EM algorithms for independent component analysis

This paper presents a new approach to the blind source separation problem. In our approach, each source density is described by a model that is quite general and fully adaptive. Based on this model, we derive and demonstrate unsupervised learning algorithms not only for square noiseless mixing, but also for the general case where the number of sources may differ from the number of observed mixtures and the data are noisy. These algorithms use expectation-maximization to estimate the arbitrary source densities, mixing matrix and noise covariance from the input data. An approximate algorithm, based on the variational framework, is developed for cases where exact learning is intractable