Avoiding poor local minima in training multilayer perceptrons

A training method is reported that adaptively selects a risk-averting error criterion to suit the function under approximation and the noise statistics of the training data so as to include fine features of the function and its segments under-represented in the training data. A companion paper also presented at ICONIP'01 proves that the method has the ability to avoid poor local minima of the selected criterion. Numerical examples given illustrate the efficacy of this training method.