Optical inner-product implementation of multi-layer correlation-matrix neural networks for 2-dimensional patterns

Abstract

An optical modular architecture based on an inner-product implementation scheme is presented for a multi-layer correlation-matrix associative memory. The modular architecture consists of compact modules for a single-layer feed-forward network, which may be cascaded for multi-layer and/or bidirectional networks. Provided the interconnections are defined as a correlation matrix of binary patterns such as the Hopfield model, bidirectional associative memory (BAM) model, and their multi-layer extensions, the inner-product implementation scheme utilizes only binary storage and binary spatial light modulators (SLMs). This resolves one of the biggest bottlenecks in the optical implementation of neural networks, i.e., the requirement of grey-level high-resolution SLMs, and provides a great advantage in the inner-product scheme for large-scale optical implementations. The performance of the proposed architecture and multi-layer correlation-matrix memory is demonstrated by an electro-optic inner-product implementation for the Exclusive-OR problem, which is regarded as a difficult problem for neural networks.