Title:
【発明の名称】ニューラルネットワーク学習装置の重み更新方式
Document Type and Number:
Japanese Patent JP3002563
Kind Code:
B2
Abstract:
PURPOSE:To sharply shorten a learning time in a system for updating the weight of a network in a neural network learning device. CONSTITUTION:In the learning device provided with a learning execution part 18 having a neural network 16 formed by inserting an intermediate layer 14 between an input layer 10 and an output layer 12, a weight updating part 20 for finding out weight variables among respective layers 10, 12, 14 from the output of the network 16 and a teacher signal and updating the weight of the network 16 and a learning pattern storing part 22 for outputting an input pattern to the network 16 and applying the teacher signal to the weight updating part 20, the updating part 20 includes a means 24 for storing weight variables, a means 26 for judging whether the weight of the network 16 starts to be converged or not based upon the stored weight variables and a means 28 for forecasting the convergent value of the weight from the stored weight variable.
Inventors:
Chikako Matsumoto
Tatsuro Matsumoto
Tatsuro Matsumoto
Application Number:
JP15348791A
Publication Date:
January 24, 2000
Filing Date:
June 25, 1991
Export Citation:
Assignee:
富士通株式会社
International Classes:
G06F15/18; G06G7/60; G06N3/08; G06N99/00; (IPC1-7): G06F15/18
Domestic Patent References:
| JP2231670A | ||||
| JP1271888A |
Attorney, Agent or Firm:
Giichiro Ito