Abstract

Photorefractive implementations of optical neural networks (ONN) can be divided into two architectures: stimulated and nonstimulated. In nonstimulated photorefractive ONN's, each connection weight is stored in the relatively well-defined grating that is formed by the interference of two plane or spherical waves. The connection weights are updated by superimposing new gratings on the old ones and by using an exposure schedule to compensate for grating erasure. Bragg ambiguity effects, in which a particular grating can be read out by beams different from the ones that wrote the grating, are a source of crosstalk in such architectures. The stimulated photorefractive optical neural network (SPONN) stores each connection weight in a continuum of spatially and angularly multiplexed gratings. The precise distribution of the gratings is determined by beam fanning, which is seeded by scattering of the incident beams by crystal inhomogeneities, and it evolves according to the dynamics of photorefractive two-and four-wave mixing. We have shown experimentally that sets of such multiplexed gratings generated in both self-and mutually-pumped photorefractive phase-conjugate mirrors can serve as global interconnection networks without the Bragg-ambiguity crosstalk characteristic of nonstimulated ONN's. A simple geometric argument shows that scattering off two or more gratings in series is sufficient to break the Bragg ambiguity. This permits us to arrange optical neurons in arbitrary patterns on spatial light modulators and to fully utilize the spatial light modulator's space-bandwidth product. We will present recent experimental results on crosstalk in SPONN and will discuss our progress towards implementing a complete optical neural network.

© 1990 Optical Society of America