Presentation at Microsoft Research on diffusion-based sampling

During his research stay in Cambridge, Massachusetts and following his talk at Massachusetts Institute of Technology, Dr. Lorenz Richter also shared our recent work on diffusion-based sampling with the team at Microsoft Research New England (MSR-NE).

The seminar, "A non-Markovian approach to diffusion-based sampling," focused on the challenges of sampling from high-dimensional, unnormalized densities - a task central to scientific domains where ground-truth data is often unavailable.

In his talk, Lorenz presented insights from our latest research, specifically introducing the Bridge Matching Sampler (BMS). This framework generalizes previous matching objectives as special forms of fixed-point iterations rooted in Nelson’s relation.

The full paper just got published here.

A sincere thank you to Carles Domingo-Enrich and the team at Microsoft Research for the invitation and the technical exchange.

The group at MSR NE focusses especially on generative modeling and the growing connections to image, video or text generation, but also to developments in fine-tuning, alignment, or scientific simulations.

You can watch the recorded talk here.

As always, our insights from the paper are the result of a great collaboration, so thank you also to Denis Blessing, Julius Berner, Egor Malitskiy, Gerhard Neumann.