FAI: Fair Representation Learning: Fundamental Trade-Offs and Algorithms

基于人工智能的计算机系统越来越依赖于有效的信息表示，以支持从图像识别系统到人脸识别的身份控制等领域的决策。然而，依赖传统统计数据和来自历史或人为整理数据的预测的系统也自然地继承了任何过去的偏见或歧视倾向。该奖项的总体目标是通过使用保持其效用的信息表示来缓解这一问题，同时消除可能导致对人群中子群体歧视的信息。具体来说，该项目将研究不同数据表示的效用和公平性之间的不同权衡，然后确定解决方案，以减少与最佳权衡之间的差距。然后，在这种权衡分析的指导下，开发新的表示和相应的算法。研究人员将根据已开发的理论提供性能限制，以及有效性证据，以获得公平的机器学习系统并获得社会信任。本研究的应用领域是人脸识别系统。参与该项目的本科生和研究生将接受培训，进行前沿研究，将公平融入基于人工智能的系统。这个项目的研究议程围绕着回答关于学习公平表示的两个问题，(i)数据表示的效用和公平性之间的基本权衡是什么？（ii）如何设计实用的公平表示学习算法，以减轻机器学习系统中的偏见，并可证明地实现理论上的效用-公平权衡？为了回答第一个问题，该项目将从理论上阐明和实证量化效用固有的不同权衡。这将考虑到不同的公平定义，如人口均等、均等几率和机会均等。为了回答第二个问题，该项目将开发表征学习算法，这些算法(a)在分析上易于处理且可证明是公平的，(b)减轻最坏情况偏差，而不是对实例或人口群体的平均偏差，(c)对于仅部分已知或完全未知的人口信息是公平的，以及(d)减轻由于样本不平衡以及通过最佳数据采样和投影的特征造成的人口偏差。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 based computer systems are increasingly reliant on effective information representation in order to support decision making in domains ranging from image recognition systems to identity control through face recognition. However, systems that rely on traditional statistics and prediction from historical or human-curated data also naturally inherit any past biased or discriminative tendencies. The overarching goal of the award is to mitigate this problem by using information representations that maintain its utility while eliminating information that could lead to discrimination against subgroups in a population. Specifically, this project will study the different trade-offs between utility and fairness of different data representations, and then identify solutions to reduce the gap to the best trade-off. Then, new representations and corresponding algorithms will be developed guided by such trade-off analysis. The investigators will provide performance limits based on the developed theory, and also evidence of efficacy in order to obtain fair machine learning systems and to gain societal trust. The application domain used in this research is face recognition systems.The undergraduate and graduate students who participate in the project will be trained to conduct cutting-edge research to integrate fairness into artificial intelligent based systems.The research agenda of this project is centered around answering two questions on learning fair representations, (i) What are the fundamental trade-offs between utility and fairness of data representations?, (ii) How to devise practical fair representation learning algorithms that can mitigate bias in machine learning systems and provably achieve the theoretical utility-fairness trade-offs? To answer the first question, the project will theoretically elucidate and empirically quantify the different trade-offs inherent to utility. This will be done consideringdifferent fairness definitions such as demographic parity, equalized odds, and equality of opportunity. To answer the second question, the project will develop representation learning algorithms that (a) are analytically tractable and provably fair, (b) mitigate worst-case bias, as opposed to average bias over instances or demographic groups, (c) are fair with respect to demographic information that is only partially known or fully unknown, and (d) mitigate demographic bias both due to an imbalance in samples as well as features through optimal data sampling and projection.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.