Computational learning theory

1 PAC Learning

We want to develop a theory to relate the probability of successful learning,

the number of training examples, the complexity of the hypothesis space, the
accuracy to which the target concept is approximated, and the manner in which
training examples are presented.

1.1 Prototypical Concept Learning

Consider instance space X, the set of examples, and concept space C, the set of
target functions that could have generated the examples such that there exists
a f C∈ that is the hidden target function. For example, C could be all n-
conjunctions, all n-dimensional linear functions, etc.
The hypothesis space is the set of all possible hypotheses that our learning
algorithm can choose from, where H is not necessarily equal to C. We consider
our training instances to be S×0,{ 1 }– including both positive and negative
examples of the target concept – such that training instances are generated by
a fixed unknown probability distribution D over X. Each training instance can
be thought of as a (data, label) tuple, as below

S = [(x1, f (x1)), (x2, f (x2))...(xn, f (xn))]

In this setting, our goal is to determine a hypothesis h∈ H that estimates f ,

evaluated by its performance on subsequent instances ∈ x X drawn according
to D.
Note the assumption that both training and testing instances are drawn from
from the same distribution D, as it is important in the analysis that follows.

1.2 Intuition

Consider Figure 1, showing the space predicted by target function f and hy-
pothesis function h, where points inside the circle are positive, points outside
are negative, and the functions are given by

h = x1 ∧ x2 ∧ x3 ∧ x4 ∧ x5 ∧ x100

Computational Learning Theory-1

 f = x2 ∧ x3 ∧ x4 ∧ x5 ∧ x100

Figure 1: f not equal h

In this example, we have seen x1 in all positive training instances (even though
it is not active in f ). Therefore, it is very likely that it will be active in future
positive examples. If not, it is active in only a small percentage of examples, so
the error should be small.
We can therefore consider the error as the probability of an example having dif-
ferent labels according to the hypothesis and the target function, given by

ErrorD = Prx∈D[f (x) /= h(x)]

2 Conjunctions

Consider z to be a literal in a conjunction. Let p(z) be the probability that

a D-sampled example is positive and z is false in it. In the example above,
z = x1 .

2.1 Error Bounds

Σ
Claim: Error(h) ≤ p(z)
z∈h

Proof
Consider that p(z) is the probability that a randomly chosen example is positive
and z is deleted from h1.
If z is in the target concept, then p(z) = 0; f is a conjunction and thus z can
never be false in a positive example if z ∈ f .
h will make mistakes only on positive examples. A mistake is made only if z
that is in h but not in f . In such cases, when z = 0, h will predict a negative
example while f indicates a positive example.
1Recall that h is a conjunction, and z is a literal, such that if – during training – we see a

positive example where z = 0, z is removed from h

Computational Learning Theory-2

Therefore p(z) is also the probability that z causes h to make a mistake on a
randomly drawn example from D. In a single example there could be multiple
literals that are incorrect, but since in the worst case they are seen one-by-one,
the sum of z bounds the error of h.
We can also consider literal z ∈h to be bad when p(z) > є n. A bad literal,
therefore, is a literal that is not in the target concept but has a significant
probability to appear false in a positive example.
Claim: It there are no bad literals, then Error(h) < ϵ
Proof
Σ
We have already stated that Error(h) ≤ p(z). Given that |h| ≤ n – the
z∈h
hypothesis has at most the same number of literals as there are features – then
Σ
n
є
we know that the error cannot exceed n
= ϵ.
Therefore Error(h) < ϵ.

2.2 Example Bounds

Let z be a bad literal. We want to determine the probability that z has not
been eliminated from h after seeing a given number of examples.

P (z survives one example) = 1 − P (z eliminated by one example)

≤ 1 − p(z) (1)
ϵ
< 1−
n
We can intuit that as we see more examples, the probability that bad literal z
survives decreases exponentially, given by
ϵ m
p(z survives m independent examples) = (1 − p(z))m < (1 − )
n

This is for one literal z. There are at most n bad literals, thus the probability
that some bad literal survives m examples is bounded by n(1 −n є )m
We want this probability to be bounded by δ, and thus we must choose m to
be sufficiently large. Consider that we want
ϵ m
n(1 − ) <δ
n
Using 1 − x < e−x, it is sufficient to require
−mе
ne n <δ
Therefore, we need
n 1
m> {ln(n) + ln( )}
ϵ δ

Computational Learning Theory-3

to guarantee the probability of failure (Error > ϵ) is less than δ
With probability > 1 − δ, there are no bad literals, i.e., Error(h) < ϵ

