Accuracy enhancement in optical linear algebra processors

Abstract

The most important and fundamental drawback of optical information processing systems is their low accuracy. Although various noise models have been developed and attempts to improve the accuracy have been made.1,2 a thorough and sufficiently general understanding of the statistical behavior of optical processors has not yet been established. In this paper a generic optical matrix- vector multiplier is analyzed from a statistical-optics viewpoint. The output probability density functions are derived for some special cases of interest. The signal dependence of the noise in optical processors are thus proved. Two methods, namely, nonuniform quantization, and use of a normalizing transform, are then suggested as post-processing approaches that can potentially render the noise signal-independent and thus greatly simplify the decision-making process, resulting in an overall enhancement in processor accuracy. Computer simulations of the systems’ behavior and performance analyses of the two suggested methods are included.