統計的機械学習は単純な最適化問題ではない

It might seem that minimization of  K_n(w) is equivalent to minimization of  K(w). If these two minimization problems were equivalent, then maximization of  L_n(w) would be the best method in statistical estimation. However, minimization and expectation cannot be commutative.

 E[\min_w K_n(w) ] \neq \min_w E[K_n(w)] = \min_w K(w)

Hence maximization of  L_n(w) does not mean minimization of  K(w). This is the basic reason why statistical learning does not result in a simple optimization problem.

p.7

Algebraic Geometry and Statistical Learning Theory (Cambridge Monographs on Applied and Computational Mathematics)

Algebraic Geometry and Statistical Learning Theory (Cambridge Monographs on Applied and Computational Mathematics)

広告を非表示にする