Folded Transport MCMC: Certifiable Quotient Posterior Computation for Symmetric Bayesian Models
Researchers have developed a new MCMC method called Folded Transport MCMC (FolT-MCMC) to address challenges in Bayesian models with symmetries. This method directly infers on the quotient posterior by using a learned normalizing flow to construct an independence sampler on the fundamental domain of the symmetry group. FolT-MCMC offers significant improvements in convergence diagnostics and certified lower bounds, showing gains of 2x to 145x on various mixture models and real-world data. AI
IMPACT Introduces a novel computational technique for Bayesian inference, potentially improving the efficiency and reliability of models used in AI research.