Machine Learning

10-701, Fall 2015

VC Dimension and Model

Complexity

Eric Xing

Lecture 16, November 3, 2015

Reading: Chap. 7 T.M book, and outline material
© Eric Xing @ CMU, 2006-2015 1
Last time: PAC and Agnostic
Learning
 Finite H, assume target function c ∈ H

 Suppose we want this to be at most δ. Then m examples suffice:

 Finite H, agnostic learning: perhaps c not in H

 

 with probability at least (1-δ) every h in H satisfies

© Eric Xing @ CMU, 2006-2015 2

What if H is not finite?
 Can’t use our result for infinite H

 Need some other measure of complexity for H

– Vapnik-Chervonenkis (VC) dimension!

© Eric Xing @ CMU, 2006-2015 3

What if H is not finite?
 Some Informal Derivation
 Suppose we have an H that is parameterized by d real numbers. Since we are
using a computer to represent real numbers, and IEEE double-precision floating
point (double's in C) uses 64 bits to represent a floating point number, this means
that our learning algorithm, assuming we're using double-precision floating point,
is parameterized by 64d bits

 Parameterization

© Eric Xing @ CMU, 2006-2015 4

How do we characterize
“power”?
 Different machines have different amounts of “power”.
 Tradeoff between:
 More power: Can model more complex classifiers but might overfit.
 Less power: Not going to overfit, but restricted in what it can model

 How do we characterize the amount of power?

© Eric Xing @ CMU, 2006-2015 5

Shattering a Set of Instances
 Definition: Given a set S = {x(1), … , x(m)} (no relation to the
training set) of points x(i) X, we say that H shatters S if H
can realize any labeling on S.

I.e., if for any set of labels {y(1), … , y(d)}, there exists some
hH so that h(x(i)) = y(i) for all i = 1, …, m.

 There are 2m different ways to separate the sample into two

sub-samples (a dichotomy)

© Eric Xing @ CMU, 2006-2015 6

Three Instances Shattered

Instance space X

© Eric Xing @ CMU, 2006-2015 7

The Vapnik-Chervonenkis
Dimension
 Definition: The Vapnik-Chervonenkis dimension, VC(H), of
hypothesis space H defined over instance space X is the size
of the largest finite subset of X shattered by H . If arbitrarily
large finite sets of X can be shattered by H , then VC(H)  .

© Eric Xing @ CMU, 2006-2015 8

VC dimension: examples
Consider X = R, want to learn c: X{0,1}
What is VC dimension of

 Open intervals:
H1: if x>a, then y=1 else y=0

 Closed intervals:

H2: if a<x<b, then y=1 else y=0

© Eric Xing @ CMU, 2006-2015 9

VC dimension: examples
Consider X = R2, want to learn c: X{0,1}

 What is VC dimension of lines in a plane?

H= { ( (wx+b)>0  y=1) }

© Eric Xing @ CMU, 2006-2015 10

 For any of the eight possible labelings of these points, we can find a linear
classier that obtains "zero training error" on them.
 Moreover, it is possible to show that there is no set of 4 points that this
hypothesis class can shatter.
© Eric Xing @ CMU, 2006-2015 11
 The VC dimension of H here is 3 even though there may be sets of size 3 that it
cannot shatter.
 under the definition of the VC dimension, in order to prove that VC(H) is at least
d, we need to show only that there's at least one set of size d that H can shatter.
© Eric Xing @ CMU, 2006-2015 12
 Theorem Consider some set of m points in Rn. Choose any
one of the points as origin. Then the m points can be
shattered by oriented hyperplanes if and only if the position
vectors of the remaining points are linearly independent.

 Corollary: The VC dimension of the set of oriented

hyperplanes in Rn is n+1.
Proof: we can always choose n + 1 points, and then choose one of the
points as origin, such that the position vectors of the remaining n points are
linearly independent, but can never choose n + 2 such points (since no n +
1 vectors in Rn can be linearly independent).

© Eric Xing @ CMU, 2006-2015 13

The VC Dimension and the
Number of Parameters
 The VC dimension thus gives concreteness to the notion of
the capacity of a given set of h.
 Is it true that learning machines with many parameters would
have high VC dimension, while learning machines with few
parameters would have low VC dimension?

An infinite-VC function with just one parameter!

where  is an indicator function

© Eric Xing @ CMU, 2006-2015 14

An infinite-VC function with just
one parameter
 You choose some number l, and present me with the task of finding l
points that can be shattered. I choose them to be

 You specify any labels you like:

 Then () gives this labeling if I choose  to be

 Thus the VC dimension of this machine is infinite.

