Scribd
Upload
35 views4 pages

How Many Samples To Learn A Finite Class?

1. The lecture discusses Probably Approximately Correct (PAC) learning and how it relates to concepts, hypotheses, and samples. PAC learning involves learning a concept from samples in a way that a hypothesis can predict future samples with high probability. 2. It explains that the number of samples needed for learning is related to the size of the concept class and its VC dimension. For a finite concept class C, O(|C|/ε log(1/δ)) samples are enough, while for infinite classes the VC dimension is important - if it is finite, then learning is possible. 3. Computational complexity is discussed - finding a hypothesis consistent with samples could be NP-hard, though improper learning avoids this

Uploaded by

Muhammad Al Kahfi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
35 views4 pages

How Many Samples To Learn A Finite Class?

1. The lecture discusses Probably Approximately Correct (PAC) learning and how it relates to concepts, hypotheses, and samples. PAC learning involves learning a concept from samples in a way that a hypothesis can predict future samples with high probability. 2. It explains that the number of samples needed for learning is related to the size of the concept class and its VC dimension. For a finite concept class C, O(|C|/ε log(1/δ)) samples are enough, while for infinite classes the VC dimension is important - if it is finite, then learning is possible. 3. Computational complexity is discussed - finding a hypothesis consistent with samples could be NP-hard, though improper learning avoids this

Uploaded by

Muhammad Al Kahfi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

6.080/6.

089 GITCS

1 April 2008

Lecture 20
Lecturer: Scott Aaronson

Scribe: Georey Thomas

Probably Approximately Correct Learning


In the last lecture, we covered Valiants model of Probably Approximately Correct (PAC) learn
ing. This involves:
S:
A sample space (e.g., the set of all points)
D:
A sample distribution (a probability distribution over points in the sample space)
c : S {0, 1}: A concept, which accepts or rejects each point in the sample space
C:
A concept class, or collection of concepts
For example, we can take our sample space to be the set of all points on the blackboard, our
sample distribution to be uniform, and our concept class to have one concept corresponding to
each line (where a point is accepted if its above the line and rejected if its below it). Given a set
of points, as well as which points are accepted or rejected, our goal is to output a hypothesis that
explains the data: e.g., draw a line that will correctly classify most of the future points.
A bit more formally, theres some true concept c C that were trying to learn. Given
sample points x1 , . . . , xm , which are drawn independently from D, together with their classications
c(x1 ), . . . , c(xm ), our goal is to nd a hypothesis h C such that
Pr [h(x) = c(x)] 1
. Furthermore, we want to succeed at this goal with probability at least 1 over the choice of
xi s. In other words, with high probability we want to output a hypothesis thats approximately
correct (hence Probably Approximately Correct).

How many samples to learn a nite class?


The rst question we can ask concerns sample complexity: how many samples do we need to have
seen to learn a concept eectively? Its not hard to prove the following theorem: after we see

1
|C|
m=O
log

samples drawn from D, any hypothesis h C we can nd that agrees with all of these samples
(i.e., such that h(xi ) = c(xi ) for all i) will satisfy
Pr [h(x) = c(x)] 1
with probability at least 1 over the choice of x1 , . . . , xm .
We can prove this theorem by the contrapositive. Let h C be any bad hypothesis: that
is, such that Pr [h(x) = c(x)] < 1 . Then if we independently pick m points from the sample
distribution D, the hypothesis h will be correct on all of these points with probability at most
(1 )m . So by the union bound, the probability that there exists a bad hypothesis in C that
nevertheless agrees with all our sample data is at most |C| (1 )m (the number of hypotheses,
20-1

good or bad, times the maximum probability of each bad hypothesis agreeing with the sample
data). Now we just do algebra:
= |C| (1 )m

m = log1
|C|
log /|C|
=
log 1
1
|C|

log
.

Note that there always exists a hypothesis in C that agrees with c on all the sample points:
namely, c itself (i.e. the truth)! So as our learning algorithm, we can simply do the following:
1. Find any hypothesis in h C that agrees with all the sample data (i.e., such that h(xi ) = c(xi )
for all x1 , . . . , xm ).
2. Output h.
Such an h will always exist, and by the theorem above it will probably be a good hypothesis.
All we need is to see enough sample points.

