Thomas Brox: Building world models: what, why, and how

Abstract

We have seen much progress in recent years, which culminated in so-called foundation models and large language models. Long-standing problems in pattern recognition and language understanding can be considered solved from the foundational research perspective. This leads to the question what is missing? Is it the comeback of classical symbolic AI, now that we know how to derive symbols from the raw data?

I will rather argue that it is dynamics, causality, curiosity, and continual learning that are missing. So-called world models that learn both the state space and the dynamics from time-predictive tasks and allow for neural planning are a promising direction to fill these missing parts. I will explain what is meant by a world model, why it is decisively different from foundation models, what makes it difficult to learn them for complex environments, and show some recent results from a larger consortium effort where we learn a driving world model from offline driving data.

When

Tuesday, July 14, 2025, 2:00 PM

Where:

IMBIT NEXUS Lab, Georges-Köhler-Allee 201, 79110 Freiburg

About

Thomas Brox

Host

Daniele Cattaneo

Program

https://www.brainlinks-braintools.uni-freiburg.de/events/e-id/2025/07/14/event/ical/673/