直交学習方式ニューラルネットワークの収斂性

概要

論文の詳細を見る
An orthogonal learning method is applied to pattern recognition using a multilayer neural network. Numerical simulations of neural networks are performed for recognizing plural numbers of 20-bit and 256/bit character patterns using standard three-layer networks composed of one intermediate hidden layer and input and output layers. It is confirmed that the initial synapse weights greatly influence the convergence of the learning process. Furthermore, high-order functions such as, the problem of counting the number "1" in a 4- and 6-bit binary patterns and the problem of classifying the number "1" in 9-bit patterns into 2 groups are investigated. It is shown that the convergence of learning is one of the most fundamental neural network problem to be overcome. A new learning method is proposed for accelerating the convergence of training of a neural network. The effectiveness of the new learning method is verified.
社団法人日本機械学会の論文
1990-07-25