Ensemble Learning for blind source separation
James W. Miskin, David Mackay · Cambridge University Press eBooks · 2001
We show how an Ensemble Learning version of Independent Component Analysis (EL{ICA) can be derived by approximating the true posterior distribution over the model parameters by an approximate distribution. We further extend this algorithm to include sources which can only be positive. We show that the blind deconvolution problem is similar to the blind separation problem and derive an algorithm for blind deconvolution. 1.1 Ensemble Learning In many problems we aim to infer a set of model parameters, , from a set of data, D. In the Bayesian framework this can be done by considering the posterior probability of the parameters P ( jD; H ) = P (D j; H ) P ( jH ) P (D jH ) : (1.1) Commonly the parameters are inferred by maximising the likelihood, P (D j; H ), (ML methods) or the posterior probability (MAP methods) with respect to the model parameters. These methods can over{t. The model parameters that are obtained can be too specic. Instead of nding the MAP estimate of the p...