Prof. dr. Paul Bürkner

TU Dortmund University

Amortized Bayesian inference for multilevel models

Abstract: With the advent of probabilistic programming languages, speed remains the only main limiting factor of Bayesian inference. This is because current gold-standard posterior approximators, in particular MCMC, are very slow, especially compared to optimization-based approaches. In the end, it seems we have to pay this price in order to achieve principled and accurate uncertainty quantification. Or do we? What if we could have accurate and fast Bayesian inference at the same time? This question leads us to what we call neural amortized Bayesian inference, a promising new field at the intersection of Bayesian inference and deep learning. I will highlight some or our recent advances as well as existing challenges in the field. This inspires a look into a potential future of Bayesian inference, accelerated by the learning and generalization abilities of neural networks where trustworthiness and speed are no longer conflicting goals.

Bio: Paul Bürkner is a leading statistician and methodologist in the field of Bayesian multilevel modeling. He currently is Professor of Computational Statistics in the Department of Statistics at TU Dortmund University where his research group develops novel probabilistic methods and tools for applied data analysis. He is widely recognized as the author and lead developer of the brms package for the R programming language, which provides a powerful and accessible interface to Bayesian multilevel models implemented in Stan. Through brms, Paul has played a central role in broadening the reach of Bayesian modeling within the social, behavioral, and health sciences. Paul’s scholarly contributions span topics such as Bayesian item response theory, simulation-based inference, and the integration of Bayesian workflow principles into applied research. His work has been published in leading statistical and methodological journals, and he actively contributes to the open-source software community.