Improving performance of neural classifiers via selective reduction of target levels ⋆

Reducing the level of the targets corresponding to training samples for a machine classifier using the outputs of an auxiliary classifier is interesting because it allows to save expressive power unnecessarily dedicated to increase the output level of well-classified samples. In this paper we propose an iterative form of this selective reduction of target levels with a simple linear reduction schedule. Extensive simulations show that the proposed method has not only a performance better than or equal to conventional training or using static versions of the reduction, but also with respect to support vector machines (SVM). This potential advantage is accompanied by a smaller size and a design effort not much higher than the corresponding SVM, thus making the proposed method very attractive for practical applications.