Abstract

Optical implementations of neural networks can combine advantages of neural network adaptive parallel processing and optical free-space connectivity. Binary valued Backpropagation¹, a supervised learning algorithm related to standard Backpropagation², significantly reduces interconnection storage and computation requirements. This implementation of binary valued Backpropagation used optical matrix-vector multiplication³ to represent the forward information flow between network layers. Previous analog optical network memory systems have been described⁴.

© 1991 Optical Society of America