M-Estimators based activation functions for robust neural network learning

Multi-layer feed-forward neural networks has been proven to be very successful in many applications, as industrial modeling, classification and function approximations. Training data containing outliers are often a problem for these supervised neural networks learning methods that may not always come up with acceptable performance. Robust neural network learning algorithms are often applied to deal with the problem of gross errors and outliers. Recently many researches exploited M-estimators as performance function in order to robustify the NN learning process in the presence of outliers (contaminated data). For first time we propose in our paper to present M-Estimators based activation functions (M-estimators T.Fs) to replace the traditional activation functions (conventional T.Fs).In order to improve the learning process, and hence the robustness of neural networks in presence of outliers. Comparative study between M-estimators T.Fs and conventional T.Fs was established in paper using function approximation problem.