Back-propagation training using a least mean power error function

Summary form only given, as follows. Like many gradient descent algorithms, back-propagation can become trapped in a local minimum which corresponds to a non-optimal network configuration. At a typical local minimum, most outputs are essentially correct with only a few outputs exhibiting gross errors. The authors propose a modified error function in which the output errors are raised to a power larger than the nominal two. This is intended to alleviate the local minimum problem by focusing the training process on the large output errors. Simulation results were obtained for simple computational networks which are prone to local minima