A new paradigm for classification tasks the 'race to the attractor' neural network model

Abstract

An approach to the classification of patterns is proposed, where time is the key element used to categorize specimens. The model consists of a set of independent nonlinear dynamical systems (NDS) where each NDS has a unique, globally stable attractor which is a prototype representing all patterns belonging to that class. All inputs to each NDS eventually get transformed into the class attractor, but in an amount of time which is inversely proportional to the probability of class membership for that input. By iterating an unknown input through all the NDS simultaneously, a 'race to the attractor' ensues, where the winner identifies the input as a member of the class represented by that NDS attractor. The proposed model has several advantages over traditional classification paradigms, including the ability to repair damage caused by the death of neurons and restore classification performance almost completely.