Accession Number : ADA200445

Title :   Connectionist Models for Intelligent Computation.

Descriptive Note : Annual technical rept. 1 Sep 87-31 Aug 88,

Corporate Author : MARYLAND UNIV COLLEGE PARK

Personal Author(s) : Chen, H H ; Lee, Y C

PDF Url : ADA200445

Report Date : 31 Aug 1988

Pagination or Media Count : 30

Abstract : We have continued our study of higher order neural networks. The superior processing power capacity and speed of the higher order neural network has been demonstrated for many tasks including text to speech, character recognitions, noise removal, time series prediction etc. Currently, we are applying it to the speech recognition problem. We have constructed a neural network to learn the task of stereopsis from random dot stereogram. The connection weights of the network are computed analytically from the Hebbion learning rule. The results show that the continuity and uniqueness constraints first proposed by Marr and Poggio are learned automatically. We proposed a novel scheme (PSIN) to automatically build a neural network while learning. The new scheme takes advantage of both the parallel and sequential strategies to solve a pattern classification or decision problem. We optimize an entropy measure to encourage the network to extract the best feature first to classify the pattern. Preliminary test of this new scheme shows that PSIN performs superior than the back propagation scheme in hard problems. (KR)

Descriptors :   *NEURAL NETS, CLASSIFICATION, COMPUTATIONS, DECISION MAKING, ENTROPY, NETWORKS, NOISE, PATTERNS, PREDICTIONS, PROCESSING, PROPAGATION, SEQUENCES, SPEECH RECOGNITION, STEREOSCOPES, STRATEGY, TIME SERIES ANALYSIS, VISION, WEIGHT

Subject Categories : Cybernetics

Distribution Statement : APPROVED FOR PUBLIC RELEASE