Extension of a training set for artificial neural networks and its application to brain source localization

The problem of training sets with inadequate training patterns is addressed. Learning based on such sets results in poor generalizations. We introduce an extension procedure to augment training sets in order to provide improved generalization. The original training set is used to provide hints, along with some statistical information, in the extension procedure. We show that if a mathematical model is available for a poorly observed physical process, then the extension of the inadequate training set is possible. The procedure is applied to the brain source localization problem. Our experiments results show that learning based on the extended training set is superior, with robust generalization, to learning based on the initial training set.