Data processing device with network structure and its learning processing method
US5239618A · kind A · utility
Assignee
Inventors
Key dates
| Filing date | Dec 20, 1990 |
| Grant date | Aug 24, 1993 |
| Priority date | — |
| Expiry date | Dec 20, 2010 |
Classification
- Technology area (CPC G)Physics
- CPC primaryG06N3/09
- WIPO fieldComputer technology
- WIPO sectorElectrical engineering
Abstract
An output layer in a layered neural network uses a linear function or a designated region (linear region) of a threshold function instead of the threshold function to convert an input signal to an analog output signal When the basic unit uses the linear function, a limiter for limiting the output to a region between 1.0 and 0. When the basic unit uses the designated linear region of the threshold function, a limiter limits the output to a region between 0.8 and 0.2. Upon a learning operation, the error propagation coefficient is determined as a constant value such as 1/6 and when the majority of the desired values are 1 or near 1, an error value regarding the opposite desired value 0 is amplified, and when the output values become equal to or more than 1, it is deemed that there is no error with regard to the output of more than 1 in case of many outputs 1, thereby speeding up an operation of updating the weight.
Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.