Abstract

An overview of our recent progress on the photonic implementation of neural networks is given. A photonic architecture utilizes spatial and double angular multiplexing in conjunction with an array of (individually coherent but mutually incoherent) sources to provide photonic neural interconnections in volume holographic materials. It also incorporates optoelectronic smart spatial light modulators (SLMs) for neuron unit response and training term generation. The SLM designs incorporate dual channel neuron unit inputs and outputs to provide complete bipolar functionality and yield a high bandwidth response to permit pulse train encoding or temporal multiplexing of signals. A unique feature of this architecture is the ability to copy the multiplexed holographic interconnection in a single exposure step. The presentation includes a discussion of simulations and laboratory experiments of the interconnection system implemented in photorefractive materials; fabrication and test of silicon detector, silicon analog electronic processing, and InGaAs/GaAs multiple quantum well modulator arrays; demonstration of single-step copying of a multiplexed set of mutually incoherent volume holograms; and requirements for training term SLM arrays and projected learning performance. Current and future issues in such photonic implementations are also discussed, as are specific advantages resulting from the high degree of parallelism available.

© 1992 Optical Society of America