LEARNING METHOD FOR HIERARCHICAL NEURAL NETWORK

Japanese Patent JP3114276

B2

PURPOSE: To improve the universal applicability of a neural network by reducing the learning time and the computed variable and also securing a proper number of intermediate layer units.
CONSTITUTION: A MADALINE Rule II is used to input a set of an input signal and a teacher signal and to obtain an error E between these signals in a learning method for a hierarchical neural network consisting of unit which uses the signum function as an output function. Then the trial patterns are generated by inverting the output codes of the intermediate layer units in sequence and in the order of internal states of these units closer to zero. Then an error E' is obtained between the output signal and the teacher signal when such a trial pattern is given to an output layer. The flags are set up on an intermediate layer unit table to the intermediate layer unit having; both errors E and E' different from each other (S23, S24). Then the intermediate layer unit having no coincidence of flags is deleted as a non-contributive unit in reference to the intermediate layer unit for all learning sets every time a single time of learning is over (S26).

Masari Ichikawa

JP25751191A

December 04, 2000

October 04, 1991

Click for automatic bibliography generation  

Advantest Corporation

G06F15/18; G06G7/60; G06N3/08; G06N99/00; (IPC1-7): G06G7/60; G06F15/18

【文献】「階層型ニューラルネットワークの中間層素子数を自動削減する誤差逆伝搬学習アルゴリズム」松永 豊、中出 美彰、村瀬 一之著、電子情報通信学会技術研究報告、第91巻 第25号 頁9~14、1991年
【文献】「淘汰機能を有するバックプロバケーション(中間層ユニット数の削減と収束の高速化)」萩原 将文著、電子情報通信学会技術研究報告、第89巻 第104号 頁85~90、1990年

Kusano Taku (1 person outside)

Previous Patent: MANUFACTURE OF THIN FILM IMAGE SENSOR

Next Patent: METHOD OF CONSTITUTING AUTOMATIC CELL CLASSIFYING LOGIC