地点：理B与理C中间，学生中心

时间：6月5日，15：30-16：30

主讲人：应益明

Title:  Online Learning with Pairwise Loss Functions Abstract: Pairwise learning usually refers to a learning task which involves a loss function depending on pairs of examples, among whichmost notable ones include bipartite ranking, metric learning, minimum entropy error  and AUC maximization.  Online learning  algorithms arewidely used in practice. However, most of such algorithms focused on the pointwise learning problems such as classification and regression.  A specific challenge in developing and analyzing online pairwise learning algorithms is that the objective function is usually defined over pairs of instances which is quadratic in the number of instances. In this talk, we study a general online learning  algorithm for pairwise learning in an unconstrained setting of a reproducing kernel Hilbert space (RKHS). We present convergence analysis for these algorithms in both regularized and un-regularized settings. The above general online algorithms require to store previous examples which is not memory efficient.  We show that, for AUC maximization, we can develop a truly online algorithm for which the space and per-iteration complexities only depend linearly on one datum. The key idea behind this is a novel formulation of AUC maximization as a stochastic saddle point problem.

Short Bio:  Yiming Ying is currently a tenured Associate Professor in the Department of Mathematics and Statistics at the State University of New York (SUNY) at Albany, USA. He received the PhD degree in Mathematics in 2002 from Zhejiang University under the supervision of Profs Silei Wang and Jiecheng Chen. After graduation, he worked as a Research Associate/Fellow at the City University of Hong Kong, University College London and University of Bristolbefore he became a Lecturer (Assistant Professor) in Computer Science at the University of Exeter (UK) in 2010.  His current research interests include learning theory and machine learning and their applications.