Variation of Gender Biases in Visual Recognition Models Before and After Finetuning

News Release Summary

Researchers from USC, Columbia, Meta AI, and Rice University have developed a framework to track how gender biases embedded in popular pretrained computer vision models behave when those models are fine-tuned for specific downstream tasks — a question that has received surprisingly little systematic attention despite how widely these pretrained models are reused in industry and academia. The team built curated sets of reference images drawn from the COCO and OpenImages datasets, pairing everyday objects like surfboards and cars with images of men and women, and then measured how closely a model's internal feature representations clustered those object images with one gender versus the other, both before and after fine-tuning. Testing six off-the-shelf models — including ResNet variants, CLIP, MoCo, and SimCLR — across different architectures, training dataset sizes, and supervised versus self-supervised training regimes, the researchers found three notable patterns: models supervised on large-scale datasets like ImageNet-21k tended to carry their pretraining biases stubbornly into whatever new task they were fine-tuned on; fine-tuning on a larger target dataset like OpenImages was more likely to introduce fresh biased associations than fine-tuning on a smaller one like COCO; and self-supervised models, particularly MoCo, showed less bias retention than their supervised counterparts, though this was not universal. To quantify these dynamics, the team introduced a metric called the Bias Transfer Score, based on Spearman's rank correlation, which measures how much a model's bias-related associations shift between the pretraining and fine-tuning stages. The work is practically significant because many developers adopt pretrained models as black boxes without visibility into what societal biases they may be silently importing into their applications.