半直交連想記憶におけるニューロンの活動電位の分布に関する検討

概要

論文の詳細を見る
Artificial neural networks(ANN)have been attracting a lot of attention, for they can efficiently process information in parallel. So far, a great number of ANN models have been proposed. These models can be roughly divided into two major types. One is the associative memory(AM)type and the other is the multi-layer type. Semi-orthogonal associative memory(SAM)is one of the AM models, which has a large memory capacity and a good recalling property. However, like other AM models, such as, Hopfield model, BAM, there exist spurious states in SAM. When SAM is used as an information processing system, say, pattern recognition system, it is often expected that the recalling outputs are its learned patterns. Thus, in the recalling processes of SAM, it is necessary to discriminate whether the recalling output is spurious or not for any initial input. In order to do this, it should be clearly analyzed the properties of this model at stable states(SS). In this paper, we shall discuss this problem and try to analyze the statistical properties of SAM at SS. Making use of the neuro-dynamics, we derive the neurons active potential(NAP)distribution at stable states and investigate the different statistical properties of SAM at the spurious and non-spurious stable states(SSS and N-SSS).
社団法人電子情報通信学会の論文
1996-09-18