An alternative view of variational Bayes and asymptotic approximations of free energy (Q420913)

The paper addresses two approximation schemes for Bayesian learning, i.e. it views the variational Bayes approach for statistical models with latent variables as an application of the local variational approximation method. Furthermore, by combining the presented view with the asymptotic theory of Bayesian estimation, a relationship between the generalization performance of the variational Bayes technique and the minimum variational free energy is obtained; a numerical experiment additionally supports this analysis. The study also relates to extensive previously published work of the author, which additionally confirms it as a complete and interesting reading material for those researching in the theoretical background of Bayesian learning.