In transfer learning, a learner leverages auxiliary data from a data-rich source domain to improve prediction on a data-poor target domain. We focus on classification problems in which the Bayes decision boundary changes between the source and target domain, making transfer (learning) necessary for optimal performance in the target domain. First, we study the label shift problem, in which the class conditional distributions remain the same but the (marginal) distribution of the labels are allowed to shift between the source and target domains. Second, we consider the posterior drift problem, in which the regression functions in the source and target domains are allowed to differ. We characterize the minimax risk of both problems and develop practical methods that attain the minimax risks.
Bio:
Dr. Yuekai Sun is an assistant professor in the statistics department at the University of Michigan. His research is guided by the statistical and computational challenges in machine learning. Some topics of recent interest are:
– algorithmic fairness,
– federated learning,
– learning in non-IID settings.
More broadly, Yuekai is interested in the mathematical foundations of data science. He obtained his PhD in computational mathematics from Stanford University, where he worked with Michael Saunders and Jonathan Taylor.