Abstract

With improving optical device technology it is now possible to consider constructing very large optoelectronic neural networks containing of the order of 1000 fully interconnected neurons. Whether such systems will work, though, depends on the quality of the optical devices and the network architecture. A recent study has indicated that a 200 neuron single layer polarization logic network can operate under the constraints of presently available spatial light modulators.¹ In this paper we will extend this study to examine the effect of device limitations on the performance of single- and multilayer networks containing up to 1000 fully interconnected neurons. We consider the effect of nonlinearities, contrast ratio, and quantization noise in the weight matrix and the effect of noise at the inputs and outputs on the convergence rate and number of patterns that can be classified.

© 1992 Optical Society of America