Composite squared-error algorithm for training feedforward neural networks

Abstract

A new algorithm, the so-called composite squared-error (CSE) algorithm, for training neural networks is presented. The CSE algorithm, whose roots lie in the field of adaptive IIR filtering, is able to avoid suboptimal solutions and associated saddle points, thus achieving lower values of the associated mean-squared-error function in a fewer number of iterations. For that matter, the CSE algorithm can regularly outperform other existing training schemes in most applications where neural networks are employed.