WINNER-TAKE-ALL NETWORKS AND ASSOCIATIVE MEMORY: ANALYSIS AND OPTICAL REALIZATION.

Neuromorphic systems that can be used as ACAMs (associative content-addressable memories) are considered. Several aspects of a class of discrete-valued high-order correlation ACAMs that are based on the outer-product rule for information storage are discussed. It is shown that as the order n of the correlation grows, the memory more closely approximates the behavior of a simple template matcher in which the input state is compared to a template of each of M stored memories to derive M correlation values. The closest (in Hamming distance for binary-valued states) match is chosen as the output state by some mechanism that finds the maximum among the set of M correlation values. A parallel version of the max finder can be implemented with a winner-take-all (WTA) network. It is shown that as n increases, the nth-order ACAM more closely approximates a three-layer feedforward network consisting of an input layer, a WTA network, and an output layer. A specific architecture for a WTA network composed of linear threshold units and a possible optical implementation of this network is offered. The capacity of the WTA-based memory and some tradeoffs between the outer-product ACAM are discussed.