Yue Song

Yue Song is a postdoctoral researcher at the California Institute of Technology (Caltech) in the Department of Computing and Mathematical Sciences, working under the supervision of Yisong Yue, Pietro Perona, and Max Welling. Previously, they completed their PhD at the European Laboratory for Learning and Intelligent Systems (ELLIS), affiliated with the Multimedia and Human Understanding Group (MHUG) at University of Trento (Italy) and Amsterdam Machine Learning Lab (AMLab) at University of Amsterdam (Netherlands), advised by Nicu Sebe and Max Welling.

Education:

• B.Sc. (cum laude) from KU Leuven (Belgium)
• Joint M.Sc. (summa cum laude) from University of Trento (Italy) and KTH Royal Institute of Technology (Sweden)
• Innovation & Entrepreneurship minor from European Institute of Innovation and Technology (EIT Digital)

Research focuses on structured representation learning at the intersection of AI and science. Key themes include:

• Exploiting geometric, temporal, and topological structure in scientific data
• Integrating inductive biases from physics/chemistry/biology into ML models (Science4AI)
• Using AI to uncover scientific insights (AI4Science)
• Generalizable, interpretable, and data-efficient methods

Recent publication trends show work across:

• Latent space analysis for GANs/diffusion models (ICCV 2023, CVPR 2023)
• Geometric and Lie group-based ML (ICLR 2024, ICLR 2025)
• Theoretical foundations of matrix operations in deep learning (TPAMI 2022, TPAMI 2024)
• Scientific applications in chemistry, neuroscience, and physics (NeurIPS 2024)

Major contributions include:

• 2025: Co-authored book Structured Representation Learning (Springer Nature)
• 2024: Area Chair for NeurIPS 2025

Collaborations span institutions: Caltech, University of Trento, University of Amsterdam, KU Leuven, KTH. Codebases available on GitHub demonstrate applications in:

• Fast differentiable matrix operations (ICLR 2022, TPAMI 2022)
• Householder projectors for GANs/diffusion models (ICCV 2023)
• Latent flow dynamics (ICML 2023)