Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors
This provides a more expressive variational inference method for researchers working with complex latent variable models, though it appears incremental as an adaptation of diffusion models to variational inference.
The authors tackled the problem of approximate inference in latent variable models by proposing denoising diffusion variational inference (DDVI), which uses diffusion models as flexible variational posteriors. The result was improved inference and learning performance over normalizing flows and adversarial networks on benchmarks and a biology task, where it outperformed baselines on the Thousand Genomes dataset.
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.