3 Formulating Prediction Theory

Given
• Instance space X
• Output space Y = {−1, +1}
• Distribution D, which is unknown over X × Y
• Training examples S, where each is drawn independently from D (|S| = m
we can define the following
• True Error: ErrorD = Pr(x,y)∈D[h(x)¬ = y]
Σ
• Empirical Error: ErrorS = P r(x,y)∈S [h(x)¬ = y] = [h(xi )¬ = y i ]
1,m

• Function space C, or set of possible target concepts, where f ∈ C : X → Y

• Set of possible hypotheses H
We cannot expect a learner to learn a concept exactly. There may be many
concepts consistent with the available data, and unseen examples may have any
label. Thus we must agree to misclassify uncommon examples that were not
seen during training.
Further, we cannot always expect a learner to learn a close approximation to
the target concept, since sometimes the training set does not represent unseen
examples.
Therefore, the only realistic expectation of a good learner is that it will learn a
close approximation to the target concept with high probability.

3.1 Probably Approximately Correct Learning

In Probably Approximately Correct (PAC) learning, one requires that given

small parameters ϵ and δ – with probability at least (1−δ) – a learner produces
a hypothesis with error at most ϵ.
This notion relies on the Consistent Distribution Assumption: there is one prob-
ability distribution D that governs both training and testing examples.

Computational Learning Theory-4

3.2 PAC Learnability

Consider a concept class C defined over an instance space X, and a learner L

using a hypothesis space H.
C is PAC learnable by L using H
if ∀f ∈C, ∀D over X, and fixed 0 < ϵ, δ < 1, L – given a collection of m
examples sampled independently according to D – produces with probability at
least 1 − δ a hypothesis h ∈ H with error at most ϵ where m is polynomial in
є , δ , n and |H|.
1 1

C is efficiently PAC learnable

if L can produce the hypothesis in time polynomial in 1 , 1 , n and size(H).
є δ

Two Limitations
• Polynomial sample complexity, which is also called information theoretic
constraint, governs if there is enough information in the sample to distin-
guish a hypothesis h that approximate f .
• Polynomial time complexity, also called computational complexity, which
tells if there is an efficient algorithm that can process the sample and
produce a good hypothesis h.
To be PAC Learnable, there must be a hypothesis h ∈ H with arbitrary small
error for every f ∈ C. We generally assume H is a super set of C.
The worst definition is that the algorithm must meet its accuracy for every
distribution and every target function f ∈ C.

3.3 Occam’s Razor

We want to prove the general claim that smaller hypothesis spaces are bet-
ter.
Claim: The probability that there exists a hypothesis h ∈ H that is consistent
with m examples and satisfies Error(h) > ϵ is less then |H|(1 − ϵ)m.
Proof
Let h be a bad hypothesis. The probability that h is consistent with on examples
is less than 1−ϵ. Since the m examples are independently drawn, the probability
that h is consistent with m examples is less than (1 − ϵ)m.
The probability that any one of the hypothesis in H is consistent with m exam-
ples is less than |H|(1 − ϵ)m.
Given this fact, we now want this probability to be smaller than δ, that is

|H|(1 − ϵ)m < δ

Computational Learning Theory-5

 ln(|H|) + m ln(1 − ϵ) < ln(δ)

With the fact that e−x > 1 − x, we have

1 1
m> {ln(|H|) + ln( )}
ϵ δ

This is called Occam’s razor, because it indicates a preference towards small

hypothesis space, i.e., if you have small hypothesis space, you do not have to
see too many examples.
There is also a trade-off of the hypothesis space. If the space is small, then it
generalizes well, but it many not be expressive enough.

4 Consistent Learners
Using the results from the previous section, we can get this general scheme for
PAC learning:
Given a sample of m examples, find some h ∈ H that is consistent with all m
examples. If m if large enough, a consistent hypothesis will be sufficiently close
to f . We can then check that m scales polynomially in the relevant parameters
(i.e. m is not too large). ”Closeness” guarantees
1 1
m> (ln |H| + ln )
ϵ δ

In the case of conjunctions, we used the elimination algorithm, which results in

a hypothesis h that is consistent with the training set, and we showed directly
that if we have sufficiently many examples (polynomial in relevant parameters),
then h is close to the target function.