How many samples to learn an innite class?


The formula

1
|C|
log

works so long as |C| is nite, but it breaks down when |C| is innite. How can we formalize the
intuition that the concept class of lines is learnable, but the concept class of arbitrary squiggles is
not? A line seems easy to guess (at least approximately), if I give you a small number of random
points and tell you whether each point is above or below the line. But if I tell you that these points
are on one side of a squiggle, and those points are on the other side, then no matter how many
points I give you, it seems impossible to predict which side the next point will be on.
So whats the dierence between the two cases? It cant be the number of lines versus the number
of squiggles, since theyre both innite (and be taken to have the same innite cardinality).
From the oor: Isnt the dierence just that you need two parameters to specify a line, but
innitely many parameters to specify a squiggle?
Thats getting closer! The trouble is that the notion of a parameter doesnt occur anywhere
in the theory; its something we have to insert ourselves. To put it another way, its possible to
come up with silly parameterizations where even a line takes innitely many parameters to specify,
as well as clever parameterizations where a squiggle can be specied with just one parameter.
Well, the answer isnt obvious! The idea that nally answered the question is called VCdimension (after two of its inventors, Vapnik and Chervonenkis). We say the set of points x1 , . . . , xm
is shattered by a concept class C if for all 2m possible settings of c(x1 ), . . . , c(xm ) to 0 or 1 (reject
or accept), there is some concept c C that agrees with those values. Then the VC-dimension
of C, denoted VCdim(C), is the size of the largest set of points shattered by C. If we can nd
arbitrarily large (nite) sets of points that can be shattered, then VCdim(C) = .
m

20-2

If we let C be the concept class of lines in the plane, then VCdim(C) = 3. Why? Well, we can
put three points in a triangle, and each of the eight possible classications of those points can be
realized by a single line. On the other hand, theres no set of four points such that all sixteen possible
classications of those points can be realized by a line. Either the points form a quadrilateral, in
which case we cant make opposite corners have the same classication; or they form a triangle and
an interior point, in which case we cant make the interior point have a dierent classication from
the other three points; or three of the points are collinear, in which case we certainly cant classify
those points with a line.
Blumer et al.1 proved that a concept class is PAC-learnable if and only if its VC-dimension is
nite, and that

VCdim(C)
1
m=O
log

samples suce. Once again, a learning algorithm that works is just to output any hypothesis h in
the concept class that agrees with all the data. Unfortunately we dont have time to prove that
here.
A useful intuition is provided by a corollary of Blumer et al.s result called the Occams Razor
Theorem: whenever your hypothesis has suciently fewer bits of information than the original
data, it will probably correctly predict most future data drawn from the same distribution.

Computational Complexity of Learning


Weve seen that given a nite concept classor even an innite class with nite VC-dimension
after seeing enough sample points, you can predict the future just by nding any hypothesis in the
class that ts the data. But how hard is it as a computational problem to nd a hypothesis that
ts the data? This has the general feel of something that might be NP-complete! In particular, it
feels similar to satisabilitynd some hypothesis that satises certain xed outputsthough its
not quite the same.
Here we need to make a subtle distinction. For proper learningwhere the goal is to output
a hypothesis in some xed format (like a DNF expression), its indeed possible to prove in some
cases that nding a hypothesis that ts the data is NP-complete. For improper learningwhere
the hypothesis can be any polynomial-time algorithm so long as it predicts the datato this day
we dont know whether nding a hypothesis is NP-complete.
On the other hand, the learning problem is certainly in N P , since given a hypothesis its easy to
check whether it ts the data or not. This means that if P = N P , then all learning problems are in
P and are computationally tractable. Think about what that means: we could ask our computer to
nd the shortest ecient description of the stock market, the patterns of neural rings in a human
brain, etc., and thereby solve many of the hardest problems of AI! This is yet another reason to
believe P = N P .

Blumer, Ehrenfeucht, Haussler, Warmuth, Learnability and the Vapnik-Chervonenkis dimension, JACM, 1989

20-3

MIT OpenCourseWare
http://ocw.mit.edu

6.045J / 18.400J Automata, Computability, and Complexity


Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

You might also like