Summary
the 2014 International Symposium on Nonlinear Theory and its Applications
2014
Session Number:C2L-C
Session:
Number:C2L-C5
Effectivity of Randomness in Memory Patterns of Recurrent Neural Network Model
Hiroto Hatano, Jousuke Kuroiwa, Tomohiro Odaka, Izumi Suwa, Haruhiko Shirai,
pp.524-527
Publication Date:2014/9/14
Online ISSN:2188-5079
Summary:
In this paper, we investigate effectivity of the randomness in memory patterns of a recurrent neural network model referred as RNN hereafter. We have shown that for the memory patterns with a certain structure, their basin volumes and furthermore visiting measures of the basins become smaller. In realizing a function of the memory search based on chaotic wandering in a chaotic neural network model referred as CNN, it is important to ensure that basin volumes of the memory patterns and visiting measures of the basins are sufficiently large. Therefore, we investigate how to construct the memory patterns which gives sufficiently large basin volumes of theirs in RNN, focusing on the randomness in the memory patterns. We apply 11 kinds of the memory patterns with changing the ratio of the randomness. As the randomness increases, basin volumes of the memory patterns increase. The basin volumes of the memory patterns without the randomness is quite smaller than those of pseudo memory patterns. Thus, the randomness in the memory patterns is practical in ensuring that their basin volumes are sufficiently large.