Training multilayer perceptrons in the presence of measurement outliers

Instead of the robust estimation criteria from statistics, a new training method using a continuum of modified risk-seeking criteria with a negative risk-sensitivity index is proposed for training neural networks with data containing outlying measurement noises. In contrast to the ordinary methods using a fixed training criterion or a fixed annealing schedule for the training criterion in a training session, the new method continues adjusting adaptively the risk-sensitivity index to tune to the measurement outliers for best reducing their effects on the training