收藏本站 | 设为首页 | English
当前位置:首页 -> 瀛︽湳绉戠爺
(6月20日)A New Paradigm for Supervised Learning with Average Top-k Loss
来源: 陈黎   发布时间:2017-6-19   阅读次数:355

题目: A New Paradigm for Supervised Learning with Average Top-k Loss

报告人: Yiming Ying,State University of New York (SUNY) at Albany

时间: 2017620日(星期二)下午 3:30

地点:    工商楼报告厅200-9

 

摘要: The process of supervised learning aims to find an optimal function $f$ from a hypothesis space by minimizing  the ensemble loss $L( \ell_1(f), \cdots,\ell_n(f) )$. Here,  the individual loss $\ell_i(f)$ measures the fitting of $f$ on the $i$-th sample  and the ensemble loss $L(\cdot)$  aggregates all individual losses and represents them collectively.  There are two common choices for the ensemble loss $L$, which is either the {\em average loss} $ {1\over n} \sum_{i=1}^n  \ell_i(f)$ 

or the {\em max loss} $ \max_{i=1}^n \ell_i(f) .$ Both are extensively studied in machine learning. 

In this talk, I will present our recent proposal to  minimize the average of $k$-largest individual losses which is referred to as the Average Top-k model (AT$_k$). It instantiates both the average loss ($k=n$) and the max loss ($k=1$).  The formulation of AT$_k$ is reformulated as a convex optimization problem which can be solved efficiently.  In particular, we show that AT$_k$ with the hinge loss as an individual loss is equivalent to the $\nu$-SVM when the data is normalized.   I will then present  some Learning Theory results of this new learning model, which gives nice implications on how to choose the critical model parameter $k$.  Finally,  some experimental results will be shown to validate the effectiveness of AT$_k$.  This is joint work with Siwei Lyu and Yanbo Fan. 

 


报告人简介: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. He then worked as a Research Associate/Fellow at the City University of Hong Kong, University College London and University of Bristol before 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. 

 

联系人:郭正初 guozc@zju.edu.cn

 欢迎感兴趣的老师和同学们参加!

Copyright © 2003-2017,浙江大学数学系 保留所有权利
联系我们:mathadmin@zju.edu.cn 邮编:310027 电话:0571-87953867