Boundedness of weight elimination for BP neural networks

Abstract

Weight elimination can be usefully interpreted as an assumption about the prior distribution of the weights trained in the backpropagation neural networks (BPNN). Weight elimination based on different scaling of weight parameters is of a general form, with the weight decay and subset selection methods as special cases. The applications of this method have been well developed, however, only few references provides more comprehensive theoretical analysis. To address this issue, we investigate the uniform boundedness of the trained weights based on a descriptive proof. © 2014 Springer International Publishing.