A method to challenge symmetries in data with self-supervised learning

TL;DR

This paper presents a general self-supervised approach to test suspected symmetries in data, effectively handling filtered data and providing a practical tool for physical and experimental analysis.

Contribution

The authors introduce a novel self-supervised method to challenge symmetries in data, capable of working with filtered datasets and minimal external input.

Findings

Successfully challenges symmetries in various datasets.

Handles filtered data without bias.

Provides interpretable model predictions.

Abstract

Symmetries are key properties of physical models and of experimental designs, but any proposed symmetry may or may not be realized in nature. In this paper, we introduce a practical and general method to test such suspected symmetries in data, with minimal external input. Self-supervision, which derives learning objectives from data without external labelling, is used to train models to predict 'which is real?' between real data and symmetrically transformed alternatives. If these models make successful predictions in independent tests, then they challenge the targeted symmetries. Crucially, our method handles filtered data, which often arise from inefficiencies or deliberate selections, and which could give the illusion of asymmetry if mistreated. We use examples to demonstrate how the method works and how the models' predictions can be interpreted. Code and data are available at https://zenodo.org/record/6861702.