Abstract

Two-dimensional neural networks based on the Hopfield model were proposed by Farhat and Psaltis¹ and Caulfield.² Their models require a large size of spatial light modulator or a large number of holographic elements. We propose a similar neural architecture, which uses a high-resolution video monitor as the programmable associative memory mask, and a moderate size SLM as an input matrix. An array of lenslets is used to provide the interconnections between the associative memory mask and the input matrix. With this architecture, an adaptive neural network of large size can be easily synthesized, and the limitations of low spatial bandwidth product and small dynamic range of available SLMs can be alleviated. To increase the system capacity, accuracy and speed, orthogonal projection algorithm, adaptive masking, and thresholding are used. This proposed system has the capability of improving the robustness and fault tolerance.

© 1988 Optical Society of America