Neural Networks with Node Gates which Divide and Conquer Problems Based on Local Difficulties

DOI Web Site
  • MURATA Junichi
    Faculty of Information Science and Electrical Engineering, Kyushu University
  • KAJIWARA Yoshitatsu
    Research and Development Laboratory, Panasonic Communications Co., Ltd.
  • HIRASAWA Kotaro
    Graduate School of Information, Production and Systems, Waseda University

Bibliographic Information

Other Title
  • 難易度に基づく分割統治機能をもつゲート付きニューラルネットワーク

Abstract

A neural network is proposed based on a divide-and-conquer scheme. The network has gates which control firing of its hidden nodes. By opening and closing the gates depending on input values, the network divides the input space into sub-regions and assigns its nodes to each of them to produce the desired output in that region. The division mechanism is constructed by learning. A new learning method is proposed which divides the space in accordance to the difficulties; areas with larger errors are divided into smaller sub-regions. Thus, the nodes in the network are more densely assigned to areas with higher difficulties to ‘conquer’ the areas appropriately. Function approximation examples are provided to illustrate the validity of the proposed network.

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top