Abstract: Approximate Bayesian inference estimates descriptors of an intractable target distributions. Markov Chain Monte Carlo (MCMC) and variational inference (VI) are two conventional approaches of Bayesian inference, but MCMC requires long simulation to extract large amounts of random samples, and VI infers from essentially incorrect, simpler distributions. We propose a particle-based deterministic scheme that transports a set of particles towards the target distribution, while emulating the behavior of MCMC samplers. Specifically, we first find an ordinary differential equation (ODE) whose evolution of marginal densities corresponds to a diffusion-based MCMC dynamics, and then project the drift term of the ODE to a reproducing kernel Hilbert space (RKHS). Such dynamics discretize into generalized SVGD (GSVGD), a Stein-type deterministic particle sampler, with particle updates coinciding with applying the diffusion Stein operator to a kernel function. We demonstrate empirically that GSVGD can de-randomize complex MCMC dynamics, which combine the advantages of auxiliary momentum variables and Riemannian structure, while maintaining the high sample quality from an interacting particle system.

Speakers:  Zheyang Shen

Zheyang Shen is a doctoral student from Probabilistic Machine Learning group, Dept. of computer science, Aalto University.

Affiliation: Aalto University

Place of Seminar:  Zoom
Meeting ID: 694 4226 1666
Passcode: 675107