Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors

TL;DR

This paper introduces denoising diffusion variational inference (DDVI), a novel black-box inference method using diffusion models as flexible posteriors, improving latent variable model inference and learning across benchmarks and a biological task.

Contribution

It presents a new diffusion-based variational posterior class and a regularized ELBO training method, enhancing inference in deep latent models.

Findings

DDVI outperforms normalizing flows and adversarial posteriors.

It improves inference accuracy on standard benchmarks.

It successfully infers latent ancestry from human genomes.

Abstract

We propose denoising diffusion variational inference (DDVI), a black-box variational inference algorithm for latent variable models which relies on diffusion models as flexible approximate posteriors. Specifically, our method introduces an expressive class of diffusion-based variational posteriors that perform iterative refinement in latent space; we train these posteriors with a novel regularized evidence lower bound (ELBO) on the marginal likelihood inspired by the wake-sleep algorithm. Our method is easy to implement (it fits a regularized extension of the ELBO), is compatible with black-box variational inference, and outperforms alternative classes of approximate posteriors based on normalizing flows or adversarial networks. We find that DDVI improves inference and learning in deep latent variable models across common benchmarks as well as on a motivating task in biology -- inferring latent ancestry from human genomes -- where it outperforms strong baselines on the Thousand Genomes dataset.