On Simple Adaptive Momentum

Abstract

Simple Adaptive Momentum was introduced as a simple means of speeding the training of multi-layer perceptrons (MLPs) by changing the momentum term depending on the angle between the current and previous changes in the weights of the MLP. In the original paper, the weight changes of the whole network are used in determining this angle. This paper considers adapting the momentum term using certain subsets of these weights. This idea was inspired by the author's object oriented approach to programming MLPs, successfully used in teaching students: this approach is also described. It is concluded that the angle is best determined using the weight changes in each layer separately.