摘要
ThispaperaddressestheproblemofBlindSourceSeparation(BSS)andpresentsanewBSSalgorithmwithaSignal-AdaptiveActivation(SAA)function(SAA-BSS).Bytakingthesumofabsolutevaluesofthenormalizedkurtosesasacontrastfunction,theobtainedsignal-adaptiveactivationfunctionautomaticallysatisfiesthelocalstabilityandrobustnessconditions.TheSAA-BSSexploitsthenaturalgradientlearningontheStiefelmanifold,anditisanequivariantalgorithmwithamoderatecomputationalload.ComputersimulationsshowthattheSAA-BSScanperformblindseparationofmixedsub-Gaussianandsuper-Gaussiansignalsanditworksmoreefficientlythantheexistingalgorithmsinconvergencespeedandrobustnessagainstoutliers.
出版日期
2006年03月13日(中国期刊网平台首次上网日期,不代表论文的发表时间)