Title: Consistent Diffusion Models and Learning from Corrupted Data with Ambient Diffusion
Speaker: Giannis Daras (UT-Austin)
Time: Oct 20, 2023, 12:30 PM – 1:30 PM CT
Location: ECB M1059 (M floor)
Abstract: In this talk, we will explore recent algorithmic innovations that generalize diffusion models and tackle the limitations of the standard formulation. In the first part of the presentation, I will introduce the problem of diffusion sampling drift that arises from the recursive nature of the generation process and the propagation of errors due to imperfect score-matching. I will show how to mitigate this problem by enforcing a consistency property (NeurIPS 2023) which states that predictions of the model on its own generated data are consistent across time. Empirically our training objective yields state-of-the-art results for conditional and unconditional generation. Theoretically, we show that if the score is learned perfectly on some non-drifted points and if the consistency property is enforced everywhere, then the score is learned accurately everywhere. In the second part of the talk, we will focus on the problem of learning a generative model using only lossy measurements. This problem arises in scientific applications where access to uncorrupted samples is impossible or expensive to acquire. I will present Ambient Diffusion (NeurIPS 2023), the first diffusion-based framework that can learn an unknown distribution using only highly corrupted samples. Ambient Diffusion can be used to finetune a pre-trained foundation model or train a model from scratch.
Bio: Giannis Daras is a fourth-year Ph.D. student in the Computer Science Department of The University of Texas at Austin, supervised by Prof. Alexandros Dimakis. Giannis is also a Research Scientist Intern at NVIDIA. His research is supported by the Bodossaki Fellowship, the Onassis Fellowship, and the Leventis Fellowship. Giannis is doing research on generative modeling. He is interested in learning generative models from lossy measurements and developing algorithms to use generative priors for solving inverse problems.