Summary
2007 International Symposium on Nonlinear Theory and its Applications
2007
Session Number:19PM1-D
Session:
Number:19PM1-D-4
An Analysis of Non-periodic Oscillation During a Learning of a Complex-valued BAM
Ryuuichi Sasaki, Masaharu Adachi,
pp.537-540
Publication Date:2007/9/16
Online ISSN:2188-5079
DOI:10.34385/proc.41.19PM1-D-4
PDF download (303.2KB)
Summary:
Bidirectional associative memories have been proposed by Kosko[1]. Kosko’s model consists of two layers of neurons with synaptic connections between the layers. In the original model, the synaptic weight connections are determined by a correlation matrix of stored patterns. However, when the training pattern vectors are not orthogonal, cross-talk noise may occur in the system. To solve such problem, Oh.et.al. have been proposed Pseudorelaxation learning algorithm for BAM (PRLAB) which uses a projection method[2]. PRLAB does not require orthogonality or any special encoding of the training pairs and extremely increase the storage capacity. We extend the real-valued PRLAB algorithm to complex value. In the case that the complex-valued BAM trained by the extended PRLAB, we show that the storage capacity increases as the same as in the real-valued PRLAB. Moreover, it is also shown that when the complex-valued BAM fails to store the patterns to have high storage capacity, the time series of directional cosine during the training show non-periodic oscillation. We analyze the non-periodic oscillation in the present paper. As a result the non-periodic oscillation has a low-dimensional structure.