Abstract

We examine two incoherent electrooptical neural network processors. Both schemes use optical matrix-vector multiplication with feedback to model the alternating projection neural network (APNN). In the first, the Stanford matrix-vector multiplier architecture using a photographic negative as the spatial light modulator (SLM) was used. Feedback was accomplished with a detector array leading to a light emitting diode (LED) array whose intensities represent the state of the neurons. In the second architecture, active SLMs allow for the updating of the neuron interconnection matrix in real time. A Semetex magnetooptic SIGHTMOD encodes the matrix onto a light beam which is projected onto the write side of a Hughes liquid crystal light valve. Light from the LED array is reflected from the read side and the modulated intensity is detected and fed back to update the LEDs. Convergence for both implementations is assured if the stored data meet certain requirements¹ and if the thermal dark current is electrically subtracted after each iteration. Although their computing power and accuracy are currently limited by the physical size of the components and the SLM display resolution, respectively, the high operating speed of the electrooptical systems make them worth studying.

© 1989 Optical Society of America