Optical quadratic perceptron neural network

Abstract

Optical quadratic neural networks are currently being investigated because of their advantages with respect to linear neural networks.1 A quadratic neuron has previously been implemented by using a photorefractive barium titanate crystal.2 This approach has been improved and enhanced to realize a neural network that implements the perceptron learning algorithm. The input matrix, which is an encoded version of the input vector, is placed on a mask, and the interconnection matrix is computer-generated on a monochrome liquid-crystal television. By performing the four-wave mixing operation, the barium titanate crystal effectively multiplies the light fields representing the input matrix by those representing the interconnection matrix to produce an analog output. This output is then digitized by a computer, thresholded, and compared to a specified target vector. An error signal representing the difference between the target and thresholded output is generated, and the interconnection matrix is iteratively modified until convergence occurs. The characteristics of this quadratic neural network will be presented and discussed.