New perspective for structural learning method of neural networks

Junzo Watada*

*この研究の対応する著者

    研究成果: Conference contribution

    抄録

    A neural network is developed to mimic a human brain. The neural network consists of units and links that connect between units. Various types of neural networks are categorized into two classes: (1) back-propagation hierarchical neural network and (2) mutual-connected neural network. Generally speaking, it is hard to fix the number of units to build a neural network for solving problems. So the number of units is decided on the basis of experts' experience. In this paper, we explain a learning method how to decide the structure of a neural network for problems. The learning method is named structural learning. Even if we give a sufficient number of units, the optimal structure will be decided in the process of learning. The objective of the paper is to explain the structural learning of both hierarchical and mutual connecting neural networks. Both networks obtained and showed the sufficiently good results. In the stock forecast by a general neural network, the operation and the system cost are very large because a lot of numbers of hidden layer units in the network are used. This research tried the optimization of the network by the structured learning, and evaluated the practicality.

    本文言語English
    ホスト出版物のタイトルLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    ページ231-240
    ページ数10
    4529 LNAI
    出版ステータスPublished - 2007
    イベント12th International Fuzzy Systems Association World Congress, IFSA 2007 - Cancun
    継続期間: 2007 6 182007 6 21

    出版物シリーズ

    名前Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    4529 LNAI
    ISSN(印刷版)03029743
    ISSN(電子版)16113349

    Other

    Other12th International Fuzzy Systems Association World Congress, IFSA 2007
    CityCancun
    Period07/6/1807/6/21

    ASJC Scopus subject areas

    • コンピュータ サイエンス(全般)
    • 生化学、遺伝学、分子生物学(全般)
    • 理論的コンピュータサイエンス

    フィンガープリント

    「New perspective for structural learning method of neural networks」の研究トピックを掘り下げます。これらがまとまってユニークなフィンガープリントを構成します。

    引用スタイル