Abstract

There has been a recent resurgence of interest in associative-processing/neural-network architectures as a model for parallel processing, particularly in such difficult problem domains as image understanding and artificial intelligence. Optics offers unique capabilities for implementing the massive parallel interconnection between successive 1-D and/or 2-D information arrays required by these architectures. Furthermore, these systems are often low-precision (e.g., binary) and hence are compatible with the limited dynamic-range capabilities of optics. We are developing optical architectures consisting of multiple adaptive, associative modules¹ interconnected and cascaded in particular configurations, according to a variety of organizational principles. The experimental performance of actual optical implementations of the individual adaptive associative modules will be discussed. These optical modules adapt as they are exposed to associated information pattern vectors (u,v), so that subsequent presentation of one pattern, u, results in recall of the other, v. The set of associations is dynamically learned in real time and stored as an electron charge distribution on an electrooptic crystal in a spatial light modulator. These architectures include feedback, which makes them self-correcting to a variety of optical system aberrations, and they have the capacity to store N associated pairs of N-element vectors, where N of a few hundred appear feasible.

© 1985 Optical Society of America