Wieland Brendel: Robustness and Generalisation in the Era of Web-Scale Data

Abstract

Understanding the robustness and generalisation capabilities of ImageNet-era models provided crucial insights into their decision-making, limitations, and internal processes. However, today's foundation models, trained on massive and diverse web-scale datasets, have largely escaped such systematic analysis. This talk presents methods to adapt and apply rigorous robustness testing to these foundation models, revealing important findings about their internal mechanisms and actual resilience to distribution shifts. I then examine theoretical frameworks for specific generalisation types—particularly compositionality and rule extraction—and sketch how these principles could guide development of next-generation foundation models with more reliable and human-aligned reasoning abilities. This work bridges empirical testing with theoretical understanding to improve how we evaluate and develop advanced machine learning systems.

Discussion

The lecture will be followed by informal discussions with light snacks and drinks.

Where:

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

More details at:

https://www.brainlinks-braintools.uni-freiburg.de/events/e-id/2025/05/21/event/ical/661/

Host

Thomas Brox