next up previous
Next: Bibliography Up: Fast Algorithms for Bayesian Previous: Experiments


First we considered the EM-algorithm for finding independent components with low noise. The problem of slow convergence was noted and an improvement was proposed. When finding the sources one at a time, the contributions of the unwanted sources was treated as noise, which leads to faster convergence. Although the approach was found to be implicitly the same as in the FastICA algorithm, it is valid for other situations too. In Bayesian ICA for i.i.d. sources, the modification can be applied as proposed. Other possibilities include finding groups of components that are not mutually independent but are independent related to other components not in the group. The independent components are then projections to multidimensional subspaces instead of one-dimensional projections. This has been proposed e.g. in [3,4]. The modification proposed in this paper applies to this case too, since the contributions of sources not in the group can be regarded as approximately Gaussian noise.

Further work includes finding more general principles, where the modification could be derived for other cases, such as time-dependent sources or nonlinear ICA.

Harri Lappalainen