Accession Number : ADA320927
Title : Photonic Technology Development for Densely Interconnected Neural Networks: Augmentation Award.
Descriptive Note : Final rept. 1 Sep 93-31 Aug 96,
Corporate Author : UNIVERSITY OF SOUTHERN CALIFORNIA LOS ANGELES DEPT OF ELECTRICAL ENGINEERING
Personal Author(s) : Jenkins, B. K.
PDF Url : ADA320927
Report Date : 24 JAN 1997
Pagination or Media Count : 4
Abstract : Technical accomplishments under the grant, 'Photonic Technology Development for Densely Interconnected Neural Networks: Augmentation Award' (AFOSR Grant No. F49620-93-1-0445), B. K. Jenkins, P.I. are described. They include an analysis of convergence conditions and properties of backward error propagation learning in photorefractive based optical neural networks. The analysis includes implementations based on fully coherent single source architectures and on incoherent/coherent multiple source architectures. Also analyzed in terms of their effect on optical neural network learning are spatial light modulator limitations such as finite modulator contrast ratio, detector noise, and limited detector dynamic range. Additionally, we have developed a technique for compensating for photorefractive grating decay during neural network learning, by varying two parameters, spatial light modulator gain and photorefractive crystal exposure energy, according to a prescribed schedule.
Descriptors : *NEURAL NETS, *COMPUTER ARCHITECTURE, *PHOTOREFRACTIVE MATERIALS, SPATIAL DISTRIBUTION, CONTRAST, PARAMETERS, LIMITATIONS, DYNAMIC RANGE, LIGHT MODULATORS.
Subject Categories : Electrooptical and Optoelectronic Devices
Computer Programming and Software
Distribution Statement : APPROVED FOR PUBLIC RELEASE