Core Insights - The research published in "Nature" presents a database of over 10,000 human images aimed at assessing and correcting biases in AI models within the visual domain, marking a significant step towards more trustworthy AI [1][4] - The "Fair Human-Centric Image Benchmark" (FHIBE) was developed by Sony AI, utilizing ethically sourced data with user consent, which allows for precise evaluation of human-centered computer vision tasks [1][4] Group 1 - FHIBE includes 10,318 images from 1,981 individuals across 81 countries and regions, with comprehensive annotations on demographic and physiological characteristics such as age, pronoun category, ancestry, hair color, and skin color [1][2] - The dataset adheres to best practices in consent mechanisms, diversity, and privacy, making it a reliable resource for assessing AI biases [1][2] - The research team compared FHIBE with 27 existing human-centric computer vision datasets, finding that FHIBE has higher standards for diversity and reliable consent, effectively reducing biases [2] Group 2 - The creation of the dataset is acknowledged to be challenging and costly, indicating potential barriers to widespread adoption [3] - The study represents a benchmark in AI ethics, transforming the abstract principle of "fairness" into actionable and verifiable technical standards and workflows [4] - This exploration is seen as crucial for shifting AI development from merely enhancing performance to becoming a trustworthy partner for humanity [4]

该论文介绍，计算机视觉广泛应用于自动驾驶车辆到面部识别技术等领域。许多计算机视觉使用的AI 模型，其训练数据存在缺陷，可能未经同意收集，经常来自网络大规模图像抓取。人们也已发现，AI 模型可能会反映出延续性别歧视、种族歧视或其他刻板印象的偏见。 在本项研究中，索尼AI团队构建了一个图像数据集，在同意机制、多样性和隐私等多方面努力采取最 佳实践：FHIBE包含81个国家或地区中1981个个体的10318张图像，该数据库涵盖人口统计和生理特征 的全面标注，包括年龄、代词类别、祖先血统、发色与肤色等；参与者获得关于项目和潜在风险的详细 信息，帮助他们做出知情同意，过程符合全面数据保护法规。这些特征使该数据库成为评估AI偏见的 可靠资源。 研究团队将FHIBE和27个现有以人类为中心的计算机视觉应用数据集进行比较，发现FHIBE数据集在多 样性与AI评估的可靠同意方面标准更高，同时还有效减少了偏见，它包含的参与者自我申报标注信息 超过其他数据集，还包括了相当比例通常代表性不足的人群。 本项研究的相关图像数据(图片来自论文)。施普林格·自然 供图 国际最新研究构建超万张人类图像数据库 评估人工智能视觉偏见 中新网北京1 ...