Optical perceptron-based quadratic neural network

Abstract

Optical quadratic neural networks are currently being investigated because of their advantages over linear neural networks.1 Based on a quadratic neuron already constructed,2 an optical quadratic neural network utilizing four-wave mixing in photorefractive barium titanate (BaTiO3) has been developed. This network implements a feedback loop using a charge-coupled device camera, two monochrome liquid crystal televisions, a computer, and various optical elements. For training, the network employs the supervised quadratic Perceptron algorithm to associate binary-valued input vectors with specified target vectors. The training session is composed of epochs, each of which comprises an entire set of iterations for all input vectors. The network converges when the interconnection matrix remains unchanged for every successive epoch. Using a spatial multiplexing scheme for two bipolar neurons, the network can classify up to eight different input patterns. To the best of our knowledge, this proof-of-principle experiment represents one of the first working trainable optical quadratic networks utilizing a photorefractive medium.