© Eric Xing @ CMU, 2006-2015 15

Sample Complexity from VC
Dimension
 How many randomly drawn examples suffice to -exhaust
VSH,S with probability at least (1 - )?

ie., to guarantee that any hypothesis that perfectly fits the training data is
probably (1-δ) approximately (ε) correct on testing data from the same
distribution

m  1 (4 log2 (2 /  )  8VC ( H ) log2 (13 /  ))

Compare to our earlier results based on |H|:

m  21 2 (ln H  ln(1 /  ))

© Eric Xing @ CMU, 2006-2015 16
Mistake Bounds
So far: how many examples needed to learn?
What about: how many mistakes before convergence?

Let's consider similar setting to PAC learning:

 Instances drawn at random from X according to distribution D
 Learner must classify each instance before receiving correct
classification from teacher
 Can we bound the number of mistakes learner makes before
converging?

© Eric Xing @ CMU, 2006-2015 17

Statistical Learning Problem
 A model computes a function:

 Problem : minimize in w Risk Expectation

 w : a parameter that specifies the chosen model

 z = (X, y) are possible values for attributes (variables)
 Q measures (quantifies) model error cost
 P(z) is the underlying probability law (unknown) for data z

© Eric Xing @ CMU, 2006-2015 18

Statistical Learning Problem (2)
 We get m data from learning sample (z1, .. , zm), and we suppose
them iid sampled from law P(z).
 To minimize R(w), we start by minimizing Empirical Risk over this
sample :

 We shall use such an approach for :

 classification (eg. Q can be a cost function based on cost for misclassified points)
 regression (eg. Q can be a cost of least squares type)

© Eric Xing @ CMU, 2006-2015 19

Statistical Learning Problem (3)
 Central problem for Statistical Learning Theory:

What is the relation

between Risk Expectation R(W)
and Empirical Risk E(W)?

 How to define and measure a generalization capacity

(“robustness”) for a model ?

© Eric Xing @ CMU, 2006-2015 20

Four Pillars for SLT
 Consistency (guarantees generalization)
 Under what conditions will a model be consistent ?

 Model convergence speed (a measure for generalization)

 How does generalization capacity improve when sample size L grows?

 Generalization capacity control

 How to control in an efficient way model generalization starting with the only given
information we have: our sample data?

 A strategy for good learning algorithms

 Is there a strategy that guarantees, measures and controls our learning model
generalization capacity ?

© Eric Xing @ CMU, 2006-2015 21

Consistency

A learning process (model) is said to be consistent if

model error, measured on new data sampled from
the same underlying probability laws of our original
sample, converges, when original sample size
increases, towards model error, measured on
original sample.

© Eric Xing @ CMU, 2006-2015 22

Consistent training?

%error

Test error

Training error
number of training examples
%error

Test error

Training error
number of training examples

© Eric Xing @ CMU, 2006-2015 23

Vapnik main theorem
 Q : Under which conditions will a learning model be
consistent?
 A : A model will be consistent if and only if the function h that
defines the model comes from a family of functions H with
finite VC dimension d

 A finite VC dimension d not only guarantees a generalization

capacity (consistency), but to pick h in a family H with finite
VC dimension d is the only way to build a model that
generalizes.

© Eric Xing @ CMU, 2006-2015 24

Model convergence speed
(generalization capacity)
 Q : What is the nature of model error difference between
learning data (sample) and test data, for a sample of finite
size m?
 A : This difference is no greater than a limit that only depends
on the ratio between VC dimension d of model functions
family H, and sample size m, i.e., d/m

This statement is a new theorem that belongs to Kolmogorov-

Smirnov way for results, i.e., theorems that do not depend on
data’s underlying probability law.

© Eric Xing @ CMU, 2006-2015 25

Agnostic Learning: VC Bounds
 Theorem: Let H be given, and let d = VC(H). Then with
probability at least 1- , we have that for all h  H,

or

recall that in finite H case, we have:

© Eric Xing @ CMU, 2006-2015 26

 Model convergence speed

% error

Test data error

Confidence
Interval
Learning sample error

Sample size m

© Eric Xing @ CMU, 2006-2015 27

How to control model
generalization capacity
Risk Expectation = Empirical Risk + Confidence Interval

 To minimize Empirical Risk alone will not always give a good

generalization capacity: one will want to minimize the sum of
Empirical Risk and Confidence Interval

 What is important is not the numerical value of the Vapnik

limit, most often too large to be of any practical use, it is the
fact that this limit is a non decreasing function of model family
function “richness”

© Eric Xing @ CMU, 2006-2015 28

Empirical Risk Minimization
 With probability 1-, the following inequality is true:

 where w0 is the parameter w value that minimizes Empirical Risk:

© Eric Xing @ CMU, 2006-2015 29

Minimizing The Bound by
Minimizing d
 Given some selection of learning machines whose empirical risk is
zero, one wants to choose that learning machine whose associated
set of functions has minimal VC dimension.

 By doing this we can attain an upper bound on the actual risk. This does not prevent a
particular machine with the same value for empirical risk, and whose function set has
higher VC dimension, from having better performance.
 What is the VC of a kNN?
© Eric Xing @ CMU, 2006-2015 30
Structural Risk Minimization
 Which hypothesis space should we choose?

