スパースコーディングを用いた指数忘却を持つ連想記憶モデルの相互情報量による最適化

摘要

本研究では，近年，脳の機能に実在する可能性が高いとされる指数忘却を取り入れた連想記憶モデルについて研究する．この連想記憶モデルにはスパースコーディーングを適用し，指数忘却係数の最適化などを相互情報量を用いた数値計算により行う．そして，従来の方法である，方向余弦を用いた最適化と相互情報量を用いた最適化の比較を行う．%According to the previous researches, the associative memory model with forgetting process could avoid overloading due to exponential decay of its connection weights at each learning step for embedding a new pattern. However, the previous researches decided whether the network could retrieve the stored pattern correctly in an arbitrary manner, in other words, the dirrection cosine between the memory pattern and the stationary state of the network. Recently, it has been shown that the mutual information should be the most appropriate concept to measure the retrieval quality, especially for sparsely encoded networks but also in general. In this paper, from the view of information theory, we analyze the sparsely encoded associative memory model with forgetting process. We searched the optimal decay rate to maximize the mutual information at each ring rate by the numerical simulations. Moreover, we compare optimization of the decay rate utilizing mutual information with the ordinary one utilizing the direction cosine.