0:00 / 0:00
event
Keynote Michael Bronstein ICAI Day October 27, 2021
ICAI Innovation Center for Artificial Intelligence15 June 2026Watch on YouTube
Part of series
Ep. 5 · 2022 Day Icai
View the seriesWhat you'll learn
- Symmetry considerations can derive the most influential deep learning architectures, from CNNs to transformers, from a single principle, modelled on Felix Klein's Erlanger programme.
- The Weisfeiler-Leman test from the 1960s, developed for molecule classification, precisely bounds the expressive power of modern graph neural networks.
- Graph methods see concrete use in drug development, from predicting water solubility to designing antibiotic-like compounds and biological drugs.
- Classical anonymisation falls short for interaction data; the structure of a network alone suffices to re-identify individuals.
- Bronstein shows how graph studies can decouple the computational architecture from the input graph, improving performance and expressive power.
Frequently asked questions
What is geometric deep learning in one sentence?
It is the approach that derives deep learning architectures from the symmetry of the data domain, by analogy with the Erlanger programme that defined geometry as the study of symmetry and invariance, as Bronstein argued at ICAI Day.
Why does the Weisfeiler-Leman test matter for graph neural networks?
This colour refinement method from the 1960s, developed to classify molecules, turns out to bound exactly the expressive power of graph neural networks. Researchers therefore know which tasks such a network can and cannot solve.
What does Bronstein mean by the '5G' of geometric deep learning?
It is his shorthand for the five domains to which the symmetry principles apply: groups, graphs, geodesics and gauges. During the keynote he focuses mainly on graphs.
What risks does Bronstein see in graph methods?
He warns that classical anonymisation does not suffice for interaction data. The structure alone of, say, phone calls allows individuals to be re-identified, with consequences for data protection on network-like data.
Topics
In this video
Read next
Sources
What is known about this topic outside the broadcast, and where it says so.
Description from the channel
Keynote presentation by Michael Bronstein at ICAI Day (October 27, 2021), organized by the Innovation Center for Artificial Intelligence, a Dutch research and innovation platform.