Training multilayer perceptron with multiple classifier systems

Abstract

An idea of training multilayer perceptron (MLP) with multiple classifier systems is introduced in this paper. Instead of crisp class membership i.e. {0, 1}, the desired output of training samples is assigned by multiple classifier systems. Trained with these samples, the network is more reliable and processes better outlier rejection ability. The effectiveness of this idea is confirmed by a series of experiments based on bank check handwritten numeral recognition. Experimental results show that for some recognition applications where high reliability is needed, MLP trained with multiple classifier systems label samples is more qualified than that of MLP trained with crisp label samples. © Springer-Verlag 2004.