Abstract

Optical techniques have shown promise for massively parallel interconnection of simple processors. These techniques may offer advantages in parallel computing and neural network implementation. The lenslet array processor is an incoherent optoelectronic processor that offers matrix algebraic operations with 2-D arrays of input and output data. Inner and outer products on bipolar analog data can be performed, as well as binary interconnections of 2-D arrays. Bipolar data is accommodated by using a combination of optical and electronic processing steps. These operations form the basis for implementation of a wide variety of neural network models. Incorporation of SLMs allows implementation of adaptive networks; competitive learning and multiple parallel Perceptron network implementations are described. An experimental system using video technology and incoherent, polychromatic light is described. A frame-grabber card in a PC stores and processes images being sent to and received from the optics. Subtraction is performed electronically in the PC. A liquid crystal TV controls the input to the network, and a CRT writes the interconnection weights onto a liquid crystal light valve. A CCD camera captures the result of the weighted optical interconnection.

© 1991 Optical Society of America