Effects of Imperfection in Spatial Optical Devices on Backpropagation Learning Capability of Optoelectronic Neural Network

Abstract

Recently there has been a great deal of work on neural network. Offering massive parallelism, high speed, and crosstalk-free interconnection, optical implementation has been sought to fully exploit the parallel characteristics of neural networks(1,2). In those optoelectronic neural networks, "two-dimensionally(2D) extended" and "discretely arrayed" devices, i.e. spatial light devices such as SLMs(spatial light modulators), arrayed light sources and detectors are utilized to enjoy the high parallelism of optics. However, at present it is not easy to obtain such perfect spatial light devices with characteristics satisfying theoretical performance; some have lack of spatial uniformity while others show limited signal-to-noise ratio and so on. Therefore, it is important to investigate the effects of imperfection of spatial optical devices on the system capability of optoelectronic neural network. The influence of interconnection weight discretization and noise in an optoelectronic neural network was reported(3), but only by computer simulation.