Jonas Arruda (University of Bonn, Germany): Diffusion Models in Simulation-Based Inference: Post-Hoc Guidance and Composition
Time and place:
June 4th, 14–15
ROOM TBA (Exactum), University of Helsinki
Abstract: Diffusion models have recently emerged as a powerful framework for simulation-based inference (SBI), enabling flexible estimation of complex posterior distributions via denoised score-based learning. While SBI methods perform well in moderate-dimensional settings, their application to large-scale, structured datasets remains limited by the cost of generating sufficient simulations across high-dimensional and hierarchical data regimes.
This talk reviews recent advances in diffusion-based SBI and presents a new approach for compositional inference applied to hierarchical Bayesian models. The central idea is to exploit model structure by learning local inference components that can be composed across levels of the hierarchy, rather than amortizing over entire datasets end-to-end. This avoids the need to repeatedly simulate full datasets and instead enables reuse of learned components across different groups and levels within a hierarchical dataset.
This approach enables efficient and accurate posterior inference in settings where traditional amortized methods break down, particularly when data are large. These results suggest that diffusion models provide a viable path toward scalable amortized inference for complex Bayesian models, bridging a gap between SBI methods and large-scale applications in science and machine learning.
Bio: Jonas Arruda has studied Mathematics at the University of Bonn since 2017 and is soon to finish his PhD in Mathematics, which he started in 2022. He is supervised by Prof. Jan Hasenauer at the Bonn Center of Mathematical Life Sciences. His research focuses on simulation-based inference and inverse problems with applications in computational life sciences. He has published his work at ICML, ICLR, and npj Systems Biology and Applications.