https://github.com/kurtisdavid/metaorthogonalization
Code for Debiasing Convolutional Neural Networks via Meta Orthogonalization (NeuRIPS 2020 AFCI Workshop).
https://github.com/kurtisdavid/metaorthogonalization
computer-vision deep-learning fairness-ml interpretability machine-learning
Last synced: about 2 months ago
JSON representation
Code for Debiasing Convolutional Neural Networks via Meta Orthogonalization (NeuRIPS 2020 AFCI Workshop).
- Host: GitHub
- URL: https://github.com/kurtisdavid/metaorthogonalization
- Owner: kurtisdavid
- Created: 2020-11-15T03:29:19.000Z (almost 5 years ago)
- Default Branch: main
- Last Pushed: 2020-12-07T01:14:14.000Z (almost 5 years ago)
- Last Synced: 2025-07-31T13:31:13.738Z (2 months ago)
- Topics: computer-vision, deep-learning, fairness-ml, interpretability, machine-learning
- Language: Python
- Homepage: https://arxiv.org/abs/2011.07453
- Size: 514 KB
- Stars: 4
- Watchers: 1
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# MetaOrthogonalization
[Kurtis Evan David](https://kurtisdavid.github.io),
[Qiang Liu](https://www.cs.utexas.edu/~lqiang/),
[Ruth Fong](https://ruthcfong.github.io/)
## [[Paper]](https://arxiv.org/abs/2011.07453)
## Citation
```bibtex
@article{david2020debiasing,
title={Debiasing Convolutional Neural Networks via Meta Orthogonalization},
author={Kurtis Evan David and Qiang Liu and Ruth Fong},
year={2020},
journal={arXiv preprint arXiv:2011.07453},
}
```
## Overview
This is the official implementation of Meta Orthogonalization in [PyTorch](https://pytorch.org/).
Our main idea is to debias convolutional networks by making downstream concepts be orthogonal to a learned bias direction.
This is directly inspired by similar methods in NLP by [Bolukbasi et al. (2016)](https://arxiv.org/abs/1607.06520).
## Required Libraries
To successfully run the code in this repo, make sure you have the following libraries installed:
* PyTorch 1.5.0 (CUDA 10.1)
* [higher](https://github.com/facebookresearch/higher)
* [bam](https://github.com/google-research-datasets/bam)
## Acknowledgement
We would like to thank Tianlu Wang for collaborating and providing more details with [adversarial debiasing](https://arxiv.org/abs/1811.08489) and their models. Please also cite their paper if you use this codebase for COCO, as we use theirs provided [here](https://github.com/uvavision/Balanced-Datasets-Are-Not-Enough).
## Contact
If you have any questions, please contact us through kurtis.e.david(at)gmail.com.