CoSy seminar: "Training-free guidance of generative AI models as Bayesian inference"

Everyone is welcome and the first 40 people to register will be treated to a free lunch sandwich. If you do not want lunch, you are still welcome to join.

Register for a lunch sandwich (before Sunday 9 November).

Abstract: Diffusion and flow-based models have emerged as a powerful tool in generative AI, enabling the creation of high-quality images and audio, but also other forms of data appearing in the natural sciences, with applications from materials discovery to weather forecasting. These models can be “controlled” to simulate from desired conditional distributions, for instance to guide generated molecules towards desired properties in the context of drug discovery, or to perform data assimilation by conditioning a forecasting model on observations. However, standard guidance techniques require paired data, expensive training or fine-tuning, and is limited to specific types of conditioning. Recent advancements have introduced training-free guidance techniques that allow for more flexible and efficient control over the generation process. This is based on interpreting the conditional generation as a Bayesian inference problem, where the generative model acts as a prior and the conditioning is incorporated via a likelihood term. This talk will explore the fundamentals of diffusion and flow-based generative models and introduce some of our recently developed Bayesian inference methods for training-free guidance of such models.

This is a lecture in the seminar series held by CIM (Centre for Interdisciplinary Mathematics).