专利名称:Neural network发明人:Takami, Junichi, 4-506,
Takanoharaekinishidanchi
申请号:EP90108123.2申请日:19900427公开号:EP0395112B1公开日:19991027
摘要:PURPOSE:To construct a neural network having high performance with the smallnumber of times of learning by providing a means for controlling integrally each weightprovided between each unit. CONSTITUTION:In the case of an example in which theneural network of exclusive OR is constructed by three layers of two units of an inputlayer, two units of an intermediate layer, and one unit of an output layer, units u1 - u5 arecoupled by weights W1 - W6, and each unit u1 - u5 is allowed to have an input/ outputcharacteristic shown by a sigmoid function having gain coefficients mu1 - mu5. Also, ameans for controlling integrally the weight provided between the units, for instance, theoriginal non-linear coefficient of the unit for giving a function so as to make the outputvalue of the whole of the neural network approach to a correct solution value by the gaincoefficient mu is formed in each unit. In such a manner, the number of times of learning isdecreased, and also, the neural network having high performance is constructed.
申请人:VICTOR COMPANY OF JAPAN
地址:JP
国籍:JP
代理机构:Grünecker, Kinkeldey, Stockmair & Schwanhäusser Anwaltssozietät
更多信息请下载全文后查看
因篇幅问题不能全部显示,请点此查看更多更全内容