Associative storage and retrieval of highly correlated natural pattern sets in diluted Hopfield networks

A model for the recognition of natural highly correlated video images is proposed. The problems of storing all patterns as individual attractors and of building classes are discussed. It is shown that a sparsely connected associative memory of the Hopfield type is able to fulfill both tasks. A design strategy for the construction of a suitable locally connected architecture is suggested, using a statistical analysis of an arbitrary given pattern set. The duality between learning and dilution is employed, and different learning respectively dilution schemes are discussed. The practical use and the efficiency of the model are shown in simulations of a large network (N=12288). A set of natural patterns with high interpattern correlations and a high site correlation within each pattern is used, in which the correlations are given and not constructed by special rules as for highly correlated random pattern sets