Optimized feature extraction and the Bayes decision in feed-forward classifier networks

摘要: The problem of multiclass pattern classification using adaptive layered networks is addressed. A special class networks, i.e., feed-forward with a linear final layer, that perform generalized discriminant analysis discussed, This sufficiently generic to encompass the behavior arbitrary nonlinear networks. Training network consists least-square approach which combines inverse computation solve for layer weights, together optimization scheme parameters nonlinearities. general analytic form feature extraction criterion derived, and it interpreted specific forms target coding error weighting. An important aspect exhibit how priori information regarding nonuniform membership, uneven distribution between train test sets, misclassification costs may be exploited in regularized manner training phase >