Abstract
Continuous normalizing flows (CNFs) learn the probability path between areference distribution and a target distribution by modeling the vector fieldgenerating said path using neural networks. Recently, Lipman et al. (2022)introduced a simple and inexpensive method for training CNFs in generativemodeling, termed flow matching (FM). In this paper, we repurpose this methodfor probabilistic inference by incorporating Markovian sampling methods inevaluating the FM objective, and using the learned CNF to improve Monte Carlosampling. Specifically, we propose an adaptive Markov chain Monte Carlo (MCMC)algorithm, which combines a local Markov transition kernel with a non-local,flow-informed transition kernel, defined using a CNF. This CNF is adaptedon-the-fly using samples from the Markov chain, which are used to specify theprobability path for the FM objective. Our method also includes an adaptivetempering mechanism that allows the discovery of multiple modes in the targetdistribution. Under mild assumptions, we establish convergence of our method toa local optimum of the FM objective. We then benchmark our approach on severalsynthetic and real-world examples, achieving similar performance to otherstate-of-the-art methods, but often at a significantly lower computationalcost.