Interesting talk by Columbia's Stephan Mandt on large scale Bayesian inference with stochastic variational inf. No parallelization, though.