Debiasing Vision-Language Models via Biased Prompts
This addresses bias issues in vision-language models for downstream applications like zero-shot classification and text-to-image generation, offering a practical, incremental improvement.
The paper tackles the problem of biases in vision-language models inherited from uncurated internet data, proposing a method that debiases text embeddings via projection to reduce social bias and spurious correlation in classifiers and generative models without extra data or training.
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.