Abstract

Although backward error propagation learning in photorefractive crystals has been previously investigated by simulation and experiment, theoretical results governing convergence have been lacking. In this paper we prove analytically that such learning in multilayer neural networks implemented using photorefractive crystals can have similar convergence properties to those of an ideal backward error propagation network. Further, we derive relationships between two learning parameters that will ensure these convergence properties are satisfied under the assumption of small weight-update sizes, and we relate these parameters to spatial light modulator gain and holographic grating update exposure energy.

© 1995 Optical Society of America