Achieving optimal Bayesian classification performance using a novel approach: the 'race to the attractor' neural network model

Abstract

It is possible to classify patterns by using a set of nonlinear dynamical systems (NDS) where each NDS specializes in classifying inputs as IN or OUT of the specific class they represent. Inputs are iterated through each NDS and converge along a trajectory towards a globally stable attractor which is the prototype for the class represented by that NDS. In the 'race to the attractor' neural network model (RTA NN), neural nets learn a convergence rate function for each input so that the time required to converge to the attractor increases as the probability of class membership of the input decreases. By iterating all NN simultaneously with the same input, a 'race to the attractor' ensues, where the winner 'r' identifies the unknown input as a member of class 'r'. The optimal classification performance of a Bayesian classifier can be achieved by the RTA NN when certain constraints (discussed in the article) are met.