4.1 Examples

Conjunctions
For conjunctions, the size of the hypothesis space is 3n, since there are 3 possible
values for each of the n features (appear negative, positive, or not at all. There-
fore, the number of examples we need according to the PAC learning framework,
m, is given by
1 n 1 1 1
m > ϵ { ln(3 ) + ln( δ )} = ϵ {n ln 3 + ln( δ )}

Thus, if we want to guarantee a 95% chance of learning a hypothesis (1 − δ) of

at least 90% accuracy (1 − ϵ), with n = 10 boolean variables, m > 140.

Computational Learning Theory-6

If we change to n = 100, m > 1130, which shows m is linear with n.
However, changing the confidence (1−δ) to 99% makes m > 1145, shows m is
logarithmic with δ.
k-CNF
Consider Conjunctive Normal Form functions (CNFs), which can express any
boolean function. Recall that CNFs are conjuctions of disjunctions. A subset of
this class is that of k-CNF, where each disjunction contains k terms, as in

f = ∧ i=1 (li1 ∨ li2 ∨ ... ∨ lik )

m

To determine if we can learn such a class of functions, we must know the size
k
of this hypothesis space. In this case, the hypothesis space is given by 2(2n) ,
corresponding to the number of ways to choose subsets from among the k literals,
including negations. Thus, the sample complexity is given by

ln(|k − CNF |) = O(nk)

Since k is fixed, we have an order polynomial in the number of examples and
thus h is guaranteed to be PAC learnable. Next step is to learn a consistent
hypothesis.
Now we must consider how to learn such a hypothesis. Using what we know
now, we cannot learn this directly. We can learn k-CNFs, however, if we move
to a new space.
Consider the example in which n = 4 and k = 2 for a monotone k-CNF. Here,
there are six disjunctions for which we can create a new mapping from the orig-

y1 = x1 ∨ x2 y2 = x1 ∨ x3
y3 = x1 ∨ x4 y4 = x2 ∨ x3
y5 = x2 ∨ x4 y6 = x3 ∨ x4

inal space to a new space with six features: (0000, 1) → (000000, 1)

(1010, 1) → (111101, 1)
(1110, 1) → (111111, 1)
(1111, 1) → (111111, 1)
Now we can apply a standard algorithm for learning monotone conjunctions.
Unbiased Learning
Consider the hypothesis space of all boolean function on n features. There
n
are 22 different functions and the bound (ln( |H |) is therefore exponential in
n
2 , which means that in general the set of all boolean functions are not PAC
learnable.
k-Clause CNF
Conjunctions of at most k disjunctive clauses.

f = C1 ∧ C2 ∧ ...Ck; Ci = l1 ∨ l2... ∨ lm

Computational Learning Theory-7

 | H| ) = O(kn) is linear in n and thus PAC
The size of the hypothesis space ln(
learnable.
k-DNF
Disjunctions of any number of terms where each conjunctive term has at most
k literals.
f = T1 ∨ T2... ∨ Tm; Ti = l1 ∧ l2 ∧ ... ∧ lm

4.2 k-term DNF Computational Complexity

Consider the class of k-term DNFs, or disjunctions of at most k conjunctive

terms. From the sample complexity perspective, we should be able to learn
in the same way as with k-CNFs, but computational complexity is challeng-
ing.
Consider a 2-term DNF consistent with a set of training data is NP-hard. Thus,
even though 2-term DNFs are PAC learnable, they are not efficiently PAC
learnable.
We can address this by enlarging the hypothesis space. If the hypothesis we
wish to learn can be represented in a larger hypothesis space that we know is
learnable, we can learn our desired hypothesis. In this case, we can represent k-
term DNFs as k-CNFs, since k-CNF is a superset of k-term DNF. Consider a 3-term
DNF (left) and its equivalent 3-CNF (right).
Y
T1 ∨ T2 ∨ T3 = {x ∨ y ∨ z}
x∈T1,y∈T2,z∈T3

Representation is important. Concepts that cannot be leaned using one repre-

sentation may be learned using another more expressive representation.
However, this leaves us with two problems:
How can we learn when data is not completely consistent with training
data?
How can we learn in an infinite hypothesis space?

