Break Point of H

if no k inputs can be shattered by H, call k a break point for H.

\[m_H(k) < 2^k\]

Growth function

\(m_H(N)\) : max number of dichotomies

    Break Point \(m_H(N)\)
Positive rays \(m_H(2) = 3 < 2^2\) 2 \(N + 1\)
Positive intervals \(m_H(3) = 7 < 2^3\) 3 \(N^2/2 + N/2 + 1\)
Convex sets   NA \(2^N\)
2D Perceptrons \(m_H(4) = 14 < 2^4\) 4 ?

Bounding function

B(N,k) maximum possible \(m_H(N)\) when break point = k

  • B(2,2) = 3, maximum < 4
  • B(3,2) = 4
  • B(N,1) = 1

\(B(N,k) = 2^N\) for N < k

\(B(N,k) = 2^N - 1\) for N = k

\[B(N,k) = 2\alpha + \beta \\ \alpha + \beta \le B(N-1, k) \\ \alpha \le B(N-1, k-1) \\ \rightarrow B(N,k) \le B(N-1,k) + B(N-1, k-1)\]
B(N,k) k=1 2 3 4 6 6
N=1 1 2 2 2 2 2
2 1 3 4 4 4 4
3 1 4 7 8 8 8
4 1 <=5 <=11 15 16 16
5 1 <=6 <=16 <=26 31 32
6 1 <=7 <=22 <=42 <=57 63

Bounding Function: The Theorem

\[B(N,k) \le \sum_{i=0}^{k-1} \binom{N}{i}\]

highest term: \(N^{k-1}\)


want:

\[\mathbb{P} [ \exists h \in H s.t. | E_{in}(h) - E_{out}(h) | > \epsilon ] \le 2 \ m_H(N) \ exp( -2 \epsilon^2 N )\]

when N large enough, ->

\[\mathbb{P} [ \exists h \in H s.t. | E_{in}(h) - E_{out}(h) | > \epsilon ] \le 2 \times 2 \ m_H(2 N) \ exp( -2 \frac{1}{16} \epsilon^2 N )\]

Vapnik-Chervonenkis (VC) bound:

\[\mathbb{P} [ \exists h \in H s.t. | E_{in}(h) - E_{out}(h) | > \epsilon ] \le 4 \ m_H(2 N) \ exp( - \frac{1}{8} \epsilon^2 N )\]
  • replace Eout by E’in
  • decompose H by kind
  • use Hoeffding without replacement