Reproduce “Continual Repeated Annealed Flow Transport Monte Carlo” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Alexander G. D. G. Matthews, Michael Arbel, Danilo J. Rezende, Arnaud Doucet

CRAFT trains flows between distributions in an annealing path and uses them within a sequential Monte Carlo sampler. At each transition, weighted particles estimate a KL-divergence training objective, while importance weighting, resampling, and Markov chain Monte Carlo steps correct imperfect transport. On a lattice φ⁴ test, its particle-MCMC proposal reaches lower susceptibility-estimation error faster than tested alternatives; this is experiment-specific.

Key insights

  • Optimizing local KL divergences with streaming SMC approximations provides a stable and robust way to train normalizing flows within an annealed framework, effectively integrating MCMC and resampling without gradient issues.
  • The 'continual repeated' training approach prevents overfitting and bias by continuously generating fresh samples, leading to more consistent and performant flow learning.
  • CRAFT-based Particle MCMC offers a highly efficient inner loop for high-accuracy, asymptotically unbiased expectation estimates, significantly outperforming traditional MCMC and pure VI in challenging multi-modal scenarios.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.