We propose denoising diffusion variational inference (DDVI), an approximate inference algorithm for latent variable models which relies on diffusion models as flexible variational posteriors. Specifically, our method introduces an expressive class of approximate posteriors with auxiliary latent variables that perform diffusion in latent space by reversing a user-specified noising process. We fit these models by optimizing a lower bound 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. It increases the expressivity of flow-based methods via non-invertible deep recurrent architectures and avoids the instability of adversarial methods. We use DDVI on a motivating task in biology -- inferring latent ancestry from human genomes -- and we find that it outperforms strong baselines on the Thousand Genomes dataset.
翻译:暂无翻译