hrvatski

Neiterativni postupak učenja parametara neuronskih mreža

The basic idea of this work is that in classification problems better results can be achieved if each hidden unit has its own matrix W for calculation of weighted distances. In this case hidden layer parameters (matrices W and positions of the centres t) can be learned much faster, locally, for each class of the problem separately, using well known statistical methods, than by optimization process proposed by Poggio and Girosi. With chosen hidden layer parameters optimal linear output layer weights can be calculated in single step (noniterative approach) using all training samples.

Neuronske mreže; radijalne bazne finkcije; neiterativno učenje

nije evidentirano