All-optical neural network with inhibitory neurons

Abstract

A Hopfield-type neural network with only inhibitory interconnections is shown to perform similarly to networks with both inhibitory and excitatory interconnects. With the use of only inhibitory neurons the need to subtract light intensities is eliminated, so an all-optical neural network can be realized. This realization is achieved with a liquid-crystal light valve for implementing a two-dimensional array of 16 inhibitory neurons and an array of subholograms for the interconnections. Two stable states that can be retrieved from noisy optical inputs are demonstrated.