Title:
【発明の名称】ニューラルネットワーク
Document Type and Number:
Japanese Patent JP3337504
Kind Code:
B2
Abstract:
PURPOSE:To obtain the neural network which hardly causes misrecognition, recognizes even an extremely similar pattern, and practically varies a load value by a simple update procedure even in case of variation in input. CONSTITUTION:The neural network consists of an input layer which has at least N input units, an intermediate layer which has K intermediate unit groups of N intermediate units, corresponding to the input units, corresponding to output categories and a unit applied with a bias, and a 1st output layer having K output units. The load values of connections between the respective units of the input layer and the intermediate unit groups are all 1, input/output characteristics of the intermediate units are represented with representative values by N elements of input pattern groups belong to the corresponding output categories and a projection membership function generated on the basis of statistical quantities showing variance, and the load values of connections between the intermeidate layer and the 1st output layer are determined by learning on the basis of a delta rule by using membership values outputted from the membership function as the outputs of the intermediate unit group.
More Like This:
Inventors:
Toshiharu Takei
Yasuhiro Takemura
Yasuhiro Takemura
Application Number:
JP27287192A
Publication Date:
October 21, 2002
Filing Date:
October 12, 1992
Export Citation:
Assignee:
Sumitomo Osaka Cement Co., Ltd.
International Classes:
G06F15/18; G06G7/60; G06N3/00; G06N3/06; G06N99/00; (IPC1-7): G06N3/00; G06F15/18; G06G7/60; G06N3/06
Domestic Patent References:
| JP2189635A | ||||
| JP4127239A | ||||
| JP423088A | ||||
| JP476678A | ||||
| JP492901A | ||||
| JP4536A | ||||
| JP2292602A |
Attorney, Agent or Firm:
Yutaka Kuramochi