A modified backpropagation algorithm for neural classifiers

A variation of the traditional backpropagation algorithm is presented as an alternative approach for training feedforward neural network based classifiers. This method aims to reduce the tendency of these classifiers to spend most of their time trying to achieve unnecessarily low mean square errors in very populated regions of the pattern space, while almost ignoring patterns in sparse regions of it until a large number of training steps occurs. An application of the new algorithm in particle discriminators for high energy physics experiments is shown.