Complementary aspects of topological maps and time delay neural networks for character recognition

A study comparing a back propagation-like network integrating feature selection notions introduced in neocognitron networks with a supervised learning algorithm, based on Kohonen´s self-organizing feature maps, is presented. The two methods are applied to handwritten zipcode recognition. The results achieved by the networks are reported using common training and test sets. The complementary aspects of these networks and possible cooperative schemes are discussed. It is shown that the performance obtained, even with a very simple cooperation scheme, significantly exceeds those of the networks working separately