Variational Inference for Interacting Particle Systems with Discrete Latent States

Giosue Migliorini, Padhraic Smyth

Keywords: continuous-time markov chains, spatiotemporal models, schroedinger bridge, variational inference

TL;DR: We model spatiotemporal data (images/graphs) with a latent discrete-state continuous-time dynamic, inferring it from noisy observations solving an approximate multi-marginal Schrödinger bridge problem.

Abstract: We present a novel Bayesian learning framework for interacting particle systems with discrete latent states, addressing the challenge of inferring dynamics from partial, noisy observations. Our approach learns a variational posterior path measure by parameterizing the generator of the underlying continuous-time Markov chain. We formulate the problem as a multi-marginal Schrödinger bridge with aligned samples, employing a two-stage learning procedure. Our method incorporates an emission distribution for decoding latent states and uses a scalable variational approximation.

Submission Number: 111