New, Faster Algorithms for Supervised Competitive Learning: Counterpropagation and Adaptive-Resonance Functionality

Abstract

Hecht-Nielsen's counterpropagation networks often learn to associate input and output patterns more quickly than backpropagation networks. But simple competitive learning cannot separate closely spaced input patterns without adaptive-resonance-like (ART) functionality which prevents neighboring patterns from ‘stealing’ each other's templates. We demonstrate ‘pseudo-ART’ functionality with a new, simple, and very fast algorithm which requires no pattern normalization at all. Competition can be based on either Euclidean or L1-norm matching. In the latter case, the new algorithm emulates fuzzy ART. We apply the pseudo-ART scheme to several new types of counterpropagation networks, including one based on competition among combined input/output patterns, and discuss application with and without noise.