Active Learning for Regression: Algorithms and Applications¶
Masashi Sugiyama
Associate Professor
Department of Computer Science, Tokyo Institute of Technology, Tokyo, Japan
Abstract :
Active learning is a challenging task in supervised learning: users are allows to choose input points (questions) to gather output values (answers) for better generalization. This has some analogy to human learning that asking good questions to the teacher boosts the speed of learning. Traditional active learning methods for regression assume that the model at hand is correctly specified. However, this assumption is rarely satisfied in practice and the traditional methods are not reliable without this assumption. In this talk, I introduce a recently proposed active learning method for regression. The method is shown to be valid also for misspecified models, while algorithmic simplicity is kept moderately. I also show successful real-world examples of active learning such as semi-conductor wafer alignment and robot control.
Bio:
Masashi Sugiyama received B.E., M.E., and Ph.D. degrees in Computer Science from Tokyo Institute of Technology, Japan in 1997, 1999, and 2001, respectively. In 2001, he was appointed as an Assistant Professor in the same institute and he has been an Associate Professor since 2003. His research interests include theory and application of machine learning, robot control, and optical measurement. From 2003 to 2004, he was an Alexander von Humboldt Research Fellow and stayed at Fraunhofer Institute FIRST.IDA, Berlin, Germany. In 2007, he received Faculty Award from IBM for his contribution to non-stationarity adaptation in machine learning. A part of the achievements was included in his co-editted book Dataset Shift in Machine Learning published from the MIT Press in 2009.