Diffusion-based Bayesian Experimental Design

Add to your list(s) Download to your calendar using vCal

• Jacopo Iollo (INRIA)
• Wednesday 25 June 2025, 09:15-10:15
• Seminar Room 1, Newton Institute.

If you have a question about this talk, please contact nobody.

RCLW03 - Accelerating statistical inference and experimental design with machine learning

Bayesian Experimental Design (BED) is a powerful tool to reduce the cost of running a sequence of experiments. When based on the Expected Information Gain (EIG), design optimization corresponds to the maximization of some intractable expected contrast between prior and posterior distributions. Scaling this maximization to high-dimensional and complex settings where no closed form prior is available has been an issue due to BED inherent computational complexity.  In this work, we introduce a pooled posterior distribution with cost-effective sampling properties and provide a tractable access to the EIG maximization via a new gradient expression. Diffusion-based samplers are used to compute the dynamics of the pooled posterior, and ideas from bi-level optimization are leveraged to derive an efficient joint sampling-optimization loop. The resulting efficiency leverage the well-tested generative capabilities of diffusion models to BED to scenarios that were previously impractical.  Numerical experiments and comparison with state-of-the-art methods show the potential of the approach. As a practical application, we showcase how our method accelerates Magnetic Resonance Imaging (MRI) acquisition times while preserving image quality.  This presentation will also detail how Diffuse, a new modulable Python package for diffusion models, facilitates composability and research in diffusion models through its simple and intuitive API , allowing researchers to easily integrate and experiment with various model components. Presentation based on: Bayesian Experimental Design via Contrastive Diffusions. Iollo, J., Heinkelé, C., Alliez, P.,  Forbes, F. (2025). International Conference on Learning Representations

This talk is part of the Isaac Newton Institute Seminar Series series.

This talk is included in these lists:

• All CMS events
• Featured lists
• INI info aggregator
• Isaac Newton Institute Seminar Series
• School of Physical Sciences
• Seminar Room 1, Newton Institute
• bld31

Note that ex-directory lists are not shown.