Variational Bayesian learning for parameter estimation of mixture of Gaussians

Non-Gaussian signals or systems are usually modeled by mixture of Gaussians (MoG) models containing hidden variables. A variational Bayesian learning algorithm was suggested to infer the parameters of MoG. The algorithm estimateed the parameters of MoG by iteratively maximizing the lower bound of the marginal likelihood and updating the variational parameters until the free-form distribution was sufficiently close to the true posterior. The detailed learning of variational Bayes for MoG was derived and explained. The experiments show that this method can estimate the parameters of MoG favorably with sampling method from the engineering view.