BLISS PROJECT Softwares for Blind Source Separation (Instantaneous Mixtures)
 This page contains some softwares for blind separation of instantaneous mixture of sources. They are presented in the form a demonstration which is avalable in two versions: a matlab and a java version. The matlab code can be downloaded (in the form tar gziped) here . Launch the demonstration by typing bliss. The java source and classes can be found in this subdirectory No Java demonstration since your brower doesn't support java WARNING The java demo. have been developed in the Linux environement (java version 1.4.2). It may not work correctly under Windows The demonstration covers only the case of two mixtures of two sources for simplicity, but the subroutines which drive the calculation accept any number of sources and the same number of mixtures. There are subroutines for 3 separation methods and another one for separating post nonlinear mixtures: The first method is based on the minimization of the Marginal Mutual Information criterion and exploits the non Gaussianity of the sources. The criterion is expressed in terms of the entropies which are estimated through a kernel method. References: The method is described in Blind Separation of Instantaneous Mixture of Sources via an Independent Component Analysis Fast algorithms for Mutual Information Based Independent Component Analysis Fast Algorithm for Estimating Mutual Information, Entropies and Score Functions Software: The method is implemented in the java class icainf and the matlab function of the same name. The second method is based on the Gaussian Mutual Information, which exploits both the non stationarity (time diversity) and the coloration (spectral diversity) of the sources. It is implementated via a time-frequency analysis and a joint diagonalisation algorithm. Reference: The method is described in Exploiting Source Non Stationary and Coloration in Blind Source Separation and (in the stationary case) in Blind Separation of Instantaneous Mixture of Sources via the Gaussian Mutual Information Criterion Further, the paper Joint Approximate Diagonalization of Positive Definite Matrices describes the joint approximate diagonalisation algorithm. Software: This method is implemented in the java class sepagaus and in the matlab function of the same name. There is also a matlab function jadiag, for performing joint diagonalisation. The third method is also based on the minimisation of the Marginal Mutual Information criterion, but exploits the non stationarity of the sources as well. Reference: The method is described in paper Blind Separation of Non Stationary Non Gaussian Sources. Software: this method is implmented in the java class icansng and the matlab function of the same name The ponstnonlinear algorithm is based on the minimisation of the Marginal Mutual Information criterion. Reference: The method is described in Blind source separation in postnonlinear mixtures Software: this method is implmented in the java class icainfpnl and the matlab function of the same name