5 Agnostic Learning

Assume we are trying to learn concept f using hypothesis space H, but f ∈ /

H. We therefore cannot learn a a completely consistent hypothesis, and thus
our goal is to find a hypothesis h ∈ H that has as small training error as
possible.
m
ErrT R 1 Σ
= m f (x i ) /= h(xi )
i

Computational Learning Theory-8

where xi is the ith training example.
We want to guarantee that a hypothesis with a small training error will have
good accuracy on unseen examples, and one way to do so is with Hoeffding
bounds. This characterizes the deviation between the true probability of some
event and its observed frequency over m independent trails.
2
P r[p > p̂ + ϵ] < e−2mє

To understand the intuition, consider tossing a biased coin. The more tosses,
the more likely the observed result will correspond with the expected result.
Similarly, the probability that an element in H will have training error which is
off by more than ϵ can be bounded as follows:
2
Pr[ErrD(h) > ErrT R(h) + ϵ] < e −2mє
2
If we consider δ = |H|e−2mє , we can get a generalization bound, or how much
will the true error ED deviate from the observed (training) error ET R.
For any distribution D, generating training and test instances with probability
at least 1 − δ over the choice of the training set of size m, (drawn i.i.d.), for all
s
h∈H
log |H| + log( 1 )
δ
ErrorD(h) < ErrorT R(h) +
2m

An agnostic learner which makes no commitment to whether f is in H returns

the hypothesis with least training error over at least the following number of
examples m can guarantee with probability at least 1− δ that its training error
is not off by more than ϵ from the true error. We therefore require a number of
examples given by
1 1
m > 2 {ln(|H|) + ln( )}
2ϵ δ
Learnability depends on the log of the size of the hypothesis space.

6 VC Dimension

For both consistent and agnostic learners, we assumed finite hypothesis spaces.
We know consider an infinite hypothesis space.

6.1 Learning Rectangles

Consider a target concept as an axis parallel rectangle (positive points inside,

negative outside, as given by Figure 2.

Computational Learning Theory-9

 Figure 2: Learning Rectangles

We can simply choose the maximum x and y values as well as the minimum x
and y values of the positive examples as the boundary for the rectangle. This is
generally a good algorithm because it learns efficiently, but we cannot use the
theorem from before to derive a bound because the hypothesis space is infinitely
large.
Therefore, we need to find out how to derive a bound given an infinitely large
hypothesis space.

6.2 Infinite Hypothesis Space

Just as before, where we discussed small hypothesis spaces (conjunctions) and

large hypothesis spaces (DNFs), some infinite hypothesis spaces are larger (more
expressive) than others; rectangles and general convex polygons have different
levels of expressiveness.
We therefore need to measure the expressiveness of an infinite hypothesis space.
The Vapnik-Chervonenkis dimension – or VC dimension – provides such a
measure. Analogous to H | ,| there are bounds for sample complexity using
V C(H).

6.3 Shattering

The key notion behind VC-dimension is that of shattering. Assume a set of

points. We want to use a function to separate all possible labelings of the set
of points. In Figure 3, a linear function (green) can separate the two points,
regardless of how they’re labeled.
For two points on a plane, there are two different ways of labeling the two points.
A line can separate those two points no matter how they are labeled.

Computational Learning Theory-10

 Figure 3: A linear function can shatter 2 points

Figure 4: A set of multiple points

However, if there are many points in the set on a plane, as in Figure 4, no

straight line can separate any labeling of these points. Linear functions are not
expressive enough to shatter 13 points, but more expressive functions are.
We say that a set S of examples is shattered by a set of functions H if – for
every partition of the examples in S into positive and negative examples – there
is a function in H that gives exactly there labels to the examples. The intuition
is that a rich set of functions shatters a large set of points.

(a) A function example (b) function fails to shatter

Consider the function in which left bounded intervals on the real axis for some
number is positive ([0, a), for some a > 0).
It is trivial for this function to shatter a single point. However, in any set of two
points on the line, the left can be labeled negative and the right positive, which

Computational Learning Theory-11

this H cannot label correctly. Thus, left bounded intervals cannot shatter two
points.
Similarly, if we consider the class of functions for which real numbers b > a and
points within [a, b] is positive, all sets of one or two points are shatterable, but
no set of three points can be shattered.

Figure 6: Half space

As a final example, consider half-spaces on the plane. We can trivially shatter

all sets of one point and two points. We can shatter some sets of three points,
but cannot shatter any set of four points. If the 4 points form a convex polygon,
then by labeling each point different from its neighbors, the four points cannot
be shattered. If, on the other hand, they do not form a convex polygon, then
one point is inside the convex hull defined by the other three, and if that point
is negative, there is no way to shatter them.

