Sparsely connected Hopfield networks for the recognition of correlated pattern sets

Abstract

A sparsely connected Hopfield network for the recognition of natural, highly correlated video images is proposed. A general design mechanism for the construction of a local neighbourhood structure using a statistical analysis of an arbitrary given pattern set is suggested. The duality between learning and dilution is employed and different learning and dilution schemes are discussed. The practical use and the efficiency of the model are shown in simulations of a large network (N=12288). The authors use a set of natural patterns with high inter pattern correlations and a high site correlation within each pattern, in which the correlations are given and not constructed by special rules as for highly correlated random pattern sets. The results obtained are analysed for different coding types of the binary pattern set.