Abstract

After the first optical implementation of the 1-dimensional Hopfield model had been reported [1], extensive researches have been conducted for 2-dimensional input/out patterns.[2,3] Page-oriented holograms [4] may achieve fairly large fixed interconnections, while volume holograms [5] or lenslet arrays with spatial light modulators (SLMs) [6] achieve adaptive interconnections. However the volume hologram still requires further researches, especially on fixing and copying, and SLM resolution is major limiting factor for large scale implementation for the latter. Recently we had developed an adaptive neural network architecture, TAG (Training by Adaptive Gain), which utilizes fixed global interconnections and adaptive local gains.[7,8] Performance with both random and pre-trained interconnections were investigated. In the previous papers we had used page- oriented holograms for the fixed interconnections. In this paper we show possibility of using ground glass for random interconnections with much higher diffraction efficiency and interconnection density, and report smallscale optical implementation of a classifying neural network.

© 1993 Optical Society of America