20160706-21李宇峰:Safe Semi-Supervised Learning

报告嘉宾：李宇峰(南京大学计算机系机器学习与数据挖掘研究所（LAMDA）)

报告时间：2016年7月6日（星期三）晚21:00（北京时间）

报告题目：Safe Semi-Supervised Learning

主持人：  张利军（南京大学）

报告摘要：It is often expected that when labelled data is few, semi-supervised learning model with the use of a large amount of unlabelled data can improve the performance. With the deepening of research, semi-supervised learning model using a large amount of unlabelled data sometimes not only cannot get performance improvement, and even lead to performance degradation. Therefore, safe semi-supervised learning, which does not degenerate performance when using unlabelled data, has become an important issue in semi-supervised learning. This talk introduces our progress on this problem, presented by studying some potential reasons for the performance degradation in semi-supervised learning. Specifically, in the respect of optimization, not-high quality of optimization solution may degenerate the performance. We present a convex yet scalable algorithm to control the quality of the solution particularly for the scenarios where the scale of data becomes large. In the respect of model selection, the uncertainty of model selection in semi-supervised scenario may degenerate the performance. We present a safe semi-supervised SVM using the worst-case analysis. In the respect of performance evaluation, the diversity of evaluation may degenerate the performance. We present a safe semi-supervised model for multiple performance evaluations with an efficient optimization algorithm.

参考文献：Y.-F. Li, I. Tsang, J. Kwok and Z.-H. Zhou. Convex and Scalable Weakly Labeled SVMs. Journal of Machine Learning Research, 14:2151-2188, 2013. 

Y.-F. Li and Z.-H. Zhou. Towards making unlabeled data never hurt. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(1):175-188, 2015. 

Y.-F. Li, J. Kwok and Z.-H. Zhou. Towards safe semi-supervised learning for multivariate performance measures. In: Proceedings of the 30th AAAI conference on Artificial Intelligence (AAAI'16), Phoenix, AZ, 2016.

报告人简介：李宇峰博士，南京大学计算机系机器学习与数据挖掘研究所（LAMDA）助理研究员。他主要对机器学习中半监督学习及相关领域开展研究，在机器学习相关重要期刊会议如JMLR、TPAMI、ICML、IJCAI、AAAI等发表论文20余篇。获2013年中国计算机学会优秀博士学位论文奖、2014年江苏省优秀博士论文奖。受邀担任了国际会议IJCAI’15高级程序委员、以及多个国际会议的程序委员如ICML’16、NIPS’16、KDD’16、CVPR’16、AAAI’16等。入选2015年CCF青年人才发展计划。

[Slides]