Determining input features for multilayer perceptrons

Abstract A method of selecting salient features from a superset of features is developed. Two metrics are used to measure the average saliency of injected noise. A confidence interval constructed around this average allows for the identification of features that contribute little to classification. This feature selection method is applied to an exclusive-or (XOR) problem containing noise and a four-class problem. This method of determining input features is shown to result in more accurate, faster training multilayer perceptron classifiers.