Reproduce “Learning Deep Generative Models with Annealed Importance Sampling” on OpenResearch

About this paper

Xinqiang Ding, David J. Freedman

The authors introduce annealed importance sampling for training latent-variable generative models, moving samples from an inference model toward the posterior through intermediate distributions and Hamiltonian Monte Carlo, then weighting them to estimate gradients. In a matched-computation Omniglot comparison, the method reports an approximate test log marginal likelihood of −102.45, versus −103.85 for IWAE-DReG; this result is limited to that setting.

Key insights

• AIS provides a unified framework that generalizes both Importance Weighted Autoencoders (IWAEs) and certain MCMC-based learning algorithms, bridging variational inference and MCMC paradigms.
• The annealing process, involving multiple short MCMC chains, allows for a more effective exploration of complex posterior landscapes and more accurate gradient estimation for a given computational budget.
• The method offers a flexible way to trade computational cost (by varying annealing steps or number of samples) for model accuracy without increasing memory requirements.

Summary by alphaXiv.