Event Information

print  
Title: Roni Khardon, Tufts University
Sharing: Public
Start Time: Friday October 20, 2017 10:00 AM
End Time: Friday October 20, 2017 11:00 AM
Location: Indiana Memorial Union (Union Building)  
Contact: Predrag Radivojac
Url: http://cs-colloq.soic.indiana.edu/
Free/Busy: busy
Description:

The School of Informatics, Computing, and Engineering (SICE) CS Colloquium Distinguished Speaker Series

Speaker:   Roni Khardon, Tufts University

Where:   State Room East, IMU

When:  10 am, Friday, October 20, 2017

Topic:  Probabilistic Models in Machine Learning: Theory, Algorithms and Applications

Abstract:  The talk will give an overview of my work on Bayesian graphical models.  I will start by reviewing work on applications in astrophysics, analyzing luminosity data from light sources in the sky. This work motivated the development of the Bayesian Grouped Multi-Task (GMT) model. GMT captures multiple related prediction problems, where each task is formed from group-related characteristics and individual characteristics, a natural form that fits many problems in science, medicine and engineering.

Solving the machine learning problem in GMT requires novel scalable inference algorithms in probabilistic models. Our algorithmic work has focused on a general family of models which includes Gaussian processes, matrix factorization, topic models and more. I will briefly describe our algorithms, that use fixed-point updates and their stochastic variants to solve the variational inference approximation, and which significantly improve the convergence speed over competing approaches.

The final part of the talk will address a fundamental theoretical question for Bayesian models. When should we prefer a Bayesian procedure over a simpler one? especially in cases where the model itself may be mis-specified and where the algorithms are limited to approximate inference? Focusing on the same family of models, our recent results show that the expected error of certain variational approximations can be quantitatively bounded relative to the error of the best meter estimate in the same model. This gives a strong justification for the use of such methods.

Biography:  Roni Khardon is a professor in the Department of Computer Science at Tufts University. He holds a Ph.D. in Computer Science from Harvard University, and M.Sc. and B.Sc. degrees from the Technion. Prior to moving to Tufts he held a faculty position at the University of Edinburgh. Khardon's interests are in Machine Learning and Data Mining, Artificial Intelligence, and Efficient Algorithms and his research explores theoretical questions, empirical questions, and applications.  He is serving as an associate editor for the Machine Learning Journal and the Artificial Intelligence Journal, and he has served as an associate editor for the Journal of Artificial Intelligence Research during 2011-2017. He regularly serves on program committees for leading conferences in AI and machine learning.

Poster

Below is the link to the webinar for Roni Khardon’s talk.   

https://iu.zoom.us/j/808476513

 

Contact Email: predrag@indiana.edu
Share this Event:
I'm going copy to my calendar back