Denoising Diffusion Variational Inference
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: visualization, vae, diffusion models, representation learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Latent variable methods are a powerful tool for representation learning that greatly benefit from expressive variational posteriors, including generative models based on normalizing flows or adversarial networks.
In this work, we propose denoising diffusion variational inference, which relies on diffusion models---recent generative algorithms with state-of-the-art sample quality---to fit a complex posterior by performing diffusion in latent space. Our method augments a variational posterior with auxiliary latent variables via a user-specified noising process that transforms a complex latent into a simple auxiliary latent. The approximate posterior then reverses this noising process by optimizing a lower bound on the marginal likelihood inspired by the wake-sleep algorithm. Our method can be used to fit deep latent variable models, which yields the DiffVAE algorithm. This algorithm is especially effective at dimensionality reduction and representation learning, where it outperforms methods based on adversarial training or invertible flow-based posteriors. We use this algorithm on a motivating task in biology---inferring latent ancestry from human genomes---and show that it outperforms strong baselines on the 1000 Genomes dataset.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7583
Loading