Yang Liu
Assistant Professor
Computer Science and Engineering
UC Santa Cruz
Abstract: Learning with noisy labels is a prevalent challenge in machine learning: in supervised learning, the training labels are often solicited from human annotators, which encode human-level mistakes; in semi-supervised learning, the artificially supervised pseudo labels are immediately imperfect. The list goes on. Existing approaches with theoretical guarantees often require practitioners to specify a set of parameters controlling the severity of label noises in the problem. The specifications are either assumed to be given or estimated using additional approaches.
In this talk, I introduce peer loss functions, which enable learning from noisy labels and do not require a priori specification of the noise rates. Peer loss functions associate each training sample with a specific form of “peer” sample, which helps evaluate a classifier’s predictions jointly. We show that, under mild conditions, performing empirical risk minimization (ERM) with peer loss functions on the noisy dataset leads to the optimal or a near-optimal classifier as if performing ERM over the clean training data. Peer loss provides a way to simplify model development when facing potentially noisy training labels. I will also discuss extensions of peer loss and some emerging challenges concerning biased data.
Bio: Yang Liu is currently an Assistant Professor of Computer Science and Engineering at UC Santa Cruz (2019 – present). He was previously a postdoctoral fellow at Harvard University (2016 – 2018). He obtained his Ph.D. degree from the University of Michigan, Ann Arbor in 2015, advised by Professor Mingyan Liu. He is interested in weakly supervised learning and algorithmic fairness. He is a recipient of the NSF CAREER Award and the NSF Fairness in AI award. He has been selected to participate in several high-profile projects, including DARPA SCORE and IARPA HFC. His recent works have won four best paper awards at relevant workshops.
