Variational Bayes for signal processing

Abstract

Bayesian inference provides a means to design optimal filters that deal with uncertainty in a principled way. But exact Bayesian inference cannot be applied to all models because integrals for normalization might not have analytic solutions. In that case, variational Bayes provides a fast method to obtain approximate posterior distributions. It can be applied to simultaneously estimate states, parameters, noise and other unknown variables in dynamical systems. As such, it is a prime method for designing signal processing systems.