SIMULTANEOUS MI­BASED ESTIMATION OF INDEPENDENT COMPONENTS AND OF THEIR DISTRIBUTIONS

Lu'is B. Almeida
luis.almeida@inesc.pt

In ICA and BSS methods, the mutual information (MI) of the estimated components is one of the most desirable measures of statistical dependence, for use as an objective function. However, its use requires the estimation of the statistical distributions of the components. Previous MI­ based ICA methods have resorted to an a­priori knowledge of those distributions or to their approximation by trun­ cated series expansions or by means of kernels. This pa­ per presents a method for simultaneously estimating those distributions and performing MI­based ICA, using a sin­ gle network trained with a single objective function. The method is able to correctly handle mixtures of subgaussian and supergaussian sources.