PAIRS
收藏资源简介:
PAIRS数据集由加拿大国家研究委员会创建,包含200张AI生成的平行图像,每张图像在背景和视觉内容上高度相似,但在性别和种族上有所不同。该数据集用于研究大型视觉-语言模型中的性别和种族偏见,通过展示相同场景下不同性别和种族的人物图像,观察模型的响应差异。数据集的应用领域主要集中在评估和解决AI模型中的社会偏见问题,旨在通过科学的方法揭示并纠正这些偏见,以促进AI技术的公平性和透明度。
The PAIRS dataset was developed by the National Research Council Canada, and contains 200 AI-generated parallel images. Each paired set of these images shares highly similar backgrounds and visual content, yet differs in terms of the gender and race of the portrayed individuals. This dataset is designed to study gender and racial biases in large vision-language models: by displaying images of individuals with different genders and races within identical scenes, it enables the observation of discrepancies in model responses. The primary application domains of this dataset focus on evaluating and mitigating societal biases in AI models, with the goal of uncovering and correcting these biases through scientific methods to advance fairness and transparency in AI technology.

- 1Examining Gender and Racial Bias in Large Vision-Language Models Using a Novel Dataset of Parallel Images加拿大国家研究委员会 · 2024年



