How to learn from complex data II: simulation-based inference in Physics and Life Sciences
A hands-on workshop
when: Sep 9-10, 2025 (09:15-16:00)where: Faculty of Physics, Friedrich-Hund-Platz 1, 37077 Göttingen, room: C 00.106 (PC pool)
available seats: 20
lecturers: Prof. Dr. Michael Wibral (CIDBN, Dep. Data-driven Analysis of Biological Networks), Dr. Richard Gao (Wilhelm Schickard Institute of Computer Science WSI, University of Tübingen)
organizers: Dr. Bernhard Bandow (GWDG / CIDBN), Prof. Dr. Fred Wolf (CIDBN, Dep. Physics of Biological Systems)
prerequisites for participation: familiarity with linux and python, basic knowledge of propability and statistics
registration: via e-mail to cidbn@uni-goettingen.de by Sep 2, 2025.
(Please state your full name and affiliation when registering. Thanks!)
day 1
lecturer: Michael Wibral
On the first day we will introduce the fundamentals of Bayesian inference that are necessary to soundly use it and interpret its results - most importantly, well introduce the Bayesian interpretation of probabilities as plausibilties of statements about the world, the differences between parameter inference, model inference and out of sample prediction. We'll learn how to use the widely used pymc software to employ samplers for solving practical real-world Bayesian inferences problems.
day 2
lecturer: Richard Gao
abstract:
On the second day, we will cover simulation-based inference (sbi) using neural density estimators, and work through examples applying sbi to infer parameters of spiking neural network models. Theoretical components include a review of maximum likelihood, KL divergence, and (conditional) density estimation, while introducing inference as an optimization problem (i.e., variational inference). We will then cover how automatic differentiation packages (e.g., pytorch) can be applied to optimize distributions parameterized by deep neural networks (e.g., mixture density networks,
normalizing flows). Finally, we will demonstrate how these methods can be applied to spiking neural network models with brain recordings.
Participants are encouraged to bring their own data and/or simulator models.