CS Events

Computer Science Department Colloquium

Lift-and-Project for Statistical Machine Learning Models

 

Download as iCal file

Thursday, April 20, 2023, 10:30am - 11:30am
 

Speaker: Periklis A. Papakonstantinou

Bio

My research interests are in theoretical computer science, cryptography and data privacy, and machine learning theory. I am also interested in understanding the interfaces between these areas.

I have been with the faculty of the Department of Management Science as an assistant and associate professor since September 2015. Since then I have cosupervised Hafiz Asif, a Ph.D. student in my department and now an assistant professor at Hofstra University. Currently, I am also the advisor of Nathaniel Hobbs whose expected graduation from the Ph.D. program is in August 2023. Hafiz did his Ph.D. in the theoretical foundations of Data Privacy. Nathaniel is doing his Ph.D. in problems in the intersection of Machine Learning and Cryptography, in particular in obfuscating and interpreting deep networks. Before Rutgers, I was (February 2010 - July 2015) an assistant professor at Andrew Yao's Institute, where four Ph.D. students graduated under my direct supervision (I was habilitated/Ph.D. supervisor of the duration of my appointment at Tsinghua). Bangsheng Tang did his Ph.D. with me in proof complexity and is now with Facebook Research, Hao Song did his PhD with me in communication complexity and is now an engineer at Pony.AI, Guang Yang did his PhD with me in cryptography and is an assistant professor in the Chinese Academy of Sciences (Institute of Computing Technology), and Shiteng Chen did his Ph.D. with me in circuit complexity and is now an associate professor in the Chinese Academy of Sciences (Institute of Software). I also supervised numerous diploma and MSc theses. These students continued their PhDs in Computer Science at Princeton, Harvard, and CMU and are now postdoctoral fellows, research assistant professors, and assistant professors at CMU, UPenn, and elsewhere.

Location : Core 301

: 

Event Type: Computer Science Department Colloquium

Abstract: In supervised learning, the prediction accuracy is critically bounded by learning errors. We introduce Lift-and-Project (LnP), a meta algorithm for probabilistic models that boosts multi-class classification accuracy. Unlike previous learning error reduction methods, LnP maps each class into a number of new classes and learns new class distributions "lifted" to a higher dimension. Specifically, instead of estimating the probability of a class c given an instance x, we estimate the probability of (c,c') given x, where (c,c') indicates that c is more likely to be the correct label for x than c', and c' encodes errors of the standard model. By marginalizing the new distributions for c, we "project" the lifted model back to the form of the original problem. We prove that in principle our method reduces the learning error exponentially. Experiments demonstrate significant improvements in prediction accuracy on standard datasets for discriminative and generative models.

: 

Contact  Eric Allender

Join Zoom Meeting
https://rutgers.zoom.us/j/96589562877?pwd=bENIS1NWbW91M0hmVE42ZGl6L3dOZz09

Join by SIP


Meeting ID: 965 8956 2877
Password: 962559
One tap mobile
+13126266799,,96589562877# US (Chicago)
+16465588656,,96589562877# US (New York)