Events

Modern digital systems powered by machine learning have permeated various aspects of society, playing an instrumental role in many high-stakes areas such as medical care and finance. Therefore, it is crucial to ensure that machine learning algorithms are deployed in a “responsible” way so that digital systems are more reliable, transparent, and aligned with societal values. In this talk, I will introduce my research on building practical theories to guide real-world responsible deployment of machine learning. First, I will introduce our recent work [1] on distribution-free uncertainty quantification for a rich class of statistical functionals of quantile functions to avoid catastrophic outcomes and unfair discrimination in the deployment of black-box models. The power of our framework is shown by end-to-end applications on large language models. Second, I will describe an extension [2] to the previous framework on the group-based fairness notions so as to protect every group that can be meaningfully identified from data. This result further strengthens our previous framework and enables us to address some challenges left open by the previous literature. Finally, I will further demonstrate how to ensure fairness generalization for complex white-box models, such as neural networks, under data imbalance by a theory-inspired algorithm [3]. I will conclude my talk with future directions.