The Learning Problem and Regularization

9.520 Class 02, 13 February 2006

Tomaso Poggio

 Plan

• Learning as function approximation

• Empirical Risk Minimization
• Generalization and Well-posedness
• Regularization
• Appendix: Sample and Approximation Error

 About This Class

Theme We introduce the learning problem as the problem

of function approximation from sparse data. We define
the key ideas of loss functions, empirical error and gen-
eralization error. We then introduce the Empirical Risk
Minimization approach and the two key requirements
on algorithms using it: well-posedness and consistency.
We then describe a key algorithm – Tikhonov regular-
ization – that satisfies these requirements.

Math Required Familiarity with basic ideas in probability

theory.
 Data Generated By A Probability

Distribution

We assume that X and Y are two sets of random variables.

We are given a training set S consisting n samples drawn

i.i.d. from the probability distribution μ(z) on Z = X × Y :

(x1, y1), . . . , (xn, yn)

that is z1, . . . , zn

We will make frequent use of the conditional probability

of y given x, written p(y|x):
μ(z) = p(x, y) = p(y|x) · p(x)

It is crucial to note that we view p(x, y) as fixed but un-

known.
 Probabilistic setting

X Y

P(y|x)

P(x)
 Hypothesis Space

The hypothesis space H is the space of functions that we

allow our algorithm to provide. For many algorithms (such
as optimization algorithms) it the space the algorithm is
allowed to search. As we will see, it is often important to
choose the hypothesis space as a function of the amount
of data available.
Learning As Function Approximation From

Samples: Regression and Classification

The basic goal of supervised learning is to use the training

set S to “learn” a function fS that looks at a new x value
xnew and predicts the associated value of y:

ypred = fS (xnew )

If y is a real-valued random variable, we have regression.

If y takes values from an unordered finite set, we have

pattern classification. In two-class pattern classification
problems, we assign one class a y value of 1, and the other
class a y value of −1.
 Loss Functions

In order to measure goodness of our function, we need a

loss function V . In general, we let V (f, z) = V (f (x), y)
denote the price we pay when we see x and guess that the
associated y value is f (x) when it is actually y.
 Common Loss Functions For Regression

For regression, the most common loss function is square

loss or L2 loss:

V (f (x), y) = (f (x) − y)2

We could also use the absolute value, or L1 loss:

V (f (x), y) = |f (x) − y |

Vapnik’s more general -insensitive loss function is:

V (f (x), y) = (|f (x) − y| − )+

Common Loss Functions For Classification

For binary classification, the most intuitive loss is the 0-1 loss:
V (f (x), y) = Θ(−yf (x))

where Θ(−yf (x)) is the step function. For tractability and other rea-
sons, we often use the hinge loss (implicitely introduced by Vapnik) in
binary classification:
V (f (x), y) = (1 − y · f (x))+
 The learning problem: summary so far

There is an unknown probability distribution on the prod-

uct space Z = X × Y , written μ(z) = μ(x, y). We assume
that X is a compact domain in Euclidean space and Y a
closed subset of IR.

The training set S = {(x1, y1), ..., (xn , yn)} = {z1, ...zn}
consists of n samples drawn i.i.d. from μ.

H is the hypothesis space, a space of functions f : X → Y .

A learning algorithm is a map L : Z n → H that looks

at S and selects from H a function fS : x → y such that
fS (x) ≈ y in a predictive way.
 Empirical error, generalization error,

generalization

Given a function f , a loss function V , and a probability distribution μ

over Z, the expected or true error of f is:
�
I[f ] = IEz V [f, z] = V (f, z)dμ(z)
Z

which is the expected loss on a new example drawn at random from

μ.

We would like to make I[f ] small, but in general we do not know μ.

Given a function f , a loss function V , and a training set S consisting

of n data points, the empirical error of f is:

1�
IS [f ] = V (f, zi)
n
 Empirical error, generalization error,

generalization

A very natural requirement for fS is distribution independent general-

ization

∀μ, lim |IS [fS ] − I[fS ]| = 0 in probability

n→∞

In other words, the training error for the solution must converge to the
expected error and thus be a “proxy” for it. Otherwise the solution
would not be “predictive”.

