Title:
【発明の名称】多値ニューラルネットワーク学習方式
Document Type and Number:
Japanese Patent JP3367295
Kind Code:
B2
Abstract:
PROBLEM TO BE SOLVED: To fast and stably acquire a desired output signal in a learning system where an optimum weight coefficient updating error signal is generated from the least absolute error value or an output unit of an incorrect answer and then the weight coefficient is updated. SOLUTION: In a binary neural network which performs the learning by means of a binary teacher signal, the weight coefficient updating error signals having the inverse polarity to the error signal and inversely proportional to the error signal are generated by the weight coefficient updating error signal generators 10 to 10M against the output unit which transmits the binary output unit signal equal to the binary teacher signal when the absolute value of the error signal obtained by subtracting the output unit signal from the teacher signal is smaller than the 1st threshold. The generators 10 to 10M generate the weight coefficient updating error signals having the smaller amplitudes than the error signal and the same polarity as the error signal when the absolute value of the error signal is larger than the 1st threshold. Then, a local minimum capture state detector 15 controls the amplitudes and also the 1st threshold. Thus, the weight coefficient is updated by means of at least the generators 10 to 10M and the detector 15 and based on the weight coefficient updating error signal.
Inventors:
Yatsuka Yotaro
Application Number:
JP22701495A
Publication Date:
January 14, 2003
Filing Date:
August 14, 1995
Export Citation:
Assignee:
KDDI Co., Ltd.
International Classes:
G06G7/60; G06F15/18; G06N3/08; (IPC1-7): G06N3/08
Domestic Patent References:
| JP415860A | ||||
| JP4251392A | ||||
| JP3218559A |
Other References:
【文献】八塚陽太郎,「誤差摂動による2値3層ニューラルネットワークの高速収束化」,電子情報通信学会1995年総合大会講演論文集,日本,社団法人電子情報通信学会,1995年 3月10日,第1分冊,pp.24
Attorney, Agent or Firm:
Keiichi Yamamoto