Network Analysis of an Optically Implemented Connectionist Architecture

Abstract

We investigate the system error tolerance of optical connectionist architectures, which utilize liquid crystal spatial light modulators to represent neurons and weights in a neural network configuration. Experimental and computer simulation results show training on one layer networks compensates for errors in the optical and electronic hardware. The mathematical description of the least mean square algorithm used to train these networks shows that this is to be expected. This tolerance, however, is not shown in multiple layer networks implemented with similar architectures. To overcome problems associated with linear dependence between training patterns and maintain the error tolerance of the one layer machines (without going to multiple layer networks), an architecture with higher level interconnections is proposed. This fully interconnected second order architecture requires only a single optical pass before updating the weights. It also uses a very similar training algorithm to that of the single layer network.