Nov 9, 2006: Learning with Gaussian Processes
Filed in: Colloquium
Dr. Wei Chu, Columbia University
On basis of our prior knowledge and observations, Bayesian learning is capable of adapting to the characteristics of realistic data, judging the uncertainty of predictions and discovering the structural knowledge. Gaussian processes provide a Bayesian framework of flexible non-parametric models. In this talk, I will start with Bayesian linear models to introduce Gaussian processes, and then focus on relational learning with Gaussian processes. Although correlation between instances is often modeled via a kernel function using input attributes of the instances, relational knowledge, as a typical data resource in many Bioinformatics and data mining problems, can further reveal additional pairwise correlations between variables of interest. We develop a class of models which incorporates both reciprocal relational information and input attributes. This approach provides a novel data-dependent kernel function for supervised learning tasks. We also apply this framework to semi-supervised learning and directed linkage prediction, and discuss the potential of this approach.
Bio
Wei Chu received the B.E. degree in automatic control from Harbin Engineering University, China, in 1995, the M.E. degree in inertial navigation systems from the 3rd Research Academy of the China Aerospace Cooperation, in 1998, and the Ph.D. degree from the Control Division of the Department of Mechanical Engineering, National University of Singapore, in 2003. From February 2003 to January 2006, he was a research fellow at the Gatsby Computational Neuroscience Unit, University College London, UK. He is currently an associated research scientist at the Center for Computational Learning Systems, Columbia University, USA. His research interests include machine learning and bioinformatics.
Abstract
On basis of our prior knowledge and observations, Bayesian learning is capable of adapting to the characteristics of realistic data, judging the uncertainty of predictions and discovering the structural knowledge. Gaussian processes provide a Bayesian framework of flexible non-parametric models. In this talk, I will start with Bayesian linear models to introduce Gaussian processes, and then focus on relational learning with Gaussian processes. Although correlation between instances is often modeled via a kernel function using input attributes of the instances, relational knowledge, as a typical data resource in many Bioinformatics and data mining problems, can further reveal additional pairwise correlations between variables of interest. We develop a class of models which incorporates both reciprocal relational information and input attributes. This approach provides a novel data-dependent kernel function for supervised learning tasks. We also apply this framework to semi-supervised learning and directed linkage prediction, and discuss the potential of this approach.
Bio
Wei Chu received the B.E. degree in automatic control from Harbin Engineering University, China, in 1995, the M.E. degree in inertial navigation systems from the 3rd Research Academy of the China Aerospace Cooperation, in 1998, and the Ph.D. degree from the Control Division of the Department of Mechanical Engineering, National University of Singapore, in 2003. From February 2003 to January 2006, he was a research fellow at the Gatsby Computational Neuroscience Unit, University College London, UK. He is currently an associated research scientist at the Center for Computational Learning Systems, Columbia University, USA. His research interests include machine learning and bioinformatics.