Resampled Datasets Are Not Enough: Mitigating Societal Bias Beyond Single Attributes
This addresses societal bias in AI datasets for fairness-critical applications like image classification and captioning, representing a novel extension beyond single-attribute methods.
The paper tackles societal bias in image-text datasets by removing spurious correlations between protected groups and both labeled and unlabeled image attributes, using text-guided inpainting models and data filtering. Evaluations on multi-label image classification and image captioning tasks show the method effectively reduces bias without compromising performance across various models.
We tackle societal bias in image-text datasets by removing spurious correlations between protected groups and image attributes. Traditional methods only target labeled attributes, ignoring biases from unlabeled ones. Using text-guided inpainting models, our approach ensures protected group independence from all attributes and mitigates inpainting biases through data filtering. Evaluations on multi-label image classification and image captioning tasks show our method effectively reduces bias without compromising performance across various models.