Decision theoretic generalizations of the PAC model for neural net and other learning applications (Q1198550)

The paper introduces an extension of the probably approximately correct model of learing from examples. The class of functions on an instance space that usually are defined only on the set \(\{0,1\}\) has been changed so that values are members of arbitrary sets. The new extended model is based on statistical decision theory. The results of the paper are based on a proposed generalized notion of dimension of Vapnik-Chervonenkis that is applicable to real-valued functions. Due to the extended model, it is possible to formulate distribution-independent upper bounds on the size of the instance space for learning in feed-forward neural networks.