Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors

2024-01-05

AI Summary

This paper addresses the challenge of posterior inference in deep latent variable models—particularly in biological applications such as human genomic ancestry inference—by proposing Diffusion-based Deep Variational Inference (DDVI), a black-box variational inference framework grounded in diffusion processes. DDVI directly models a learnable diffusion process as an iterative refinement of the variational posterior in latent space, marking the first use of diffusion models as highly flexible and expressive posterior approximators. To balance inference fidelity and training stability, DDVI introduces a wake-sleep-inspired regularized ELBO objective. Evaluated on standard benchmarks and the 1000 Genomes Project dataset for ancestry inference, DDVI significantly outperforms strong baselines—including normalizing flows and adversarial posteriors—in both inference accuracy and model learning efficacy.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

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.

Problem

• Develops diffusion-based variational posteriors for latent variable models
• Introduces regularized ELBO for training expressive approximate posteriors
• Improves inference in deep latent variable models and biological tasks

Innovation

• Uses diffusion models as variational posteriors
• Trains with regularized ELBO inspired wake-sleep
• Outperforms normalizing flows adversarial networks