CAREER: Fair Artificial Intelligence for Intelligent Humans: Removing the Barriers to Deployment of Fair AI Technologies

越来越多的人意识到，人工智能（AI）和机器学习系统在某些情况下可能会以不公平和歧视性的方式行事，在许多领域（包括刑事司法，招聘，医学和大学招生）造成有害后果。确保AI公平性的技术在AI文献中受到了很多关注。 然而，这些技术还没有在真实的系统中看到相当程度的部署，这因此限制了它们在现实世界中的影响。这可能部分是由于部署公平AI技术的几个实际挑战。首先，传统观点认为，公平性会带来预测性能的成本，这可能会影响组织的底线。其次，很难知道人工智能公平性的数学定义是适当的，因为这些定义相互冲突，并编码不同的价值体系。最后，还有一个鸡和蛋的问题，即要求组织将公平考虑纳入人工智能系统的公众压力只会在其他地方成功证明后才会增加。这项研究将开发技术解决方案，以解决这些面向人类的障碍，以采用AI公平技术，从而增加部署和后续的积极现实世界的影响。为了解决公平AI技术的实际局限性，这项研究将以人为本的考虑纳入公平AI算法的设计和执行，连接和推进统计机器学习，公平AI，以人为本的AI 该项目的第一个轨道将开发获得“免费公平”的方法，其中在牺牲很少或没有性能的情况下找到最公平的解决方案。 研究人员将设计黑盒、灰盒和白盒方法来完成这项任务。 然后，研究的第二个轨道将专注于开发可解释的人工智能和数据可视化技术，以帮助人类评估和权衡不同公平竞争概念的后果。 实现这一目标的关键步骤是创建一个统一的公平性框架，该框架系统地编码可能的公平性度量空间。 最后，在该项目的第三个轨道中，研究人员将为人工智能公平的几个现实应用开发实用的解决方案，包括医疗资源的分配和基于人工智能的职业咨询。这些解决方案将涉及应用和基础人工智能研究，并将促进对前两个轨道中开发的方法的评估。 该项目还包括推广活动，扩大科学，技术，工程和数学（STEM）领域的参与，培训和教育研究生，以及课程开发。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

There is growing awareness that artificial intelligence (AI) and machine learning systems can in some cases behave in unfair and discriminatory ways with harmful consequences in many areas including criminal justice, hiring, medicine, and college admissions. Techniques for ensuring AI fairness have received a lot of attention in the AI literature. However, these techniques are yet to see a substantial degree of deployment in real systems, which has thus-far limited their real-world impact. This is likely due in part to several practical challenges for deploying fair AI technologies. Firstly, the conventional wisdom is that fairness brings a cost in prediction performance which could affect an organization's bottom-line. Secondly, it is difficult to know which mathematical definition of AI fairness is appropriate to adopt since the definitions conflict with each other and encode different value systems. Finally, there is a chicken-and-egg problem, in that public pressure for an organization to adopt fairness considerations into an AI system only increases after this has been successfully demonstrated elsewhere. This research will develop technical solutions to resolve these human-facing barriers for the adoption of AI fairness techniques, thereby increasing deployment and the subsequent positive real-world impact.To resolve the practical limitations of fair AI techniques, this research incorporates human-centered considerations into the design and execution of fair AI algorithms, connecting and advancing the state of the art in statistical machine learning, fair AI, and human-centered AI. The first track of the project will develop methods for obtaining “fairness for free,” in which the fairest possible solution is found when sacrificing little-to-no performance. The researchers will design black-box, gray-box, and white-box approaches to this task. Then, the second track of the research will focus on developing explainable AI and data visualization techniques to help humans assess and trade off the consequences of different competing notions of fairness. A key step to accomplish this is to create a unifying fairness framework which systematically encodes the space of possible fairness metrics. Finally, in the third track of the project, the researchers will develop practical solutions to several real-world applications of AI fairness, including the allocation of medical resources, and AI-based career counseling. The solutions will involve both applied and fundamental AI research, and will facilitate the evaluation of the methods developed in the first two tracks. The project also includes initiatives for outreach, broadening participation in science, technology, engineering, and mathematics (STEM) fields, training and educating graduate students, and curriculum development.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.