6.4 Definition

An unbiased hypothesis space H shatters the entire instance space X if it is able

to induce every possible partition on the set of all possible instances.
The larger the subset X that can be shattered, the more expressive a hypothesis
space is (i.e. less biased). The VC dimension of hypothesis space H over instance
space X is the size of the largest finite subset X (even if there is only one subset)
that is shattered by H.
If there exists a subset of size d that can be shattered, then V C(H) ≥ d.
If no subset of size d can be shattered, then V C(H) < d.
• V C(Half intervals) = 1; No subset of size 2 can be shattered
• V C(Intervals) = 2; No subset of size 3 can be shattered
• V C(Half-spaces in the plane) = 3; No subset of size 4 can be shattered

Computational Learning Theory-12

6.5 Sample complexity and VC dimension

VC dimension serves the same role as the size of the hypothesis space. Using
VC dimension as a measure of expressiveness, we can give an Occam algorithm
for infinite hypothesis spaces.
Given a sample D of m examples we will find h ∈ H that is consistent with all
m examples, if
1 13 2
m > {8V C(H) log + 4 log( )}
ϵ ϵ δ
then with probability at least 1 − δ, h has error less then ϵ.
We consider that the hypothesis space has to be infinite if we want to use this
bound. If we want to shatter m points, then H has to be at least 2m in order
to shatter any configurations of those m examples.
Thus |H| > 2m , log(|H|) ≥ V C(H).

6.6 Axis-Parallel Rectangles, Continued

Consider again the problem of learning axis-parallel rectangles. To determine

if axis-parallel rectangles are PAC-learnable, we must determine sample com-
plexity. Here, we show that V C(H) ≥ 4. In Figure 7a, it is trivial to find an

(a) A set that can be shattered (b) a set that cannot be shattered

axis-parallel rectangle to shatter the points, regardless of their labeling. Though

Figure 7b illustrates a set of four that is not shatterable, it is sufficient to find
any set of four to prove the VC dimension is greater or equal to 4.
Next we need to argue that no set of five points can be shattered. For any layout
of five points, we just need to show one kind of labeling that cannot be shattered.
In this case, we can say that – of the five points – there must be a minimum and
maximum x and y. There must by definition be a point within those minimum
and maximum bounds, and thus by labeling the extreme four points as positive
but the fifth, internal point as negative, axis-parallel rectangles cannot shatter
five points.

Computational Learning Theory-13

Thus, from sample complexity perspective, axis-parallel rectangles are PAC
learnable.
To determine if this hypothesis class is efficiently PAC learnable, we need an
efficient algorithm to find the rectangle. Here, we can find the smallest example
rectangle that contains all positive examples. This is likely not the best rectangle
– it cannot generalize to new positive examples – so in the ideal case we would
like a margin, that is, a rectangle slightly larger than the minimum one we’ve
seen during training.
Given such an algorithm, we have shown that axis-parallel rectangles are effi-
ciently PAC learnable.

6.7 Sample Complexity Lower Bound

We’ve discussed upper bounds on the number of examples; that is, if we have
seen m examples, we can be reasonably sure our algorithm will perform well on
new examples. There is also a general lower bound on the minimum number of
examples necessary for PAC learning.
Consider any concept class C such that V C(C) > 2. For any learner L and
small enough ϵ, δ, there exists a distribution D and a target function in C such
that if L observes less than
1 1 V C(C) − 1
m = max[ log( ), ]
ϵ δ 32ϵ
examples, then with probability at least δ, L outputs a hypothesis having
error(h) > ϵ.
This is the inverse of the bound algorithm we have seen before.
Ignoring constant factors, the lower bound is the same as the upper bound,
except for the extra log( 1є ) factor in the upper bound.

7 Conclusion

The PAC framework provides a reasonable model for theoretically analyzing the
effectiveness of learning algorithms.
We discussed that the sample complexity for any consistent learner using the
hypothesis space, H, can be determined from a measure of Hs expressiveness
(|H|, V C(H)).
We discussed consistent and agnostic learners, showing that the log of the size
of a finite hypothesis space is most important, and then extended this notion
to the infinite hypothesis space.

Computational Learning Theory-14

We also discussed sample and computational complexity, showing that if sample
complexity is tractable, the computational complexity of finding a consistent
hypothesis governs the complexity of the learning problem.
Many additional models have been studied as extensions of the basic one: learn-
ing with noisy data, learning under specific distributions, learning probabilistic
representations, etc..
An important extension is PAC-Bayesians theory, where the idea is that – rather
than simply assume that training and test are governed by the same distribution
– assumptions are also made about the prior distribution over the hypothesis
space.
It’s important to note that though the bounds we compute are loose, they
can still guide model selection. A lot of recent work is on data dependent
bounds.
The impact COLT has had on practical learning system in the last few years has
been very significant: SVMs, Winnow (Sparsity), Boosting, and Regularization,
to name a few.

Computational Learning Theory-15