Incorporating prior knowledge in ICA

Abstract

We propose the modification of the cost function that is optimised by the independent component analysis (ICA) algorithm, so that prior knowledge about some statistical properties of one of the components is incorporated in the optimisation process, for this component to be identified first. Once this component is removed from the recordings, the remaining components may be identified in the usual way. We demonstrate our idea using simulated data, for which the true mixing coefficients are known. We show that the incorporation of the constraint broadens the region around the true mixing values over which, if we start the algorithm, it will converge to the desired solution.