Dynamic competitive learning in the differentiator

A modified version of the differentiator, which is an unsupervised pattern classifier, is described. The learning rule is based on three neurobiologically inspired processes: sensitization, habituation, and recovery. By incorporating several features, the differentiator is able to overcome the problems of a simple competitive learning method in finding clusters of patterns. This is achieved through an enhanced dynamic competition involving all the weight vectors. It is shown by simulation that this network performs better than the simple winner-take-all method of competitive learning.<
>