REGULARIZED SECOND ORDER SOURCE SEPARATION

I. Schießl 1 ,H. Sch˜oner 1 , M. Stetter 1 , A. Dima 1 and K. Obermayer 1
D­10587 Berlin, Germany; email: ingos@cs.tu­berlin.de

The task of separating signals from experimentally mea­ sured linear mixtures is often complicated by the presence of noise sensor noise and statistical dependencies between the original sources, which often makes standard indepen­ dent component analysis (ICA) algorithms fail [1, 2]. One way to overcome these problems is to introduce additional knowledge we have about the mixing process and the sig­ nals themselves. Here we suggest to add a regularization term to the cost function of multishift extended spatial decorrelation (multi­ shift ESD, [2]) which contains prior information about the time­course of one or more original sources. Using an arti­ ficial toy dataset and a dataset that contains prototype sig­ nals obtained from optical recording of brain activity we show that the regularization term improves the separation results at different noise levels.