Batch-sequential algorithm for neural networks trained with entropic criteria

Abstract

The use of entropy as a cost function in the neural network learning phase usually implies that, in the back-propagation algorithm, the training is done in batch mode. Apart from the higher complexity of the algorithm in batch mode, we know that this approach has some limitations over the sequential mode. In this paper we present a way of combining both modes when using entropic criteria. We present some experiments that validates the proposed method and we also show some comparisons of this proposed method with the single batch mode algorithm. © Springer-Verlag Berlin Heidelberg 2005.