Bipolar pattern association using a recurrent winner take all network
John E. McInroy, Bogdan M. Wilamowski · Proceedings of International Conference on Neural Networks (ICNN'97) · 2002
A neural network for heteroassociative (or autoassociative) pattern recognition of input bipolar binary vectors is proposed. By combining the advantages of feedforward and recurrent techniques for heteroassociation, a simple network with guaranteed error correction is found. The heart of the network is based on a new, recurrent method of performing the winner-take-all function. The analysis of this network leads to design rules which guarantee its performance. The network is tested on a character recognition problem utilizing the entire IBM CGA character set.