Application of the Network Discriminant Function to Learning

Abstract : This introduces methods of applying the Network Discriminant Function (NDF) to training and automatic architecture selection of feed-forward multiple layer perceptrons (MLPs) used for Pattern Recognition (PR). Knowledge of the NDF may be used to impact MLP architecture choices. The NDF gives a new insight into MLP performance in that it reflects how well the MLP is clustering the data. If clustering is an intrinsic capability of the MLP which makes it perform well, then the NDF should correlate with other performance measures such as the Sum Squared Error. Architecture choices and training technique need to be selected as to aid the clustering ability of the MLP. The major goals of the research are to establish the NDF as a useful criteria and to compare it to other criteria such as Cascade-Correlation's error correlation. The NDF-Cascade (NDFC) learning architecture and the MLP NDF are proposed, tested and evaluated.