Pim de Haan

About

Pim de Haan is a researcher at the forefront of geometric deep learning and causal representation learning. His work bridges the gap between abstract mathematical structures and practical machine learning architectures. In his highly influential paper "Weakly Supervised Causal Representation Learning" (2022, 13 citations), de Haan proved that high-level causal representations can be identified from low-level pixel data under weak supervision—a breakthrough that challenges long-held assumptions about the impossibility of causal discovery from observational data alone. This work provides a rigorous theoretical foundation for learning causal models from unstructured inputs, opening new avenues for robust and interpretable AI systems. More recently, de Haan introduced the "Geometric Algebra Transformer" (2023, 9 citations), a novel architecture capable of handling diverse geometric data types—points, vectors, rotations, and translations—within a single unified framework. This innovation has broad implications for fields ranging from robotics and computer vision to physics and chemistry, where geometric structure is paramount. De Haan’s contributions are shaping the next generation of machine learning models that are both geometrically aware and causally grounded.

Research Focus

Key Achievements

Top Papers

1. 1
   Weakly supervised causal representation learning

   13 citations · 2022
2. 2
   Geometric Algebra Transformer

   9 citations · 2023

Key Collaborators

Contact & Links