Using self organising feature maps for feature selection in supervised neural networks

The definition of the input-output pairs used to train a supervised artificial neural network (ANN) is at the very first stage of the development of an ANN model, and is fundamental to the usefulness of the trained model. This paper considers an automatic method for selecting the two `best´ measurements, from a potentially large set of available inputs, on which a ANN model can be trained. The inputs that are chosen are `best´ in the sense that they provide more information on the value of the output than any other two way combination of input vectors. The method is sensitive to linear, nonlinear and nonlinearly separable input-output relationships. The data selection technique is applied before any supervised training is undertaken, reducing the number of speculative models to be entertained, and serves to guide the developer through the crucial first few steps of model development