Thursday, April 30, 2026, 10:30 - 16:00

TBA

Preliminary Schedule:

10:30 Resolving Ill-Conditioning in Scientific Machine Learning (Dr. Marius Zeienhofer, abstract see below)

11:30 Lunch

13:00 Natural policy gradient methods: From geometric foundations towards safe RL (Dr. Johannes Müller, abstract see below)

14:00 Coffee Break

14:30 Grounding Diffusion Models with Optimal Control (Jasper Hoffmann, abstract see below)

 

Abstracts

Resolving Ill-Conditioning in Scientific Machine Learning (Dr. Marius Zeienhofer)
We present an 'optimize-then-project' approach for scientific machine learning that mitigates ill-conditioning arising from differential operators in loss functions. The key idea is to design algorithms at the infinite-dimensional level and subsequently discretize them in the tangent space of the neural network ansatz. We illustrate this approach in the context of physics-informed neural networks and the variational Monte Carlo method for quantum many-body problems, where neural networks have recently emerged as promising ansatz classes. We conclude with considerations related to the scalability of these schemes.

From geometric foundations towards safe RL (Dr. Johannes Müller) 
Natural policy gradient and trust-region methods are central to modern reinforcement learning. In this talk, we present a unified perspective on policy optimization that connects information geometry, regularization, and constrained optimization. We first examine the geometric structure underlying natural policy gradient methods and show that they are induced by the conditional entropy regularization of a linear program. This allows us to study the error induced by entropy regularization in discounted Markov decision processes, yielding optimal convergence rates. Further, this shows that the natural policy gradient method prefer maximum entropy policies thereby characterizing the implicit bias of natural policy gradient methods. Building on this geometric framework, we construct trust-regions for constrained Markov-decision processes. This yields a safety-aware optimization scheme that enforce safety through geometric constraints

Grounding Diffusion Models with Optimal Control (Jasper Hoffmann) 
While diffusion models have achieved remarkable success in robot learning and autonomous driving, their application to complex planning remains an evolving frontier. This talk begins with a tutorial on diffusion through the lens of flow matching, framing it as a powerful, simulation-free alternative to traditional continuous normalizing flows (CNFs). We then explore how modern research ensures these models respect underlying system dynamics and hard safety constraints. Finally, we present a novel framework that integrates model predictive control (MPC) directly into the generative process, enabling more robust and feasible trajectory synthesis.