Optoelectronic array that computes error and weight modification for a bipolar optical neural network

Abstract

The design, fabrication, and results of an optoelectronic device that computes the weight changes required by the delta-rule learning algorithm and encodes the result on a pair of optical beams are presented. This very-large-scale-integrated ferroelectric liquid-crystal array was designed specifically to permit bipolar optoelectronic neural networks to learn without the limitations of an external controlling computer. The device contains 64 smart pixels, which represent the processing elements in an artificial neural network. Each processing element consists of two photodetectors, a current-to-voltage converter, two comparators, two switches, and two output liquid-crystal modulating pads. The device has a measured contrast ratio of 10:1, a 10%-to-90% rise time of 350 micros, and a 90%-to-10% fall time of 150 micros.