On Uncertainty Calibration for Equivariant Functions

This repository contains all code for On Uncertainty Calibration for Equivariant Functions by Edward Berman, Jacob Ginesin, Marco Pacini, and Robin Walters.

Running Experiments

First, clone the repo (and the submodules!) via git clone --recurse-submodules https://github.com/EdwardBerman/EquivariantUQBounds.git

Vector Field Regression

Refer to the README.md in the folder for the experiment

Swiss Roll

Instructions are in the ext_theory Git submodule on how to run the "spiral" experiment

Galaxy Morphology

Instructions are in the SIDDA Git submodule on how to run, including a script for producing the noised version of the dataset and a zenodo link to the default .npy files

Chemical Scalar Properties

Run the run.py script in /uq_bleading/

Chemical Vector Properties

See the SEGNN submodule. Specific training scripts will be added upon completion of a follow-up work.

Citation

@misc{berman2025uncertaintycalibrationequivariantfunctions,
      title={On Uncertainty Calibration for Equivariant Functions}, 
      author={Edward Berman and Jacob Ginesin and Marco Pacini and Robin Walters},
      year={2025},
      eprint={2510.21691},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2510.21691}, 
}