INDEPENDENT COMPONENT ANALYSIS THROUGH DIRECT ESTIMATION OF THE MUTUAL INFORMATION

Georges A. Darbellay 1 and Petr Tichavsky 2, 2 Institute of Information Theory and Automation, Academy of Sciences
e-mail: georges.darbellay@epfl.ch, e-mail: tichavsk@utia.cas.cz

We introduce an algorithm for the estimation of the mu- tual information between several signals. This algorithm also provides a convenient way of estimating the entropy of scalar signals. The algorithm uses a data-dependent par- titioning of the observation space and has a relatively low computational cost. Matlab code is available. The useful- ness of this approach in solving the independent component analysis problem is demonstrated on the separation of a lin- ear mixture of two and three real-world speech data sets.