Optical Implementation of Neural Networks with Fixed Global Interconnection and Local Adaptive Gain-Control

Abstract

A new adaptive learning algorithm has been developed for optical implementation of large-sized neural networks. In this model N 4 interconnections for N × N 2-dimensional neurons are composed of two different types, i.e. global fixed N 4 interconnections and N 2 adaptive gain control. The former may be achieved by a multifacet(N 2 facets) hologram, where each holographic facet stores N 2 interconnections. The latter may be done by spatial light modulators (SLMs) of N 2 elements. This model allows us to implement neural networks of N × N neurons with SLMs of only N × N elements. The adaptive learning algorithm is based on gradient descent and error back-propagation, and easily extendable to multilayer structures. Performance of this model has been investigated, and its electro-optical implementation will be proposed.