A new composite ICA algorithm and its application in fMRI data processing

Independent component analysis (ICA) is a new technique in signal processing to extract statistically independent components from an observed multidimensional mixture of data. A composite algorithm of Newton iteration and natural gradient descent (CNN) is presented to implement ICA by maximizing the sum of marginal Negentropies which is equivalent to minimizing the mutual information of independent signals. The CNN algorithm avoids the singularity of Newton iteration. At the same time, it possesses higher convergence speed than the natural gradient algorithm. We specifically apply the algorithm to functional magnetic resonance imaging (fMRI) data, and the results lend validity to the proposed method as providing a reasonable physiological explanation for the fMRI data.