Achieving linear scaling in ML for particle physics: CARDINAL

Jul 30, 2025, 1:30 PM
20m
Barus and Holley 168 (Brown University)

Barus and Holley 168

Brown University

184 Hope St, Providence, RI 02912
Oral Presentation Plenary Talks

Speaker

Alexander Bogatskiy (University of Chicago)

Description

State-of-the-art machine learning models in particle physics, such as PELICAN [1] and Particle Transformer [2], exhibit disadvantageous $\mathcal{O}(N^2)$ scaling with the multiplicity $N$ of the jets. Fundamentally, this has to do with the tendency for highly expressive models to involve complicated permutation-equivariant layers to model pairwise interactions. Similar limitations arise in other fields that use graph neural networks, such as chemistry. I will present a radically simpler architecture, CARDINAL, which balances universal expressivity and ultra-low computational cost by employing a permutation-invariant (and also Lorentz-invariant) embedding which allows for $\mathcal{O}(N)$ scaling. In common benchmark tests, it is tens or hundreds of times faster and less memory-demanding than comparable popular architectures.

[1] Alexander Bogatskiy, Timothy Hoffman, David W. Miller, Jan T. Offermann & Xiaoyang Liu, Explainable equivariant neural networks for particle physics: PELICAN, 2024.
[2] Huilin Qu, Congqiao Li, Sitian Qian, Particle Transformer for Jet Tagging, 2022.

Author

Alexander Bogatskiy (University of Chicago)

Presentation materials