Research
Four pillars of fundamental AI research
Our programme is deliberately theory-first: we work on the questions whose answers change how the next generation of systems is built.
14 active projects
Foundations of Deep Learning
Why do over-parameterised networks generalise? We study optimisation geometry, feature learning dynamics, and scaling behaviour from first principles.
Active threads
- Implicit bias of gradient methods
- Feature learning and representation collapse
- Empirical scaling laws under distribution shift