Probabilistic Inference and Generative Modeling

The Probabilistic Inference and Generative Modeling group focuses on the development of fundamental algorithms for learning and inference in complex probabilistic models. The group’s research is guided by a principled, mathematically sound approach, with a strong emphasis on robustness and scalability to high-dimensional settings.

A central goal of the group is to design algorithms that serve as reliable building blocks for downstream applications, including reinforcement learning and imitation learning. To this end, the group primarily investigates modern generative modeling techniques, with a particular focus on diffusion models and flow-based models.

In addition, the group actively researches trust-region methods, which aim to provide stable and well-behaved update schemes for training generative models. These methods play a key role in improving convergence, stability, and reliability, especially in large-scale or sensitive learning scenarios.

Key Areas

• Sampling methods
• Diffusion and flow models
• Trust-region methods