Adaptation of the Regularization Parameters in the Nm-Delta Networks

The paper describes an application of regularization techniques to an automatic choice of parameters driving the learning process in the NM-Delta neural network architecture. The heterogeneous learning algorithm is identified as very similar to the Levenberg-Marquardt method but with a considerably smaller computational cost and different justification of parameter selection. The performance of the modified algorithm proves to be comparable with that of the Levenberg-Marquardt.