Optical neural networks with random interconnections by cheap ground glass

Abstract

Ground glass, which provides high efficieny and very high density random interconnections, is proposed for large scale implementation of optical neural networks for 2-dimensional input and output patterns. In recently developed TAG (Training Adaptive Gain) neural network model the random fixed interconnections provide global connectivity, while local gains are trained for adaptive learning and implemented by 2-dimensional SLM (spatial light modulator). Compared to popular neural network models with global adaptive interconnections, this model requires much less SLM elements and suitable for very large scale implementation. Performance of the TAG model with random interconnections are first investigated theoretically, and possibility of using cheap ground glass for very large scale implementation of these random interconnections are also studied. Two attempts, output node selection and increased local adaptive connectivity, to improve performance of this TAG model with ground glass are also introduced.