Robust Parallel Fast-ICA Algorithms Using Batch and Adaptive MMSE Estimators

FRANCISCO MESSINA; BRUNO CERNUSCHI FRÍAS

BUENOS AIRES

Congreso; 41 JAIIO, Simposio Argentino de Tecnología (AST 2012); 2012

SADIO

All the algorithms for ICA require high-order statistics to estimatethe independent components. This is because second-order informationis insufficient to assess that two random variables are independentof each other. It is known that the robustness of the high-order sampleestimators is poor, meaning that a few outliers can change dramaticallyits value. In this paper, we generalize the alternative robust statisticsfor moments and cumulants introduced by Welling [1] presenting theMMSE-robust moments. Then we present a batch and adaptive versionsof an algorithm for estimating the parameters that define the estimator.Finally, we modify two FastICA algorithms of ICA based on kurtosis andnegentropy to apply the MMSE robust estimators and show some experimentswith supergaussian sources to demonstrate the improvement.