Debiasing Vision-Language Models via Biased Prompts

TL;DR

This paper introduces a method to reduce social biases in vision-language models by projecting out biased directions in text embeddings, improving fairness and robustness without extra data or training.

Contribution

The authors propose a simple, closed-form debiasing technique for vision-language models that effectively minimizes biases in both discriminative and generative tasks.

Findings

Reduces social bias and spurious correlations in models

Effective for both classification and generation tasks

No additional data or training required

Abstract

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.