Generalization performance of complex adaptive tasks
E.M. EISENSTEIN, Ido Kanter · Physical Review Letters · 1993
Optimal strategies for predicting correctly the output of a few new random inputs, when various feedforward networks are trained by noise-free random training examples, are examined analytically and numerically. The existence of a universal strategy for various generalization tasks is discussed, and indicates that the Bayes algorithm is not always the optimal strategy.