A desirable additional requirement is universal consistency

� �
∀ε > 0 lim sup IPS I[fS ] > inf I[f ] + ε = 0.
n→∞ μ f ∈H

Remark: We will add in future classes the requirement of uniform

convergence in both definitions.
 A reminder: convergence in probability

Let {Xn} be a sequence of bounded random variables. We

say that

lim Xn = X in probability
n→∞

if

∀ε > 0 lim IP{|Xn − X| ≥ ε} = 0.

n→∞

or

if for each n there exists a εn and a δn such that

IP {|Xn − X| ≥ εn} ≤ δn,
with εn and δn going to zero for n → ∞.
 3. ERM and conditions for generalization

(and consistency)

Given a training set S and a function space H, empirical risk minimiza-

tion (Vapnik) is the algorithm that looks at S and selects fS as

fS = arg min IS [f ]
f ∈H

The solution to this problem does not in general exhibit generalization

and the problem may be ill-posed, depending on the choice of H.

If the minimum does not exist we can work with the infimum.

Notice: For ERM generalization and consistency are equivalent

 Generalization and Well-posedness of

Empirical Risk Minimization

For the solution of ERM to be useful in the context of

learning, the solution must

• “generalize”

• exist, be unique and be “stable” (well-posedness).

 Here is a graphical example for

generalization: given a certain number of

samples...

f(x)

x
suppose this is the “true” solution...

f(x)

x
... but suppose ERM gives this solution!

f(x)

x
 How can I guarantee that for a sufficient

number of examples the ERM solution will

converge to the true one?

f(x)

x
Classical conditions for consistency of ERM

Uniform Glivenko-Cantelli Classes

L = {H ◦ V } is a (weak) uniform Glivenko-Cantelli (uGC) class

if

� �
∀ε > 0 lim sup IPS sup |I[] − IS []| > ε = 0.
n→∞ μ ∈L

ˇ
Theorem [Vapnik and Cervonenkis (71), Alon et al (97), Dudley, Gin´
e, and Zinn
(91)]

A necessary and sufficient condition for consistency of ERM is that L is uGC.

Thus, as we will see later, a proper choice of the hypothesis space H ensures gen-
eralization of ERM (and consistency since for ERM generalization is necessary and
sufficient for consistency and viceversa). We will be exploring the uGC definition
(and equivalent definitions) in detail in 9.520.
 Well-posedness of ERM

ERM is in general an ill-posed problem. It can be made

well-posed by an appropriate choice of H.

As we will see later, well-posedness is mainly used to mean

stability of the solution: fS depends continuously on the
training set S. In particular, changing one of the training
points should affect less and less the solution as n goes to
infinity.
 General definition of Well-Posed and

Ill-Posed problems

A problem is well-posed if its solution:

• exists
• is unique
• depends continuously on the data (e.g. it is stable)

A problem is ill-posed if it is not well-posed.

Here is a graphical example for stability:

given 10 samples...

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
...we can find the smoothest interpolating

polynomial (which degree?).

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

But if we perturb the points slightly...

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

 ...the solution changes a lot.

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
If we restrict ourselves to degree two

polynomials...

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
...the solution varies only a small amount

under a small perturbation.

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
 Regularization

The basic idea of regularization (originally introduced in

-
dependently of the learning problem) is to restore well
-
posedness of ERM by constraining the hypothesis space

H. The direct way – minimize the empirical error subject

to f in a ball in an appropriate H – is called Ivanov reg-
ularization. The indirect way is Tikhonov regularization
(which is not ERM).
 Ivanov and Tikhonov Regularization

ERM finds the function in (H,

·
) which minimizes
n
1 �
V (f (xi ), yi)
n i=1
which in general – for arbitrary hypothesis space H – is
ill-posed. Ivanov regularizes by finding the function that
minimizes
n
1 �
V (f (xi ), yi)
n i=1

while satisfying

f
2 ≤ A
Alternatively, Tikhonov regularization minimizes over the
hypothesis space H, for a fixed positive parameter λ, the
regularized functional
n
1 �
V (f (xi ), yi) + γ
f
2
K, (1)
n i=1
where
f
is the norm in H – the Reproducing Kernel
Hilbert Space (RKHS), defined by the kernel K.
 Well-posed and Ill-posed problems

Hadamard introduced the definition of ill-posedness. Ill-posed prob-

lems are typically inverse problems.
As an example, assume g is a function in Y and u is a function in
X, with Y and X Hilbert spaces. Then given the linear, continuous
operator L, consider the equation
g = Lu.
The direct problem is is to compute g given u; the inverse problem is
to compute u given the data g. In the learning case L is somewhat
similar to a “sampling” operation.
The inverse problem of finding u is well-posed when

• the solution exists,

• is unique and

• is stable, that is depends continuously on the initial data g.

