Classifier's Complexity Control while Training Multilayer Perceptrons

We consider an integrated approach to design the classification rule. Here qualities of statistical and neural net approaches are merged together. Instead of using the multivariate models and statistical methods directly to design the classifier, we use them in order to whiten the data and then to train the perceptron. A special attention is paid to magnitudes of the weights and to optimization of the training procedure. We study an influence of all characteristics of the cost function (target values, conventional regularization parameters), parameters of the optimization method (learning step, starting weights, a noise injection to original training vectors, to targets, and to the weights) on a result. Some of the discussed methods to control complexity are almost not discussed in the literature yet.