WINNER-TAKE-ALL NETWORKS FOR PHYSIOLOGICAL MODELS OF COMPETITIVE LEARNING

The winner-take-all (WTA) property is essential to competitive-learning systems. This article discusses WTA-type neural networks composed of nonlinear dynamic model neurons, characterized by a nonlinear loss term. It is shown by mathematical analyses that these networks have the WTA property even when their neurons have nonidentical characteristics and the interconnections have nonidentical strengths. The class of WTA networks is further generalized to allow explicitly modeled interneurons between the principal cells. A model of a cyclically operating WTA system, capable of handling changing inputs by automatically inactivating the winner, is then set up and demonstrated by computer simulations.