Ill-posed problems fail to satisfy one or more of these criteria. Often

the term ill-posed applies to problems that are not stable, which in a
sense is the key condition.
 Tikhonov Regularization

As we will see in future classes

• Tikhonov regularization ensures well-posedness eg ex-

istence, uniqueness and especially stability (in a very
strong form) of the solution

• Tikhonov regularization ensures generalization

• Tikhonov regularization is closely related to – but dif-

ferent from – Ivanov regularization, eg ERM on a hy-
pothesis space H which is a ball in a RKHS.
 Next Class

• In the next class we will introduce RKHS: they will be

the hypothesis spaces we will work with.

• We will also derive the solution of Tikhonov regular-

ization.
 Appendix: Target Space, Sample and

Approximation Error

In addition to the hypothesis space H, the space we allow

our algorithms to search, we define...

The target space T is a space of functions, chosen a priori

in any given problem, that is assumed to contain the “true”
function f0 that minimizes the risk. Often, T is chosen to
be all functions in L2, or all differentiable functions. Notice
that the “true” function if it exists is defined by μ(z), which
contains all the relevant information.
Sample Error (also called Estimation Error)

Let fH be the function in H with the smallest true risk.

We have defined the generalization error to be IS [fS ] − I[fS ].

We define the sample error to be I[fS ] − I[fH], the difference in true

risk between the best function in H and the function in H we actually
find. This is what we pay because our finite sample does not give us
enough information to choose to the “best” function in H. We’d like
this to be small. Consistency – defined earlier – is equivalent to the
sample error going to zero for n → ∞.

A main topic of this course is “bounding” the generalization error.

Another topic – the main one in classical learning theory and statis-
tics – is bounding the sample error, that is determining conditions
under which we can state that I[fS ] − I[fH] will be small (with high
probability).

As a simple rule, we expect that if H is “well-behaved”, then, as n

gets large the sample error will become small.
 Approximation Errror

Let f0 be the function in T with the smallest true risk.

We define the approximation error to be I[fH ] −I[f0], the

difference in true risk between the best function in H and
the best function in T . This is what we pay because H is
smaller than T . We’d like this error to be small too. In
much of the following we can assume that I[f0] = 0.

We will focus less on the approximation error in 9.520, but

we will explore it.

As a simple rule, we expect that as H grows bigger, the

approximation error gets smaller. If T ⊆ H – which is a
situation called the realizable setting –the approximation
error is zero.
 Error

We define the error to be I[fS ] − I[f0], the difference in

true risk between the function we actually find and the
best function in T . We’d really like this to be small. As we
mentioned, often we can assume that the error is simply
I[fS ].

The error is the sum of the sample error and the approxi-
mation error:

I[fS ] − I[f0] = (I[fS ] − I[fH]) + (I[fH ] − I[f0])

If we can make both the approximation and the sample

error small, the error will be small. There is a tradeoff
between the approximation error and the sample error...
 The Approximation/Sample Tradeoff

It should already be intuitively clear that making H big

makes the approximation error small. This implies that we
can (help) make the error small by making H big.

On the other hand, we will show that making H small will

make the sample error small. In particular for ERM, if H is
a uGC class, the generalization error and the sample error
will go to zero as n → ∞, but how quickly depends directly
on the “size” of H. This implies that we want to keep H
as small as possible. (Furthermore, T itself may or may
not be a uGC class.)

Ideally, we would like to find the optimal tradeoff between

these conflicting requirements.
 Error Definition: a caveat

We define the error to be I[fS ]−I[f0]. In the literature, the

true risk of the function we find, I[fS ] is sometimes called
the error. In the case where I[f0] = 0, the two definitions
are equivalent.
 ...mapping notation and results in

CuckerSmale...

(f ) ←→ I(f )

z (f ) ←→ IS (f )

Thus
Lz ←→ I(f ) − IS (f )

For ERM
fz ←→ fS

Theorem B (for H compact) ←→ generalization, see Theorem b (Na-

ture; for general algorithms and general H)

Theorem C (eg H (fz ) → 0) ←→ Theorem a (Nature; consistency of

ERM) where H(f ) = (f ) − (fH ),