Understanding MCMC Dynamics as Flows on the Wasserstein Space

About

It is known that the Langevin dynamics used in MCMC is the gradient flow of the KL divergence on the Wasserstein space, which helps convergence analysis and inspires recent particle-based variational inference methods (ParVIs). But no more MCMC dynamics is understood in this way. In this work, by developing novel concepts, we propose a theoretical framework that recognizes a general MCMC dynamics as the fiber-gradient Hamiltonian flow on the Wasserstein space of a fiber-Riemannian Poisson manifold. The "conservation + convergence" structure of the flow gives a clear picture on the behavior of general MCMC dynamics. The framework also enables ParVI simulation of MCMC dynamics, which enriches the ParVI family with more efficient dynamics, and also adapts ParVI advantages to MCMCs. We develop two ParVI methods for a particular MCMC dynamics and demonstrate the benefits in experiments.

Chang Liu, Jingwei Zhuo, Jun Zhu · 2019

Related benchmarks

Task	Dataset	Result
Bayesian Neural Network Inference	Boston UCI (test)	Test Log-Likelihood-2.42	8
Bayesian Neural Network Inference	Concrete UCI (test)	Test Log-Likelihood-2.95	8
Bayesian Neural Network Inference	Yacht UCI (test)	Test Log-Likelihood-0.73	8
Bayesian Neural Network Inference	Power Plant (UCI) (test)	Test Log-Likelihood-2.77	8
Bayesian Neural Network Inference	Energy UCI (test)	Test Log-Likelihood-1.36	8
Bayesian Neural Network Inference	Kin8nm UCI (test)	Test Log-Likelihood1.23	8
Bayesian Neural Network Inference	YearPredictionMSD UCI (test)	Test Log-Likelihood-3.6	8
Bayesian Neural Network Inference	Protein UCI (test)	Test Log-Likelihood-3.73	8