Noise and cross talk study in an optical neural network

Abstract

The performance of software implemented artificial neural networks in many applications is well documented; however, fewer studies have been reported on networks implemented using less-than-ideal optical hardware. Modeling and characterization of hardware imperfections is essential before a transition can be made from using CPU intensive software simulated neural networks to practical, electro-optic based neural networks. In this report we study non-linear and limited accuracy components as error sources and experimentally document their effects. The electro-optic neural-network used for this research consists of two electrooptic artificial neural-network layers and one artificial neural-network layer implemented in software only. In the electro-optic layers, light emitting diodes (LEDs) are used to provide the input, liquid crystal spatial light modulators (SLM) serve as the interconnection weight matrixes, and photodiode detectors are the nonlinear thresholding elements. The error sources analyzed include: nonuniform LED illumination, optical misalignment and cross talk within the SLM, thermal drift in the SLM, and repeatability, linearity, and analog accuracy of the SLM.