Abstract

A basic feature of the optical neural networks (or neurooptic processors¹) being developed by Northrop is that each optical synapse is identified uniquely with a single-volume holographic phase grating of grating vector K. These optical synapses occupy a single-point vicinity in Fourier space (K_x,K_y,K_z), while their biological counterparts occupy a single-point vicinity in configuration space (x,y,z). This type of neurooptic architecture evokes a number of problems fundamental to real time high-throughput optical processing. In this paper we address ourselves specifically to the problem of executing adaptive neural network algorithms using multiple-exposure Bragg holography.

© 1986 Optical Society of America