Diffusion and Flow-Based Models for Inverse Problems by Professor Yee Whye Teh

Abstract

Across diverse scientific domains, scientists face the same fundamental problem: they must infer unknown images or physical states from incomplete, noisy, or indirect measurements/constraints. In such inverse problems, measurements alone are insufficient to determine a unique answer, so reliable inference must draw on prior scientific knowledge about realistic solutions and report calibrated uncertainties over multiple solutions consistent with both prior knowledge and measurements. This can be framed as Bayesian inference, where generative AI methods based on diffusions and flows represent complex scientific priors, and steering based methods aim to direct the generative process towards parts of the state space that are consistent with measurements.

In this talk, Professor Yee Whye Teh presents two pieces of work in this area. Meta flow-maps (arxiv:2601.14430) use stochastic flow maps to construct high-quality steering signals, improving reconstruction performance. Exact posterior scores (arxiv:2606.17048) demonstrate that for linear Gaussian measurements, the exact steering signal can be derived and learned via fine-tuning to produce state-of-the-art inverse problem solvers. These works are led by students Peter Potaptchik, Adhi Saravanan, and Abbas Mammadov.

About the Speaker

Professor Yee Whye Teh is a Professor of Statistical Machine Learning in the Department of Statistics at the University of Oxford and a Research Director at DeepMind working on AI research. He is also an AI Visiting Professor in Singapore, visiting NTU and NUS. He obtained his Ph.D. from the University of Toronto and completed postdoctoral research at the University of California, Berkeley and the National University of Singapore as a Lee Kuan Yew Postdoctoral Fellow.

From January 2007 to August 2012, he was a Lecturer and later a Reader at the Gatsby Computational Neuroscience Unit, UCL. His research interests span machine learning and artificial intelligence, particularly probabilistic methods, Bayesian nonparametrics, and deep learning. He develops novel models and efficient algorithms for inference and learning. He has delivered the Breiman Lecture at NeurIPS 2017, the IMS Medallion Lecture at JSM 2019, and keynote talks at MLSP 2017, KDD 2018, and UAI 2019.