 Bias / variance tradeoff

 SRM: choose H to minimize bound on true error!

unfortunately a somewhat loose bound...

© Eric Xing @ CMU, 2006-2015 31
SRM strategy (1)
 With probability 1-,

 When m/d is small (d too large), second term of equation becomes

large

 SRM basic idea for strategy is to minimize simultaneously both

terms standing on the right of above majoring equation for (h)

 To do this, one has to make d a controlled parameter

© Eric Xing @ CMU, 2006-2015 32

SRM strategy (2)
 Let us consider a sequence H1 < H2 < .. < Hn of model family
functions, with respective growing VC dimensions
d1 < d2 < .. < dn

 For each family Hi of our sequence, the inequality

is valid
 That is, for each subset, we must be able either to compute d, or to get a bound
on d itself.

 SRM then consists of finding that subset of functions which

minimizes the bound on the actual risk.
© Eric Xing @ CMU, 2006-2015 33
SRM strategy (3)

SRM : find i such that expected risk (h) becomes

minimum, for a specific d*=di, relating to a specific
family Hi of our sequence; build model using h from Hi

Risk

Best Model Total Risk

Empirical
Confidence interval
Risk
In h/L

Model Complexity
h*
© Eric Xing @ CMU, 2006-2015 34
Putting SRM into action:
linear models case (1)
 There are many SRM-based strategies to build models:

 In the case of linear models

y = <w|x> + b,

one wants to make ||w|| a controlled parameter: let us call HC the

linear model function family satisfying the constraint:
||w|| < C

Vapnik Major theorem:

When C decreases, d(HC) decreases
||x|| < R

© Eric Xing @ CMU, 2006-2015 35

Putting SRM into action:
linear models case (2)
 To control ||w||, one can envision two routes to model:

 Regularization/Ridge Regression, ie min. over w and b

RG(w,b) = S{(yi-<w|xi> - b)² |i=1,..,L} +  ||w||²

 Support Vector Machines (SVM), ie solve directly an optimization

problem (classif. SVM, separable data)

Minimize ||w||²,
with (yi= +/-1)
and yi(<w|xi> + b) >=1 for all i=1,..,L

© Eric Xing @ CMU, 2006-2015 36

The VC Dimension of SVMs
 An SVM finds a linear separator in a Hilbert space, where the
original date x can be mapped to via a transformation (x).
( )
( ) ( )
( ) ( ) ( )
(.) ( )
( ) ( )
( ) ( )
( ) ( )
( ) ( ) ( )
( )
( )

Input space Feature space

 Recall that the kernel trick used by SVM alleviates the need to
find explicit expression of (.) to compute the transformation

© Eric Xing @ CMU, 2006-2015 37

The Kernel Trick
 Recall the SVM optimization problem
1 m
m
max  J ( )    i    i j yi y j (xTi x j )
i 1 2 i , j 1
s.t. 0   i  C , i  1,, k
m

 y
i 1
i i  0.

 The data points only appear as inner product

 As long as we can calculate the inner product in the feature
space, we do not need the mapping explicitly
 Define the kernel function K by

© Eric Xing @ CMU, 2006-2015 38

Mercer’s Condition
 For which kernels does there exist a pair {H;(.)} with the
valid geometric properties (e.g., nonnegative dot-product) for
a transformation satisfied, and for which does there not?

 Mercer’s Condition for Kernels

 There exists a mapping (.) and an expansion

iff for any g(x) such that

then

© Eric Xing @ CMU, 2006-2015 39

The VC Dimension of SVMs
 We will call any kernel that satisfies Mercer’s condition a
positive kernel, and the corresponding space H the
embedding space.

 We will also call any embedding space with minimal

dimension for a given kernel a “minimal embedding space”.

 Theorem: Let K be a positive kernel which corresponds to a

minimal embedding space H. Then the VC dimension of the
corresponding support vector machine (where the error
penalty C is allowed to take all values) is dim(H) + 1

© Eric Xing @ CMU, 2006-2015 40

VC and the Actual Risk

 It is striking that the two curves have minima in the same

place: thus in this case, the VC bound, although loose, seems
to be nevertheless predictive.
© Eric Xing @ CMU, 2006-2015 41
What You Should Know
 Sample complexity varies with the learning setting
 Learner actively queries trainer
 Examples provided at random

 Within the PAC learning setting, we can bound the probability that
learner will output hypothesis with given error
 For ANY consistent learner (case where c in H)
 For ANY “best fit” hypothesis (agnostic learning, where perhaps c not in H)

 VC dimension as measure of complexity of H

 Quantitative bounds characterizing bias/variance in choice of H
 but the bounds are quite loose...

 Mistake bounds in learning

 Conference on Learning Theory: http://www.learningtheory.org

© Eric Xing @ CMU, 2006-2015 42