A supervised learning network based on adaptive resonance theory.

Abstract

A neural network architecture, fuzzy ART with logistic discrimination (ART-LD), is introduced as a method of realising the pattern recognition task in a supervised learning manner. The system is formed by the hierarchical organisation of two network modules: a fuzzy ART and a logistic discrimination. The learning consists of two separate stages. Firstly, the fuzzy ART module self-organises the input patterns into category clusters, whose operations are governed by fuzzy set theory and "competitive learning" dynamics that ensure fast and stable learning. Then the outputs, which can be interpreted as fuzzy memberships of an input pattern to the encoded categories, provide the spatial distance information that is generalised by the subsequent logistic discrimination to give the final prediction. Examples are presented, and the generalisation capabilities of ART-LD are demonstrated through two simulated and one real classification problem.