12 - MULTILAYER COOPERATIVE/COMPETITIVE NETWORKS

Abstract

This chapter focuses on multi-layer cooperative/competitive networks. The networks discussed in the chapter include Rumelhart and Zipser's competitive learning network; the Cohen and Grossberg masking fields network; Kunihiko Fukushima's Neocognition; Grossberg's Boundary Contour System; and the Hierarchical Scene Structure system of Minsky and Maren. The competitive learning approach is a variation of feedforward architecture in which lateral inhibitory connections are allowed. The concept underlying the competitive learning network is that neurons in the higher levels of a multilevel network can learn to recognize features in input patterns by using an architecture that has lateral inhibitions combined with a learning method that reinforces the connections to winning neurons. This leads to a network that can recognize different types of input patterns based on the features of those patterns. The competitive learning network undergoes unsupervised learning and creates feature classes, which lead to its own categories of output.