Presenting our insights on diffusion-based sampling at MIT

Last week, Lorenz Richter had the opportunity to visit the Center for Computational Science & Engineering (CCSE) at Massachusetts Institute of Technology to discuss our latest research on diffusion-based sampling.

The group, led by Youssef Marzouk, is at the forefront of developing rigorous methods for uncertainty quantification, inverse problems, and statistical machine learning - areas that align closely with our work at dida.

The presentation and discussion centered around how we can use controlled diffusion processes to sample from complex target distributions.

• Bridging frameworks: Discussing the variational formulations that allow us to treat diffusion-based sampling as a stochastic optimal control problem.

• Efficiency for high dimensions: How our recent work can improve convergence, shown for scientific applications.

• Scientific ML: Exploring the synergy between generative AI and classical computational physics to solve inverse problems more robustly.

It was a fantastic experience to discuss the mathematical foundations of AI with such a talented group of researchers. The exchange of ideas between the European AI landscape and the technical ecosystems here in Cambridge continues to be a great source of inspiration for our own projects.

A big thank you to Youssef Marzouk for the invitation and to Joanna Zou for the great organization and warm welcome!

We're looking forward to continuing the conversation.