Improvements to the conventional layer-by-layer BP algorithm

Abstract

This paper points out some drawbacks and proposes some modifications to the conventional layer-by-layer BP algorithm. In particular, we present a new perspective to the learning rate, which is to use a heuristic rule to define the learning rate so as to update the weights. Meanwhile, to pull the algorithm out of saturation area and prevent it from converging to a local minimum, a momentum term is introduced to the former algorithm. And finally the effectiveness and efficiency of the proposed method are demonstrated by two benchmark examples. © Springer-Verlag Berlin Heidelberg 2005.