All-optical implementation of the inverted neural network model

The Inverted Neural Network (INN) model especially designed for optical implementation is presented. This model takes into account the physical constraints imposed by conventional optical components and ensures that all the connections are positive. Thus subtraction of light intensities is not required for implementation nor are electronic computations. In the laboratory realization of the INN model a liquid crystal light valve fulfills the function of an array of neurons while an array of subholograxns serves as the interconnects. The overall network was tested with 64 neurons and four stable states. 1. INTRODTJCTION Several problems have to be solved in order to realize neural network models optically. The first problem stems from the fact that most models consist of both positive and negative interconnects. Conventional solutions to this problem usually involve electronic subtractions or dynamic threshold levels. It has been recently shown1''2 how the Hopfleld model can be modified into a model with positive values only. An all-optical realization of the modified model for 16 neurons and two stable states has also been demonstrated1 . The second problem that has to be considered is the fact that most of the optical devices that are used for realizing the neurons cannot exhibit very sharp thresholding. Although neurons with graded response have collective computational properties like those of two state neurons the memory capacity of Hopfield-type networks decreases as the slope of the neural