VC Dimension definition and impossibility result Lecturer Yishay

VC Dimension – definition and impossibility result Lecturer: Yishay Mansour Eran Nir and Ido Trivizki

VC Dimension – Lecture Overview n n n PAC Model – Review VC dimension – motivation Definitions Some examples of geometric concepts Sample size lower bounds More examples

The PAC Model - Review n n A fixed, unknown distribution D from which the examples are chosen independently. The target concept is a computable function n n Our goal – finding h such that: - accuracy parameter; - confidence parameter. An algorithm A learns a family of concepts C if for any and any distribution D , A outputs a function such that.

VC Dimension - Motivation n Question: How many examples does a learning algorithm need? For PAC and a finite concept class C we proved: We would like to be able to handle infinite concept classes – VC Dimensions will provide us a substitute to for infinite concept classes.

VC Dimension - Definitions n n Given a concept class C defined over the instance space X, let The projection of C on S is all the possible functions that C induces on S : n n A concept class C shatters S if In other words: a class shatters a set if every possible function on the set is in the class.

VC Dimension – Definitions Cont. VCdim (Vapnik-Chervonenkis dimension) of C: n The maximum size of a set shattered by C: n If a maximum value doesn’t exist then n For a finite class C:

VC Dimension – Examples n In order to show that the VCdim of a class is d we have to show: : find some shattered set of size d. : show that no set of size d+1 is n n shattered

VC Dimension – Examples: Half Lines (C 1) n The concepts are for where:

VC Dimension – Examples: Half Lines (C 1) Cont. Claim: n n : , , thus . : for any set of size 2 there is an assignment which is not in the concept class: for the assignment which lets x be 1 and y be 0 is impossible.

VC Dimension – Examples: Linear halfspaces (C 2) n The concepts are where for let. are lines in the plane where positive points above or on the line, and negative points are below.

VC Dimension – Examples: Linear halfspaces (C 2) Cont. Claim: n n : Any three points that are not collinear can be shattered. : No set of four points can be shattered: Generally: Half spaces in have VCdim of .

VC Dimension – Examples: Axisaligned rectangles in the plane (C 3) n Positive examples are points inside the rectangle, and negative examples are points outside the rectangle.

VC Dimension – Examples: Axisaligned rectangles in the plane (C 3) Claim: n : a for points set in the following shape can be shattered:

VC Dimension – Examples: Axisaligned rectangles in the plane (C 3) Claim: n : Given a set of five points in the plane, there must be some point that is neither the extreme left, right, top or bottom point of the five. If we label this non-extermal point negative and the remaining four extermal points positive, no rectangle can satisfy the assignment.

VC Dimension – Examples: A finite union of intervals (C 4) n For any set of points we could cover the positive points by choosing the intervals small enough so

VC Dimension – Examples: Convex Polygons on the plane (C 5) n n n Points inside the convex polygon are positive and outside are negative. There is no bound on the number of edges. Claim:

VC Dimension – Examples: Convex Polygons on the plane (C 5) Proof: n For every labeling of d points on the circle perimeter, there exists that is consistent with the labeling. n This is a polygon which includes all the positive examples and none of the negative. Thus the group of points is shuttered. n This holds for every d, and so

Sample Size Lower Bounds n n Goal: we want to show that for a concept class with a finite VCdim d there is a function m of such that if we sample less than points, any PAC learning algorithm would fail. Theorem: If a concept class C has VCdim d+1 then:

Sample Size Lower Bounds Proof n For contradiction: let shatters T (possible because Let D(x) be n Choose n randomly so that it’s such that C ).

Sample Size Lower Bounds – Proof Cont. n n n is in C because C shatters T. Claim: if we sample less than points out of then the error is at least. Proof: Let RARE be n n n Sample size: the expected number of points we sample from RARE is at most Error: This implies that with probability of at least 0. 5 we sample at most points of RARE and thus have error of at least.

VC Dimension – Examples: Parity (C 6) n n Let where Claim: n assignment the set and so n . The concept class is. : Let for the vectors. We get: is shattered. : There are . For any bits we choose parity functions, thus

VC Dimension – Examples: OR of n literals (C 7) n Let . The concept class is n Claim: n n : use n unit vectors (see prev. proof). : Use ELIM algorithm to show. Show the (n+1) vector cannot be assigned 1, thus no set of (n+1) vectors can be shuttered.

Radon Theorem n n Definitions: n Convex Set: A is convex if for every the line connecting is in A. n Convex Hull: The Convex Hull of S is the smallest convex set which contains all the points of S. We denote it as conv(S). Theorem (Radon): n Let E be a set of d+2 points in. There is a subset S of E such that.

VC Dimension – Examples: Hyper-Planes (C 8) n n The concept class assigns 1 to a point if it’s above or on a corresponding hyper-plane, 0 otherwise. Claim: n n : use n unit vectors and the zero vector to form a n+1 set that can be shuttered. : use Radon theorem (next page)

VC Dimension – Examples: Hyper-Planes (C 8) Cont. n n Assume a set of size d+2 points can be shattered. Use Radon Theorem to find S such that Assume there is a separating hyper-plane that classifies points in S as ‘ 1’, points not in S as 0. No way to classify points in.