Theory of Generalization
test error can approximate training error if there is enough data and growth function does not grow too fast
The Four Breaking Points
growth function $m_H(N)$:max number of dichotomies
- positive rays: $m_H(N)=N+1$
- positive intervals: $m_H(N)=frac{1}{2}N^2+frac{1}{2}N+1$
- convex sets: $m_H(N)=2^N$
- 2D perceptrons :$m_H(N)<2^N$ in some cases
Restriction of Breaking Point
近期评论