A Transfer Learning Approach to Algorithmic Fairness

在当今数据驱动的世界里，机器学习模型经常用于刑事司法、教育、贷款、医疗和许多其他领域的高风险决策。尽管用机器学习模型取代人类似乎消除了决策过程中的人类偏见，但它们可能会使训练数据中的偏见永久化，甚至加剧。当这种算法偏见对弱势群体产生不利影响时，尤其令人反感。在这个项目中，我们专注于检测和缓解由于训练数据中的采样偏差而导致的算法偏差。该项目还为研究生提供了研究培训机会。有三个目标。首先，PI确定现有算法公平实践的能力中的差距，以克服训练数据中的抽样偏差。PI还研究机器学习(ML)模型发展的当前趋势(例如，数据增强和过度参数化)如何使算法偏差永久化和加剧。其次，PI将公平的机器学习问题转化为迁移学习问题，并利用迁移学习的最新发展来检测和缓解由采样偏差引起的算法偏差。第三，PI考虑如何收集更能代表总体的训练数据集，并产生没有算法偏差的ML模型。最终目标是扩大算法公平实践在ML从业者中的吸引力和采用率。PIS计划证明，算法公平的迁移学习方法避免了实现这一最终目标的两个障碍：(I)它通过避免准确性和公平性之间的明显权衡，使算法公平的目标与ML实践者的目标保持一致，以及(Ii)它通过提供对算法公平实践的有效性的客观衡量，解决了许多ML任务中对算法公平实践选择缺乏共识的问题。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

In today's data-driven world, machine learning models are routinely used to make high-stakes decisions in criminal justice, education, lending, medicine, and many other areas. Although replacing humans with machine learning models appears to eliminate human biases in decision-making processes, they may perpetuate or even exacerbate biases in the training data. Such algorithmic biases are especially objectionable when they adversely affect underprivileged groups. In this project, we focus on detecting and mitigating algorithmic biases that are caused by sampling biases in the training data. The project also provides research training opportunities for graduate students. There are three aims. First, the PIs identify gaps in the capabilities of existing algorithmic fairness practices for overcoming sampling biases in the training data. The PIs also study how current trends in the development of machine learning (ML) models (for example, data augmentation and overparameterization) can perpetuate and exacerbate algorithmic biases. Second, the PIs cast the fair machine learning problem as a transfer learning problem and leverage recent developments in transfer learning to detect and mitigate algorithmic biases caused by sampling bias. Third, the PIs consider how to collect training datasets that are more representative of the general population and beget ML models that are free from algorithmic biases. The ultimate goal is to broaden the appeal and adoption of algorithmic fairness practices among ML practitioners. The PIs plan to demonstrate that the transfer learning approach to algorithmic fairness avoids two barriers in the way of this ultimate goal: (i) it aligns the goal of algorithmic fairness with the goals of (possibly non-altruistic) ML practitioners by avoiding the apparent trade-off between accuracy and fairness, and (ii) it addresses the lack of consensus on the choice of algorithmic fairness practice in many ML tasks by providing an objective measure of the efficacy of such practices.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.