More about HKUST
Accurate Max-margin Training for Structured Output Spaces
Speaker: Dr. Sunita SARAWAGI Department of Computer Science and Engineering Indian Institute of Technology Bombay Title: "Accurate Max-margin Training for Structured Output Spaces" Date: Tuesday, 27 May 2008 Time: 10:30am - 11:30am Venue: Lecture Theatre H (Chen Kuan Cheng Forum, near lift nos. 27/28) HKUST Abstract: In this talk, I will present new insights and objectives for max-margin training of structured models. There are two popular formulations for maximum margin training of structured spaces: margin scaling and slack scaling. While margin scaling has been extensively used since it requires the same kind of MAP inference as normal structured prediction, slack scaling is believed to be more accurate and better-behaved. I will describe an efficient variational approximation to the slack scaling method that solves its inference bottleneck while retaining its accuracy advantage over margin scaling. Further I argue that existing scaling approaches do not separate the true labeling comprehensively while generating violating constraints. I will propose a new max-margin trainer PosLearn that generates violators to ensure separation at each position of a decomposable loss function. ****************** Biography: Sunita Sarawagi researches in the fields of databases, data mining, machine learning and statistics. Her current research interests are information integration, graphical models, probabilistic databases, and domain adaptation. She is associate professor at IIT Bombay. Prior to that she was a research staff member at IBM Almaden Research Center. She got her PhD in databases from the University of California at Berkeley and a bachelor's degree from IIT Kharagpur. She has several publications in databases and data mining including a best paper award at the 1998 ACM SIGMOD conference and several patents. She is on the editorial board of the ACM TODS and ACM TKDD journals and ex-editor-in-chief of the ACM SIGKDD newsletter. She is program chair for the ACM SIGKDD 2008 conference and has served as program committee member for SIGMOD, VLDB, SIGKDD, ICDE, and ICML conferences.