Home

Awesome

Fair Attribute Classification through Latent Space De-biasing

Project Page | Paper | Colab Notebook | 2min Talk | 5min Talk | 10min Talk

This repo provides the code for our CVPR 2021 paper "Fair Attribute Classification through Latent Space De-biasing."

@inproceedings{ramaswamy2020debiasing,
author = {Vikram V. Ramaswamy and Sunnie S. Y. Kim and Olga Russakovsky},
title = {Fair Attribute Classification through Latent Space De-biasing},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2021}
}

Our work is featured in Coursera's Generative Adversarial Networks (GANs) Specialization course. Check out the colab notebook linked above for details.

Main experiments

Data processing:

Baseline:

GAN:

<!--Train a progressive GAN on celeba (code here: https://github.com/facebookresearch/pytorch_GAN_zoo), save the final model in record/GAN_model/final_model.pt (or set pretrained=True in generate_images.py)-->

Our model:

Extensions of our method

Using domain-dependent hyperplanes:

Augmenting real-images with GAN-inversion:

Augmenting two protected attributes:

Additional experiments

Acknowledgements

This work is supported by the National Science Foundation under Grant No. 1763642 and the Princeton First Year Fellowship to SK. We also thank Arvind Narayanan, Deniz Oktay, Angelina Wang, Zeyu Wang, Felix Yu, Sharon Zhang, as well as the Bias in AI reading group for helpful comments and suggestions.