Design and simulations of cellular neural-like associative memory

Abstract

A cellular-neuron associative memory (CNAM), which is an associative neural memory of the Hopfield type with a restricted number of connections, is investigated. Algorithms for designing CNAMs take advantage of the fine-grained parallelism induced both by independent cell operations and by connection locality. A very important property is the fact that learning and retrieval processes may be performed in the same cellular array. Some necessary and sufficient conditions for strong stability and k-attractability are obtained, which are expressed in terms of cell neighborhood relations of stored patterns. Simulation of learning and retrieval processes in a CNAM storing symbols drawn in thin lines showed that, for this class of patterns, it is possible to provide strong stability approximately for 2|Q| prototypes, where Q is the cardinality of the neuron neigborhood, with the capability of restoring 60-70% of 1-distortions.