Abstract

An experimental optical vector–matrix multiplier is used for the implementation of an associative memory that is based on neural network models. An acoustooptic device in conjunction with a pulsed laser diode source is used as the input device through which an input vector is entered into the system. The interconnection matrix is stored in a computer-generated optical transparency and the product vector is detected on a CCD detector array. The output vector is serially read out, electronically processed (thresholded, for example), and then fed back to the acoustooptic device where it serves as the input for subsequent iterations. This relatively simple optical system can be used to simulate several variations of neural network models. This is done by selecting different matrices which simulate different interconnections among the neurons and also by selecting the electronic processing that is done in the feedback loop which simulates the action of the individual neurons. The storage capacity and the radius of attraction (the degree of similarity between the input and stored vectors that is required for associative recall) of such optical associative memories are characterized in detail.

© 1985 Optical Society of America