学术报告——Restricted GPCA: Low-Rank Representation and its Applications (UIUC)
发布时间:2013-04-18 阅读次数:
Date:4月22日上午10：00 Venue:九教北307B
Title:Restricted GPCA: Low-Rank Representation and its Applications
Speaker:Guang-can Liu. University of Illinois at Urbana-Champaign
Dr. Guang-can Liu received the bachelor degree of mathematics from Shanghai Jiao Tong University (SJTU) in 2004，and the PhD degree of computer science from SJTU, 2010. From Sep 2010 to Jan 2012, he was a research fellow at National University of Singapore (NUS). Currently, he is an associate researcher at University of Illinois at Urbana-Champaign (UIUC). His research areas include machine learning, computer vision, image processing, multimedia and data mining.
Dr. Liu近年来在TPAMI、Neural Computaiton、TIP等高水平学术期刊以及ICML、NIPS、ICCV、CVPR、AAAI等高水平学术会议上发表多篇学术论文。尤其近期在Low rank Representation方面开展了深入研究并取得了一系列高水平成果，该方向代表性成果如下：1. Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma,Robust Recovery of Subspace Structures by Low-Rank Representation, IEEE Trans. on Pattern Analysis and Machine Intelligence (TPAMI), vol. 35, no. 1, pp. 171 -- 184, 2013.
2. Guangcan Liu, Zhouchen Lin, Yong Yu and XiaoouTang,Unsupervised Object Segmentation with A Hybrid Graph Model (HGMIEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), Pages: 910-924, Volume 32 , Issue 5, 2010.
3. Guangcan Liu and Shuicheng Yan. Active Subspace: Towards Scalable Low-Rank Learning, Neural Computation, 2012.
Talk Abstract:
Today's data-driven community is full of high-dimensional, massive and noisy datasets. This is quite different from several years ago, and thus it is critical to establish new tools for analyze the intrinsic structures of data. In this context, the most fundamental and rather old problem, the PCA problem, is new again. In this work, we consider the generalized problem of PCA, called GPCA, which was supposed to be more powerful than PCA. Unfortunately, in general case, the GPCA problem is very challenging as it concerns the chicken-and-egg coupling between subspace estimation and segmentation. Naturally, it would be much more challenging to resolve its advanced version, Robust GPCA. Because of these difficulties, the application scope of previous GPCA methods is actually much narrower than PCA. Notice that, in high-dimensional case, the sum of multiple low-rank subspaces is till low-rank. With this restriction, it will be shown that the GPCA and Robust GPCA problems are possible to be well-solved. We introduce a novel method, called Low-Rank Representation(LRR), which seeks the lowest rank representation among all the candidates that can represent the data samples as linear combinations of bases in a given dictionary. Experimentally, it is shown that LRR can achieve state-of-the-art performance in various applications, e.g, motion segmentation, image segmentation and saliency detection