Abstract

We describe a two-layer optical neural network. The system uses photorefraction holograms to implement synaptic interconnections and liquid-crystal light valves (LCLV's) to perform nonlinear thresholding operations. Kanerva's sparse, distributed memory was implemented using this system, and its ability to recognize a handwritten-character alphabet (A-Z) has been demonstrated experimentally in Kanerva's model; the first layer has fixed, random weights of interconnections, and the second layer is trained according to an outer-product scheme. A total of 104 handwritten-character patterns, with four patterns from each class, were used to train the network. When a character pattern is presented at the input of the trained network, one of the 26 output neurons is switched on, indicating which character the input pattern is. For this network to solve more complicated problems, it is desirable to make both layers adaptive. We propose a new local learning rule that trains both layers based only on local information and a global, scalar error signal. This learning rule is much easier to implement than the back-error-propagation algorithm, which requires the backpropagation of error signals through the network, but it still guarantees that the network is trained by energy descent.

© 1990 Optical Society of America