Investigating the performance of MLP classifiers where limited training data are available for some classes

The standard implementation of the back-propagation training algorithm for multi-layer Perceptron (MLP) neural networks assumes that there are equal number of samples for training each of the required classes. Where limited training data are available for one (or more) classes, sub-optimal performance may be obtained. We have demonstrated in a previous study [Parikh ct al., 1999. Proceedings of Condition Monitoring 1999, Swansea, UK] that, where unequal training class cannot be avoided, performance of the classifier may be substantially improved by duplicating the available patterns in the smaller class. In this study, we investigate whether the addition of random noise to the `duplicated' training patterns will further improve the classification performance. In the study conducted here, we conclude that the addition of noise does not give a consistent improvement in performance.