A Fast Learning Strategy Using Pattern Selection for Feedforward Neural Networks

Intelligent pattern selection is an active learning strat- egy where the classifiers select during training the most informative patterns. This paper investigates such a strat - egy where the informativeness of a pattern is measured as the approximation error produced by the classifier. The algorithm builds the training corpus starting from a small randomly chosen initial dataset and new patterns are added to the learning corpus based on their error sen- sitivity. The training dataset expansion is based on the selection of the most erroneous patterns. Our experimen- tal results on MNIST1 separated digit dataset show that only 3.26% of training data are sufficient for training pur- pose without decreasing the performance (98.36%) of the resulting neural classifier.