Analog Optical Neural Nets: A Noise Sensitivity Analysis
Abstract
Neural networks represent a promising alternative to traditional artificial intelligence approaches. The developement of analog optical implementations of neural networks such as the multilayer perceptron with learning by backward error propagation (BEP) requires an understanding of the noise sensitivity of such architectures. The objective of this program is to study the effects of component and system noise on the performance of such optical implementations. The method used is computer simulation. In this first phase of the program, the one-hidden layer perceptron with back propagation was simulated using a simplified, device-independent noise model. The results point to a distinct noise threshold above which the learning mechanism is corrupted. The efficiency of learning based on variations within back propagation on the initializing method was also studied. In the next phase, a device-dependent noise model will be used. To this end a plausible all-optical architecture capable of both the forward pass and backward error propagation steps of training data presentation has been proposed.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 30, 1990
- Accession Number
- ADA226789
Entities
People
- James J. Levy
- Michael W. Haney
- Ravindra A. Athale
Organizations
- Braddock Dunn & McDonald