decrease as the size of the training set increases.

Uploaded by

Diego Moreno
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
27 views

Lecture Notes For ECE 695-09/08/03

The document summarizes key concepts in label prediction problems. It discusses: 1) The problem setup where the goal is to learn a hypothesis h that maps inputs x to labels y based on a training dataset, while minimizing prediction errors. 2) Basic solutions involve restricting the search for h to a hypothesis class H to avoid overfitting, and providing guarantees that H is sufficiently constrained. 3) Theoretical bounds on the difference between a hypothesis' expected prediction error and its error on the training data, which decrease as the size of the training set increases.

Uploaded by

Diego Moreno
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Lecture notes for ECE 695- 09/08/03.

1 Label Prediction: Problem setup


Unknown distribution P over X {0, 1}.
Input: S {X {0, 1}}
m
, S is an i.i.d sample from P. S = {(x
i
, y
i
) : 1 i
m)} is called Training data.
Output: h : X {0, 1}. h called hypothesis or predictor.
Cost: E
P
(h) = Pr({(x, y) : h(x) = y}). E
P
(h) called true error.
The problem is that we have to nd a good h. A naive approach will
encounter overtting.
Basic solution: Limit the search to some predesignated set H called the hy-
pothesis class.
First theoretical task: Provide guarantees for a good selection of H that
avoids overtting.
First, we use basic probability considerations to give guarantees for selec-
tions from H that overt. Fix some h and assume that E
P
(h) > . The
probability that E
S
(h) = 0 is at most (1 )
|S|
. So, we can conclude that
P, H, m, with probability > 1 if some h H has E
S
(h) = 0 then,
E
P
(h) <
1
m
(ln(|H|) + ln(
1

)), (1)
where m = |S|.
Equivalently, P, H, , , with probability > 1 if some h H has
E
S
(h) = 0 and
m >
1

(ln(|H|) + ln(
1

)), (2)
then E
P
(h) < where m = |S|.
We wish to analyze the error of hs that do not necessarily have zero empirical
error. The main tool that will be used is the Cherno/Hoeding bound.
Cherno/Hoeding Bound: Let {X
i
}

i=1
be independent Bernoulli ran-
dom variables each with expectation 0 p 1. Then, m
Pr
_

1
m
m

i=1
X
i
p

>
_
< 2e
2m
2
.
1
In our case, we use it as follows: Fix any h : X {0, 1} and let
X
i
=
_
1 if h(x
i
) = y
i
0 otherwise.
Note that E(X
i
) = E
P
(h). By Hoedings inequality,
Pr
_
|E
S
(h) E
P
(h)| >
_
< 2e
2m
2
.
It can also be shown that
Pr
_
E
P
(h) E
S
(h) >
_
< e
2m
2
. (3)
Exercise: Prove that P, H, m, , with probability > 1 , for all
h H
E
P
(h) E
S
(h) <
_
(ln(|H|) + ln(1/))
2m
, (4)
where m = |S|.
Bounds depending on description length: Given H, we x a description
of the hypothesis in H as follows. Let

() : H {0, 1}

map all h H to
binary strings of nite length. Thus,

h is the string coding h and let |

h| denote
the description length of h. Assume that the above binary mapping is prex
free. Then, by Krafts inequality,

hH
2
|

h|
1. (5)
Suppose that we have dierent error tolerances for dierent h viz., let
h
be
the error tolerance for h. Set

h
=

(ln 2(|

h|) + ln(1/))
2m
,
and by (3) we get for every xed h,
Pr
_
E
P
(h) E
S
(h) >
h
_
<

2
|

h|
Using the union bound and Kraft inequality (5), we can conclude that
Pr(h H : E
P
(h) E
S
(h) >
h
) < ,
OR
With probability atleast 1 , h H, E
P
(h) E
S
(h)
h
.
2
More generally, we can say that P, H, m, , with probability > 1
for all h H
E
P
(h) E
S
(h) <

(ln 2(|

h|) + ln(1/))
2m
, (6)
where m = |S|.
Bayesian Approach: We model our prior knowledge by a prior distribution
over the hypotheses. This is equivalent to specifying a particular description of
H. For example, for a 2-adic distribution q(h) over H we can choose binary
Human coding as the description mapping. Then q(h) = 2
|

h|
and
E
P
(h) E
S
(h) <
_
(ln(1/q(h)) + ln(1/))
2m
,
Relations between dierent predictors Note that there exists some
f : X {0, 1} minimizing E
P
(). It is given by
f(x) =
_
1 if
P(x,1)
P(x,0)+P(x,1)
> 0.5
0 otherwise.
(7)
Ofcourse since P is unknown, f is unknown. So, let us constrain ourselves
to a class H of predictors. Let h

be the predictor that minimizes E


S
() and
h

be the predictor that minimizes E


P
() over H. By denition,
E
S
(h

) E
S
(h

), (8)
E
P
(h

) E
P
(h

). (9)
Then it follows from our arguments that
E
P
(h

) < E
S
(h

) + E
S
(h

) + E
P
(h

) + 2.
This gives us a handle on the dierence between the true errors of the sam-
plewise optimal predictor and the best predictor (in terms of minimizing the
true error).
3

You might also like