发明名称 |
Learning process for neural network - comparing sample signals and categorising for use in backward learning process |
摘要 |
The learning process is used with a neural network that has an input layer (1) of 7 frames together with a further two layers (2,3) and an output layer (4). Each data element has a time period of 70 ms. In a first operation a learning sample is selected and the differences to other samples is determined in a second operation. Samples are categorised and the weightings determined. When all values are obtd. a backward learning algorithm is executed to improve the results. USE/ADVANTAGE - For phoneme comparison in speech recognition system. Improves the learning rate of the network.
|
申请公布号 |
DE4208727(A1) |
申请公布日期 |
1992.10.01 |
申请号 |
DE19924208727 |
申请日期 |
1992.03.18 |
申请人 |
ATR INTERPRETING TELEPHONY RESEARCH LABORATORIES, KYOTO, JP |
发明人 |
KOMORI, YASUHIRO;SAGAYAMA, SHIGEKI, KYOTO, JP |
分类号 |
G06F15/18;G06G7/60;G06N3/08;G06N99/00 |
主分类号 |
G06F15/18 |
代理机构 |
|
代理人 |
|
主权项 |
|
地址 |
|