ELEMENTS OF THE STATISTICAL LEARNING CONCEPT
FOR A NEURAL NETWORK AND ACCURATE
PREDICTION OF ITS OPERATION
G. F. Malychina, A. V. Merkusheva
Saint-Petersburg
The learning of neural networks (NN)
for many problems (pattern recognition, nonlinear multi-parameter re-gression,
probability distribution identification) is considered in generalized form
on the basis of a concept that includes probabilistic interpretation for
the NN input—output transfer function and basic notions having a mathematically
formalized foundation: diversity (a set) of mapping being realized by NN
(and a set of loss functions isomorphic to it); characteristics of that diversity
on the basis of entropy and Vapnik—Chervonenkis dimension; risk functional
(RF) and a condition allowing RF approximation by means of an empirical risk
func-tional (ERF); the limits of the actual RF departure from ERF. The elements
of the leaning statistical theory de-scribed here provide prediction and
correction ("control") of the NN operation index after leaning, i.e. at the
stage of NN testing with the data on not participating in learning.