Optical Neural Networks Based on Distributed Holographic Gratings.
Abstract
This final report describes research in optical neural networks performed at Hughes Research Laboratories under a three year DARPA sponsored contract the advantages of optics for neural network implementations, including high storage capacity, connectivity, and very fine grained parallelism, was demonstrated. The optical neurocomputer developed under this program is based on a new type of holography which we call multiple grating holography, in which this approach reduces crosstalk and improves the utilization of the optical input device. In addition, this optical neurocomputer is the first and, to the best of our knowledge, the only one which is programmable and capable of implementing a wide variety of neural network models without any hardware adjustments. Successfully implemented neural networks included the Perceptron, Bidirectional Associative Memory, Kohonen, and backpropagation neural networks. Up to 10(4) neurons, 2 x 10(7) weights, and processing rates of 10(8) connection updates per second were achieved. Under this contract, we built an optical neurocomputer which utilizes a laser diode light source operating at 830 nm. This allowed us to reduce the size of the system. We also developed a new method for representing bipolar neural weights using coherent detection.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 13, 1996
- Accession Number
- ADA324598
Entities
People
- Yuri Owechko
Organizations
- HRL Laboratories