Discriminative least squares learning for fast adaptive neural equalization

In this paper a novel approach to learning in Recurrent Neural Networks (RNN) is introduced and applied to the problem of digital adaptive equalization. The proposed method extends to RNN a technique applied with success to feedforward NN and is based on the descent of the error functional in the space of the linear combinations of the neurons (the neuron space); it exploits the principle of discriminative learning, based on the minimization of an error functional which is a direct measure of the classification error considered in equalization problems. Main features of the new approach are higher speed of convergence and better numerical conditioning w.r.t. gradient-based approaches, while numerical stability is assured by the use of robust Least Squares solvers. Preliminary experiments regarding the equalization of PAM signals in different transmission channels are described, which demonstrate the effectiveness of the proposed approach.