LEARNING METHOD FOR LEARNING MACHINE

Japanese Patent JP3238178

B2

PURPOSE: To adjust to easy-to-converge and well-balanced output by adjusting a parameter so as to minimize the value of a loss function decreasing with the increase of the difference between an output vector and a bad example teacher within teacher signals.
CONSTITUTION: A loss function calculating means 8 is constituted of the loss function calculating means 9 by a model teacher and the loss function calculating means 10 by the bad example teacher. In this case, the bad example teacher within the teacher signals is defined as data not to be outputted as the output vector, therefore, to be avoided. Then, by the difference (x) between the bad example teacher and the output vector, the loss function (r) for decreasing along with the increase of the difference (x) is calculated with an equation (r)=exp(-x2). Then, when the value of the loss function is larger than a prescribed value, a control parameter is updated so as to reduce the value of the loss function until the value of the loss function becomes smaller than the prescribed value, the loss function is repeatedly calculated and when the calculated loss function reaches the value smaller than the prescribed value, the latest updated control parameter is integrated into the learning machine.

Koumi Yamada

JP34538791A

December 10, 2001

December 26, 1991

Click for automatic bibliography generation  

Toshiba Corporation

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

JP2170265A
JP520291A
JP5165801A

【文献】山田、田中、「教師信号との出力の差が増大する学習方式」、1991年電子情報通信学会秋季大会講演論文集(分冊6)、社団法人電子情報通信学会・発行(1991年8月15日発行、特許庁情報館受入日:平成3年11月13日)

Hidekazu Miyoshi (4 outside)

Previous Patent: Sterilization bag

Next Patent: METHOD FOR COOLING TOP ROLL FOR PRODUCING ELECTRIC RESISTANCE WELDED TUBE