Abstract

We describe a two layer optical network that is trained to recognize in real time ″Denk,″ who is a student in our group. The network is implemented with liquid crystal spatial light modulators for the neural planes and lithium niobate photorefractive crystals for the interconnections. The network has approximately 60,000 units at the input plane, 30 hidden units, and a single output unit. The network is trained with a video segment 1.5 min long, from which 180 frames were selected for learning. Specifically, each hidden unit was trained to respond to six frames. The trained network classified the rest of the training tape almost flawlessly. The system was then tested by presenting through a TV camera real time inputs of Denk and other members of our group. The system almost never mis-classifies other people as Denk and exhibits remarkable tolerance to changes in aspect, illumination, and facial expression.

© 1992 Optical Society of America