The Bias of Harmful Label Associations in Vision-Language Models
This work addresses fairness issues in AI for diverse populations, highlighting persistent biases in vision-language models despite performance improvements.
The study investigated harmful label associations in vision-language models using the Casual Conversations dataset, finding that models are 4-7 times more likely to make harmful classifications for individuals with darker skin tones, and that scaling model size increases confidence in these predictions without addressing disparities.
Despite the remarkable performance of foundation vision-language models, the shared representation space for text and vision can also encode harmful label associations detrimental to fairness. While prior work has uncovered bias in vision-language models' (VLMs) classification performance across geography, work has been limited along the important axis of harmful label associations due to a lack of rich, labeled data. In this work, we investigate harmful label associations in the recently released Casual Conversations datasets containing more than 70,000 videos. We study bias in the frequency of harmful label associations across self-provided labels for age, gender, apparent skin tone, and physical adornments across several leading VLMs. We find that VLMs are $4-7$x more likely to harmfully classify individuals with darker skin tones. We also find scaling transformer encoder model size leads to higher confidence in harmful predictions. Finally, we find improvements on standard vision tasks across VLMs does not address disparities in harmful label associations.