Learning method and neural network structure.

摘要

The invention relates to a learning method implemented within a neural network which operates in accordance with the gradient back-propagation algorithm. In order to determine the new synaptic coefficients with a minimal learning period, the invention introduces parameters which favour corrections based on the sign of the error at the start of learning, and which gradually execute less abrupt corrections. This can be supplemented with other parameters favouring a layer-by-layer strategy, accelerating learning on input layers relative to output layers. It is also possible to append a strategy operating throughout the neural network. Application: layered neural networks. <IMAGE>