A method for training recurrent neural networks for classification by building basins of attraction

Abstract

The main contribution of this report is the suggestion, development, and trial of a new way of using a Hopfield network for classification. Rather than storing the members of the training sets (sets of exemplars) as stable points, a connection matrix is determined, using perceptron-type learning, such that the exemplars are put into basins of attraction. The elements from different training sets are placed in different basins of attraction, with usually more than one basin per training class. The attractors for the basins of attraction are determined at the same time that the exemplars are placed in these basins. An arbitrary element is then classified by the basin of attraction it falls into. The method is successfully applied to artificially generated classes and to the classification of cervical cells for cancer detection.