Name: Neural Foundation Models for Causal Understanding, what Bayes Can Teach Us
Start: 2026-01-30T15:30:00+00:00
End: 2026-01-30T17:00:00+00:00
Location: Usher Building, The University of Edinburgh

Neural Foundation Models for Causal Understanding, what Bayes Can Teach Us

Dive into how neural foundation models and Bayesian thinking can up your game in causal understanding—live and in person!
About this Event

Neural Foundation Models for Causal Understanding

Join us in person for a fascinating dive into how neural foundation models can revolutionize our grasp of causal understanding.

Understanding causal relationships is crucial for understanding on how to intervene on a system in order to get a desired effect (e.g. treatments that cure patients, or what to knock out in a gene regulatory network). Determining these causal relationships is painful in practice, since causation can only be determined from “interventional data” obtained from costly controlled experiments (e.g. RCTs). On the other hand, passive “observational data” of a system are much cheaper but can only determine correlations, and it is well-known that causation cannot be determined from correlation. Nevertheless, there is a significant body of work that aims to determine as much as possible about causation from observational data, by relying on additional assumptions about the causal mechanism. In this talk, I will discuss our work that has shown that Bayesian inference is well suited to express these assumptions, and that applying what Bayes prescribes directly leads to methods that perform very well. These principles also prescribe how large neural network-based “foundation models” should be built to answer flexible causal questions for a wide range of datasets, and I will discuss some of our work to implement this.

Bio

Mark van der Wilk is an Associate Professor in the Department of Computer Science at the University of Oxford and a Tutorial Fellow at Hertford College. His research focuses on machine learning methods that enable robust generalisation, efficient model architectures, and safe, data-efficient interaction with real-world environments. His work spans Bayesian inference, equivariance, causality, meta-learning, and Gaussian processes, with applications ranging from small-data statistics to large-scale neural networks. He has published at leading venues including NeurIPS and ICML, receiving a best paper award. Prior to Oxford, he was a Senior Lecturer at Imperial College London and completed his PhD at the University of Cambridge under Prof. Carl Rasmussen.