TRUNCATED EXPONENTIALNONLINEARITIESFOR INDEPENDENTCOMPONENT ANALYSIS

Thispaperproposes exponential typenonlinearities inorder to blindly separate instantaneous mixtures ofsignals withmixedkurtosis signs. These nonlinear functions areapplied only inacertain rangearound zeroinorder toensure that therelative gradient algorithm remains locally stable. Theproposed truncated nonlinearities neutralize theeffect ofoutliers while thehigher order terms inherently present intheexponential function result infast convergenceespecially forsignals withbounded support. Byvarying the truncation threshold, signals withbothsub-Gaussian andsuperGaussian probability distributions canbeseparated. Furthermore, whenthesources consist ofsignals withmixedkurtosis signs we propose toestimate thecharacteristic function online inorder to classify thesignals assub-Gaussian orsuper-Gaussian andconsequently choose anadequate value ofthetruncation threshold. Somecomputer simulations arepresented todemonstrate theeffectiveness oftheproposed idea.