•   When: Thursday, March 25, 2021 from 11:00 AM to 12:00 PM
•   Speakers: Mingrui Liu, Postdoctoral Fellow, Rafik B. Hariri Institute for Computing, Boston University
•   Location: ZOOM
•   Export to iCal

Abstract:

Classical learning paradigm with traditional objective function is not sufficient to learn efficiently from complex data (e.g., imbalanced data, unlabeled data), since the traditional objective function is usually not a good indicator of the default performance metric people care about in practice. To bridge this gap, the learning systems have to be equipped with the ability to handle novel objectives. My research empowers the learning systems with provably efficient algorithms for optimizing several novel objectives, including maximizing the Area under the ROC Curve (AUC) and training Generative Adversarial Networks (GANs).

 This talk will showcase a few results. First, I will focus on provably efficient algorithms for online AUC maximization with both linear model and deep neural network to learn from imbalanced data. The main result is that AUC maximization can be reformulated as a min-max objective and algorithms with fast convergence rates are established. 

 Second, I will focus on principled adaptive gradient algorithms for training Generative Adversarial Networks (GANs) to learn from unlabeled data. In sharp contrast to heuristics algorithms which usually diverge and need lots of human’s tuning efforts, the proposed algorithm provably converges with a fast adaptive rate and with minimal parameter tuning.

 Brief Bio:

Mingrui Liu is a postdoctoral fellow at Rafik B. Hariri Institute for Computing at Boston University, working with Francesco Orabona. Previously he obtained his Ph.D. at The University of Iowa in August 2020 under the advise of Tianbao Yang. He has also spent time working as a research intern at IBM T. J. Watson Research Center. His research interests include machine learning, convex and non-convex optimization, deep learning, distributed learning, online learning, and continual learning. He serves as an senior program committee member of IJCAI.