Bayesian Learning via Stochastic Gradient Langevin Dynamics