Gaussian Processes for Transfer Learning
Friday, April 6, 2007
Talk: 11:00 am - 12:00 pm
Host: Rong Jin
The phenomenon of knowledge sharing between different but related learning tasks is highly common in real-world applications. Transfer learning aims to harness the shared dependency structure across many related learning tasks, and thus transfer valuable knowledge from one task to another. In this talk I will briefly review the recent research achievements in this direction, and then introduce our recent work on transfer learning using Gaussian processes. In particular, a common view to many different algorithms will be provided, to outline their connections as well as differences. The Gaussian process can be seen as a nonparametric generalization of many existing parametric models, and achieves a higher flexibility to represent the knowedge shared between tasks. In the end I will show some empirical results on collaborative filtering and sensor network analysis, and summarize some personal thoughts on future research directions.
Dr. Kai Yu is a research staff member at NEC Labs America. He received his PhD degree in computer science at University of Munich, Germany, and worked at Siemens as a senior research scientist before joining NEC Labs. His research has been focused on probabilistic modeling, Bayesian inference, general machine learning algorithms, and their applications to information retrieval, text mining, clinical data analysis, and intrusion detection. In NEC labs his current focus is transfer learning for human activity recognition and natural language processing. For more information about his research work, please visit: http://www.dbs.informatik.uni-muenchen.de/~yu_k/