Title:
LEARNING DATA STORAGE FOR NEURAL NETWORK
Document Type and Number:
Japanese Patent JP3232595
Kind Code:
B2
Abstract:
PURPOSE: To attain the effective updating learning without increasing the input data and the teacher data by degenerating two input data having the highest resemblance into a single piece of synthetic input data.
CONSTITUTION: The new input data are added to an input data group of a data base in a step 300, and an optional pair of input data are generated in a step 302. Then the resemblance is calculated between both input data in a step 304. The same processing is repeated with combination of all input data in a step 306. Then a pair of data having the highest resemblance is decided in a step 310. In a step 312, the synthetic input data that undergone the weighted average with the count value defined as a weighting coefficient is calculated together with the synthetic teacher data. Then the synthetic count value is obtained in a step 314, and two input data having the lowest resemblance are deleted out of a group of input data in a step 316 together with addition of the synthetic input data. In the same way, a corresponding group of teacher data are also updated.
Inventors:
Takao Yoneda
Tomoya Kato
Masaru Yamanaka
Hattori Shiho
Tomoya Kato
Masaru Yamanaka
Hattori Shiho
Application Number:
JP21593891A
Publication Date:
November 26, 2001
Filing Date:
July 31, 1991
Export Citation:
Assignee:
Toyota Koki Co., Ltd.
International Classes:
G06F15/18; G06G7/60; G06N3/08; G06N99/00; (IPC1-7): G06N3/08; G06G7/60
Domestic Patent References:
| JP4184668A | ||||
| JP4299442A |
Attorney, Agent or Firm:
Osamu Fujitani