Abstract

An adaptive neural-network architecture is presented that incorporates (1) a simultaneous incoherent/coherent holographic recording and reconstruction technique that permits simultaneous updates of all weights in a multiplexed volume holographic interconnection during each iteration of a neural learning algorithm¹; (2) a double-angle multiplexing arrangement in which each pixel of the object-beam spatial light modulator (SLM) is illuminated by a set of mutually incoherent beams, each at a different angle; and (3) optoelectronic SLM's for the neural input and training planes, each with dualchannel inputs and outputs.² The architecture incorporates modularity; capability for lateral, feedforward, and feedback interconnections in a multilayer network; effectively bipolar signals and weights; and capacity to implement a variety of network models and supervised or unsupervised learning algorithms. Results of computer simulations and laboratory experiments on selected aspects of the architecture will also be presented.

© 1990 Optical Society of America