online read us now
Paper details
Number 4 - December 1995
Volume 5 - 1995
On the robustness of learning in the multi-layer perceptron
Paul Williams, Andrew W.G. Duller
Abstract
The back-propagation algorithm, for training multi-layer perceptrons tends to be slow to converge to a final solution and many methods have been proposed for improving this. One technique takes advantage of an alternative training error criterion, however, we show that this reduces the robustness of the learning in the presence of outliers in the input data. Two examples are used to show the characteristics of the learning methods, one a test problem and the other from a "real-world" problem.
Keywords
-