Abstract

We implement a neural network which takes advantage of the parallelism of optics and the flexibility of electronics by using a mutually pumped, phase conjugate mirror (MP-PCM) and computer controlled feedback. This device has the potential of modeling networks with 10⁵ neurons and 10¹⁰ interconnections. The states of all neurons are represented as a 2-D array of pixels on a spatial light modulator. Each pixel is interconnected optically in 3-D via a photorefractive hologram in the MP-PCM which uses two beams input into a photorefractive crystal. Each beam forms a continuum of gratings through two-wave mixing within the crystal. The two beams mutally interact and each pixel in the input plane writes gratings with all the other pixels to form the phase conjugate beams. The gratings produced by one beam are read by the second beam, forming a multiport device. The neuron nonlinear activation functions are implemented electronically using a frame grabber and host computer. Homoassociations are performed by addressing the crystal with the same spatial information on both input beams. The mutual interaction in the crystal allows a partial version of one image to reconstruct the entire version of the second image. For heteroassociations, the crystal is addressed with different images on each beam.

© 1989 Optical Society of America