Tony Jebara, PhD

Associate Professor, Computer Science, Columbia University

Director, Columbia Machine Learning Laboratory

Recording Not Available

“Probabilistic Graphical Models and Online Learning”

Abstract:  Probabilistic graphical models are fundamental tools in machine learning and have numerous applications. They entail three canonical problems. 1) Maximum a posteriori inference finds the most likely configuration under a probability distribution. 2) Marginal inference computes the probability of a subset of variables in a given model. 3) Learning finds model parameters and is really just iterative marginal inference. All three canonical problems are NP-hard for graphical models with cycles or large tree-width.
For MAP inference, we give polynomial time guarantees by reducing the task to the maximum weight stable set problem in combinatorics. For marginal inference we improve heuristics like loopy belief propagation through polynomial time exact discrete optimization and double-cover methods. For learning, we combine Bethe approximation with a Frank-Wolfe algorithm in the convex dual which circumvents the intractable partition function.
I will show applications in friendship link recommendation, social influence estimation, computer vision, financial networks and power networks. I will also show Netflix applications in online learning where the data arrives sequentially. For instance, how to choose the best image for every movie/show for each user.

Bio: Tony Jebara heads machine learning at Netflix and is Associate Professor at Columbia University (on leave).  He has published over 100 peer-reviewed papers in conferences and journals like NIPS, ICML, UAI, COLT, JMLR, CVPR, ICCV, and AISTAT. His research has been recognized with multiple best paper awards and an NSF Career Award. His work has been featured on television (ABC, BBC, New York One, TechTV, etc.) and in the popular press (New York Times, Slash Dot, Wired, Businessweek, IEEE Spectrum, Esquire, etc.).  Jebara has served at the International Conference on Machine Learning as Program Chair in 2014 and as General Chair in 2017. He obtained his PhD in 2002 from MIT.

For more information on MIDAS or the Seminar Series, please contact midas-contact@umich.edu. MIDAS gratefully acknowledges Northrop Grumman Corporation for its generous support of the MIDAS Seminar Series.