Quasi-optimum combination of multilayer perceptrons for adaptive multiclass pattern recognition

Abstract

Standard multiclass pattern recognition requires frequent re-learning stages when the set of categories of interest evolves in time. In order to minimize the computation costs of class incorporation and removal, we divide the global multiclass recognizer into a collection of class pairwise neural dichotomizers. When a new class appears, an adequate set of dichotomizers is created and trained to discriminate the new class from the rest. If a class disappears, its associated dichotomizers are eliminated In both cases previously learned knowledge is not disturbed. The properties of neural recognizers and pairwise modularization allow an analytic quasi-optimum method for combining network outputs to obtain the global multiclass response. An incremental and distributed pattern recognition architecture is presented and its performance experimentally evaluated, obtaining better error rates and learning times than conventional multiclass recognizers using similar resources. The design is highly parallel and asyncronous, adequate for dynamic real time applications.