Abstract

We have developed an optical neurocomputer for implementing neural networks based on real-time holography in photorefractive materials. Optical techniques show great promise for implementing neural network models because of the parallelism, connectivity, and high storage capacity of optics. It is generally agreed that optics will most likely be used for implementing very large neural networks that are impractical to simulate or implement electronically. In our system weighted connections between neurons are recorded in barium titanate as photorefractive holograms. We have virtually eliminated distortion and crosstalk between holograms by using beam fanning to distribute each connection weight among many angularly and spatially distributed photorefractive gratings. We have successfully demonstrated several optical neural networks including the Perceptron Bidirectional Associative Memory (BAM) and back propagation with a single hidden layer. All of these networks were implemented on the same hardware using a single crystal and SLM by simply changing the software. We have demonstrated networks with up to 7,000 neurons, 800,000 weights, and PC bus-limited learning rates of up to 10⁶ connection updates per second.

© 1992 Optical Society of America