CAREER: Overcoming bias in computer vision: Building fairer systems and training diverse leaders

人工智能（AI）系统已经成为日常生活中不可或缺的一部分。这些系统提供对时事信息的访问，引导购物体验，并允许跨语言边界的交流。例如，计算机视觉系统（根据照片或视频中的视觉信息做出自动决策）正越来越多地部署在自动驾驶或医疗诊断等高风险应用中。然而，众所周知，自动化的人工智能系统会捕捉、传播甚至放大历史偏见、刻板印象和差异：计算机视觉中已知的问题包括面部识别中的种族偏见、物体检测中的地理偏见和活动理解中的性别偏见，等等。这个项目的重点是为计算机视觉系统开发实用的偏见缓解策略。这项工作对于确保计算机视觉在高风险应用中的道德和公平部署是不可或缺的。关于减轻人工智能系统中的社会偏见，有越来越多的文献。其中大部分研究的是带有表格或文本输入的模型的偏见，比如刑事司法记录或简历。减轻计算机视觉中的偏差需要独特的方法：由于输入令牌（单个像素）没有信息，因此在数据和模型中揭示有问题的模式尤其具有挑战性。该项目侧重于视觉保护属性与识别模型预测之间不适当的相关性形式的偏见。它采用多管齐下的方法，包括制定减轻数据偏差的策略（改进数据收集过程和利用合成数据），研究偏差如何从数据传播到下游模型（设计非常适合这一目标的新颖可解释性技术），以及制定直接减轻模型偏差的策略（利用新的基准和指标来通知模型设计）。此外，该项目还解决了当前人工智能研究人员的同质性问题，这是人工智能偏见的根源之一。除了技术创新之外，教育部分还侧重于与国家非营利组织AI4ALL合作，从高中开始为历史上代表性不足的群体的学生提供培训和领导途径。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Artificial Intelligence (AI) systems have become an integral part of daily life. These systems provide access to information about current events, guide shopping experiences, and allow communication across language boundaries. For example, computer vision systems (which make automated decisions based on visual information from photos or videos) are becoming increasingly deployed in high-stakes applications such as autonomous driving or medical diagnosis. However, automated AI systems have been known to capture, propagate and even amplify historical biases, stereotypes, and disparities: known issues in computer vision include racial bias in face recognition, geographic bias in object detection, and gender bias in activity understanding, to name a few. This project focuses on developing practical bias mitigation strategies for computer vision systems. The work is integral to ensuring the ethical and equitable deployment of computer vision in high-stakes applications. There is a rich and growing literature on mitigating social bias in AI systems generally. Much of it studies bias in models with tabular or text input, such as criminal justice records or resumes. Mitigating bias in computer vision requires unique approaches: since the input tokens (single pixels) are uninformative, revealing problematic patterns in data and models is particularly challenging. The project focused on bias in the form of inappropriate correlations between visual protected attributes and predictions of recognition models. It features a multi-pronged approach, which includes developing strategies for mitigating bias in the data (improving data collection processes and leveraging synthetic data), studying how bias propagates from data into downstream models (designing novel interpretability techniques that are well-suited for this goal), and developing strategies for directly mitigating bias in the models (leveraging novel benchmarks and metrics to inform model design). In addition, the project also tackles the problem of homogeneity among current AI researchers, which is one of the root causes of AI bias. Going beyond the technical innovations, the educational component focuses on training and providing leadership pathways for students from historically underrepresented groups starting as early as high school, in partnership with the national nonprofit AI4ALL.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.