报告题目: Kernel-based Unsupervised Learning on Large Data Sets（大规模数据上基于核的非监督学习） 时间：12月12日 19:00 地点：蒙民伟楼109报告厅 个人简介： 郭天佑教授于 1988年在香港大学电机与电子工程系获学士学位，1996 年在香港科技大学计算机系获博士学位。1996-1997年在美国新泽西州朗讯—贝尔实验室从事研究工作任顾问，1997年起在香港浸会大学计算机系任助理教授，2000年回到香港科技大学计算机系，现任计算机科学与工程系副教授。郭天佑教授主要从事机器学习、模式识别和神经网络等领域的研究工作，已在顶级国际刊物以及顶级国际会议发表论文90余篇，曾获《IEEE Transactions on Neural Networks》的2004年度杰出论文奖。郭教授发表的论文已被国际同行引用900余次，尤其是在核方法（Kernel Method）的研究方面，他的工作在国际上有重要影响。郭教授目前担任著名国际刊物《IEEE Transactions on Neural Networks》和《Neurocomputing》等的编委，多次担任国际会议的程序委员会副主席或领域主席，并担任包括顶级国际会议ICML、ECML等在内的20余次国际会议程序委员。
Unsupervised learning has long been an active and important research area in machine learning. However, with the increasingly widespread use of information services such as the Internet, this has led to the so-called information explosion problem and unsupervised learning methods typically cannot handle these large data sets. This talk focuses on two unsupervised learning methods that have been commonly used by the kernel community. The first one is the Nystrom method, which is a sampling-based technique for approximating the eigen-systems of large kernel matrices. Motivated from the integral equation defining the kernel eigenfunctions, we extend the Nystrom formulation to a more general, density-weighted version. An efficient algorithm is proposed to enforce the weights. Experiments on kernel principal component analysis, spectral clustering and image segmentation all demonstrate significant improvements in both accuracy and scalability. The second part covers a recent large margin unsupervised learning approach called maximum margin clustering (MMC). Computationally, it involves non-convex optimization and has to be relaxed to different semidefinite programs (SDP). However, SDPs are computationally very expensive. To make MMC more practical, we propose an efficient approach that performs alternating optimization directly. Experiments demonstrate that the proposed approach is often more accurate, much faster and can handle much larger data sets. |
| 今天是： |
