 |
Center for Nonlinear Studies |
I will present a complete framework for constructing stochastic gradient MCMC sampling algorithms. The idea is motivated by introducing irreversible Markov processes into stochastic gradient sampling algorithms to both increase mixing rate and decrease calculation burden. A small bias is introduced as a trade off for decrease of variance. When the bias is not tolerable, I have recently discovered a method to correct for this bias.
Host: Michael Chertkov