University of Amsterdam

The University of Amsterdam's AI footprint runs through Max Welling's AMLab — a founding home of Bayesian and geometric deep learning, whose Clifford-equivariance line (Clifford Group Equivariant Networks → Clifford-Steerable CNNs) and physical-systems transformers (Erwin, ICML 2025) target PDE and molecular simulation at scale. Welling co-founded materials-discovery startup CuspAI in 2024 and is primarily engaged there; he retains his UvA chair, and CuspAI-branded output is tracked separately, not under UvA.

The vision side is anchored by Cees Snoek's Video & Image Sense Lab and HAVA-Lab (ELLIS Amsterdam), which shipped NeoBabel — an open multilingual text-to-image foundation model with a 124M-pair multilingual dataset — and TULIP, the token-length-upgraded CLIP. Efstratios Gavves' Physical AI program (QUVA/POP-AART) drives world models as learnable digital twins (DreMa, ICLR 2025) and physical-reasoning benchmarks, while Erik Bekkers is the geometric-DL hub: the PONITA weight-sharing framework underpins Equivariant Neural Fields and its seismic Eikonal follow-up.

Maarten de Rijke's IRLab rounds out the picture with an ERC generative information-retrieval program and co-organization of the NIST TREC iKAT track for personalized conversational search. Legacy org repos (AMLab-Amsterdam, QUVA-Lab) are largely historical — active code now ships via personal repos (maxxxzdn/erwin, ivonajdenkoska/tulip, mmderakhshani/NeoBabel).