Distributed Computation for Marginal Likelihood based Model Choice

We propose a general method for distributed Bayesian model choice, using the marginal likelihood, where data set is split in non-overlapping subsets. These subsets are only accessed locally by individual workers and no shared between workers. approximate evidence full through Monte Carlo sampling from posterior on every subset generating per subset. The results combined novel approach which corrects splitting summary statistics of generated samples. Our divide-and-conquer enables choice large setting, exploiting all available information but limiting communication derive theoretical error bounds that quantify resulting trade-off computational gain loss precision. embarrassingly parallel nature yields important speed-ups when used massive sets as illustrated our real world experiments. In addition, we show how suggested can be extended to within reversible jump setting explores multiple feature combinations one run.