Selective learning using sensitivity analysis

Research on improving generalization performance and training time of multilayer feedforward neural networks has concentrated mostly on the optimal setting of initial weights, learning rates and momentum, optimal architectures, and sophisticated optimization techniques. In this paper we present an alternative approach where the network dynamically selects patterns during training. We apply sensitivity analysis to select only patterns closest to the separating hyperplanes. Experimental results of an artificial and two real world classification problems show that our selective learning method significantly reduces the training set size without decreasing generalization performance, i.e., the results presented show that the generalization is improved compared to learning with all training patterns