Symmetry and Generalisation in Neural Approximations of Renormalisation Transformations

Cassidy Ashworth, Pietro Liò, Francesco Caso · 2025

Read Symmetry and Generalisation in Neural Approximations of Renormalisation Transformations by Cassidy Ashworth, Pietro Liò, Francesco Caso (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-16591.txt.

Description

Cassidy Ashworth, Pietro Liò, and Francesco Caso probe how symmetry shapes learning in a 2025 paper. Their test task is a renormalisation step, a way to zoom out on a system. Symmetry means parts of a model can swap with no change in act. Past work hints that breaks and fixes of such symmetry drive layered learning. They use a sum rule from stats as a clean test map.

They train plain layered nets and graph nets with varied symmetry and maths forms. Too much symmetry or too much free form both hurt new-case scores. Proofs show why some tight nets fail to spread their sums well. Tests extend the same lens to graph nets and their inner steps. The work flags a sweet spot between strict form and free form.

Catalog record

CREATORCassidy Ashworth, Pietro Liò, Francesco Caso
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE68 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-16591

Browse the full library · About the Tower of Babel