ON STABLE POINTS AND CYCLES IN BINARY NEURAL NETWORKS.

The design of content-addressable memories by means of binary neural networks is tied to the synthesis of stable points with a guaranteed region of convergence for pattern storage and retrieval. The authors define a local minimum in the binary field, and suggest the usefulness of the concept in memory design. It is shown that N (nonarbitrary) stable points, left bracket N/2 right bracket local minima, and left bracket N/2 right bracket four-cycles can always be assigned. Simple procedures for selection of these points are outlined. The tradeoff between memory capacity and pattern retrievability through a definition of entropy in the memory is demonstrated.