Presentation Information
[N-1-22]On Clustering Performance of Hysteresis Neural Networks
◎Seigo Nakamura1, Toshimichi Saito1 (1. Hosei Univ.)
Keywords:
Hysteresis neural networks,Binary data,Clustering
The continuous-time hysteresis neural network (CT-HYN) is a continuous-time recurrent neural network characterized by binary hysteresis activation function. The dynamics is described by piecewise linear ordinary differential equation of real valued state variables. In this paper, we apply the CT-HYN to clustering of a binary data set. As the binary data set, a true-false chart consisting of students and problems is handled. In the clustering, changes in performance caused by appropriate parameter settings are investigated. To set the connection matrix, correlation learning and sparsification are used. Furthermore, a threshold parameter of the activation function is varied simultaneously with the sparsification of the connection matrix. In the evaluation of the clustering, the uniformity of cluster elements is used. We confirm that the clustering performance of the CT-HYN can be improved by appropriate sparsification of the connection matrix and appropriate setting of the threshold parameter.