Training a neural network using differential input
US6128609A · kind A · utility
37Cited by
3References
17Claims
0Family size
Assignee
Inventor
Key dates
| Filing date | Oct 14, 1997 |
| Grant date | Oct 3, 2000 |
| Priority date | — |
| Expiry date | Oct 14, 2017 |
Classification
- Technology area (CPC G)Physics
- CPC primaryG06N3/09
- WIPO fieldComputer technology
- WIPO sectorElectrical engineering
Abstract
A neural network is trained using a training neural network having the same topology as the original network but having a differential network output and accepting also differential network inputs. This new training method enables deeper neural networks to be successfully trained by avoiding a problem occuring in conventional training methods in which errors vanish as they are propagated in the reverse direction through deep networks. An acceleration in convergence rate is achieved by adjusting the error used in training to compensate for the linkage between multiple training data points.
Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.