Learning method, neural network and computer for simulating said neural network

摘要

The invention relates to a training method implemented in a neural network which is carried out according to the algorithm of the gradient retropropagation. In order to determine the new synaptic coefficients with a learning period, the invention introduces the parameters which favour of the corrections based on the sign of the error at the beginning of the learning process and which gradually perform the corrections less sudden. This may be supplemented by other parameters, promoting a strategy layer by layer, accelerating the learning on the layers of the input with respect to the layers of the output. It is also possible to add a strategy operating on the whole of the neural network. / p & & p & application: neural networks in layers.