Source localization in the brain remains an ill-posed problem unless further constraints about the type of sources and the head model are imposed. Human head is modeled in various ways depending critically on the computing power available and/or the required level of accuracy. Sophisticated and truly representative models may yield more accurate results in general, but at the cost of prohibitively long computer times and huge memory requirements. In conventional source localization techniques, solution source parameters are taken as those which minimize an index of performance, defined relative to the model-generated and clinically measured voltages. We propose the use of a neural network in the place of commonly employed minimization algorithms such as the Simplex Method and the Marquardt algorithm, which are iterative and time consuming. With the aid of the error-backpropagation technique, a neural network is trained to compute source parameters, starting from a voltage set measured on the scalp. Here we describe the methods of training the neural network and investigate its localization accuracy. Based on the results of extensive studies, we conclude that neural networks are highly feasible as source localizers. A trained neural network's independence of localization speed from the head model, and the rapid localization ability, makes it possible to employ the most complex head model with the ease of the simplest model. No initial parameters need to be guessed in order to start the calculation, implying a possible automation of the entire localization process. One may train the network on experimental data, if available, thereby possibly doing away with head models.
