Machine learning models have been shown to inherit biases from their training datasets. This can be particularly problematic for vision-language foundation models trained on uncurated datasets scraped from the internet. The biases can be amplified and propagated to downstream applications like zero-shot classifiers and text-to-image generative models. In this study, we propose a general approach for debiasing vision-language foundation models by projecting out biased directions in the text embedding. In particular, we show that debiasing only the text embedding with a calibrated projection matrix suffices to yield robust classifiers and fair generative models. The proposed closed-form solution enables easy integration into large-scale pipelines, and empirical results demonstrate that our approach effectively reduces social bias and spurious correlation in both discriminative and generative vision-language models without the need for additional data or training.
@article{arxiv.2302.00070,
title = {Debiasing Vision-Language Models via Biased Prompts},
author = {Ching-Yao Chuang and Varun Jampani and Yuanzhen Li and Antonio Torralba and Stefanie Jegelka},
journal= {arXiv preprint arXiv:2302.00070},
year = {2023}
}