Abstract

A reconfigurable optical crossbar switch has been integrated into a microcomputer environment for modeling neural networks that learn. A differential technique is used for obtaining bipolar inputs, outputs, and connection weights in a mask-type crossbar using an inexpensive low-contrast spatial light modulator. Neural networks based on matched-filter techniques and competitive interconnects have been implemented for adaptive associative recall as well as unsupervised learning in an autonomous controller.

© 1988 Optical Society of America