1. Vapnik-Chervonenkis Dimension Part II: Lower and Upper bounds

2. PAC Learning model There exists a distribution D over domain X Examples: Goal: –With high probability (1-  ) –find h in H such that –error(h,c ) < 

3. Definitions: Projection Given a concept c over X –associate it with a set (all positive examples) Projection (sets) –For a concept class C and subset S –  C (S) = { c  S | c  C} Projection (vectors) –For a concept class C and S = {x 1, …, x m } –  C (S) = { | c  C}

4. Definition: VC-dim Clearly |  C (S) |  2 m C shatters S if |  C (S) | =2 m VC dimension of a class C: –The size d of the largest set S that shatters C. –Can be infinite. For a finite class C –VC-dim(C)  log |C|

5. Lower bounds: Setting Static learning algorithm: –asks for a sample S of size m(  ) –Based on S selects a hypothesis

6. Lower bounds: Setting Theorem: –If VC-dim(C) =  then C is not learnable. Proof: –Let m = m(0.1,0.1) –Find 2m points which are shattered (set T) –Let D be the uniform distribution on T –Set c t (x i )=1 with probability ½. Expected error ¼. Finish proof!

7. Lower Bound: Feasible Theorem –VC-dim(C)=d+1, then m(  )=  (d/  ) Proof: –Let T be a set of d+1 points which is shattered. –Let the distribution D be: z 0 with prob. 1-8  z i with prob. 8  /d

8. Continue –Set c t (z 0 )=1 and c t (z i )=1 with probability ½ Expected error 2  Bound confidence –for accuracy 

9. Lower Bound: Non-Feasible Theorem –For two hypotheses m(  )=  ((log 1  )    ) Proof: –Let H={h 0, h 1 }, where h b (x)=b –Two distributions: –D 0 : Pr[ ]= ½ -  and Pr[ ]= ½ +  –D 1 : Pr[ ]= ½ +  and Pr[ ]= ½ - 

10. Epsilon net Epsilon bad concepts –B  ( c ) = { h | error(h,c) >  } A set of points S is an  -net w.r.t. D if –for every h  in B  ( c ) –there exists a point x in S –such that h(x)  c(x)

11. Sample size Event A: –The sample S 1 is not an epsilon net, |S 1 |=m. Assume A holds –Let h be a epsilon-bad consistent hypothesis. Sample an additional sample S 2 –with probability at least 1/2 –the errors of h on S 2 is  m/2 –for m=|S 2 |= O(1/ 

12. continues Event B –There exists h in B  ( c ) –and h consistent with S 1 –h has  m/2 errors on S 2 Pr[ B | A ]  1/2 –2 Pr[B]  P[A] Let F be the projection of C to S 1  S 2 –F=  C (S 1  S 2 )

13. Error set ER(h)={ x : x  S 1  S 2 and c(x)=h(x)} |ER(h)|   m/2 Event A: –ER(h)  S 1 =  Event B: –ER(h)  S 1 =  –ER(h)  S 2 = ER(h)

14. Combinatorial problem 2m black and white balls –exactly l black balls Consider a random partition to S 1 and S 2 The probability that all the black balls in S 2

15. Completing the proof Probability of B –Pr[B]  |F| 2 -l  |F| 2 -  m/2 Probability of A –Pr[A]  Pr[B]  |F| 2 -  m/2 Confidence  Pr[A] Sample –m=O( (1/  ) log 1/  (1/  ) log |F| ) Need to bound |F| !!!

16. Bounding |F| Define: –J(m,d)=J(m-1,d) + J(m-1,d-1) –J(m,0)=1 and J(0,d)=1 Solving the recursion Claim: –Let VC-dim(C)=d and |S|=m, –then |  C (S)|  J(m,d)