A Study of Learning Issues in Feedforward Neural Networks

Abstract

When training a feedforward stochastic gradient descendent trained neural network, there is a possibility of not learning a batch of patterns correctly that causes the network to fail in the predictions in the areas adjacent to those patterns. This problem has usually been resolved by directly adding more complexity to the network, normally by increasing the number of learning layers, which means it will be heavier to run on the workstation. In this paper, the properties and the effect of the patterns on the network are analysed and two main reasons why the patterns are not learned correctly are distinguished: the disappearance of the Jacobian gradient on the processing layers of the network and the opposite direction of the gradient of those patterns. A simplified experiment has been carried out on a simple neural network and the errors appearing during and after training have been monitored. Taking into account the data obtained, the initial hypothesis of causes seems to be correct. Finally, some corrections to the network are proposed with the aim of solving those training issues and to be able to offer a sufficiently correct prediction, in order to increase the complexity of the network as little as possible.