News
2026.10.17
Version 2 of my preprint on equivariant attention is out, now with Georg Bökman as co-author. It has a new title, Inherent Trade-offs when Enforcing Correct Data Symmetries in Multi-head Self-attention, and adds experiments.
2026.09.10
I will be hosting the workshop Geometric Viewpoints on Learning, Data, and Neural Networks, which takes place in Munich from 23 to 25 September.
2026.09.01
New preprint on the geometry of equivariant attention! I use some algebraic geometry to show that no single polynomially parameterized equivariant architecture can represent all equivariant functions representable by an unconstrained multi-head self-attention layer.