Sensitivity methods for variable selection using the MLP

Variable selection is a key factor in achieving reliable classification. One of the most commonly used traditional statistical methods of variable selection is the p-value test which for a logistic regression classifier, involves estimating the optimal value for the weights and also their accuracy. A corresponding method for the multilayer perceptron (MLP) is to select variables on the basis of a linearization of the network about sample operating points. As a result of this linearisation, we can call this class of methods sensitivity methods. In this paper it is shown that an extension of the p-value test to the MLP would not lead to a reliable method to identify the relative importance of different input variables for classification. The application of sensitivity methods to the MLP is reviewed by considering two possible implementations applied, first to the XOR problem with additional redundant inputs, and secondly to a real-world classification problem. It is concluded that sensitivity methods constitute a useful analogue to the p-value test for variable selection with the MLP. These methods have general applicability for variable selection in signal processing