Implementation of a robust feedforward neural network using the classifier structure

In this paper, we improve the performance of a feedforward neural network (FNN) through eliminating the effect of gross error by using the classifier structure. We first prove that the output of a classifier just prior to winner-take-all (WTA) represents the empirical posteriori probability, f₀(θ_i|x), of each pattern θ_i given input x. We also apply filtering approach based on robust statistics before reconstructing analog outputs. The data corrupted by noise can be rejected in this process. Finally, based on these results, we suggest a new neural network structure, named neurofilter. It consists of 3 stages, which are pattern transform, filtering, and inverse transform. Simulation results shows that the proposed structure yields consistent estimates even in the presence of noise