Once every week while school is in session, EECS invites a distinguished researcher or practitioner in a computer science or electrical and computer engineering-related field to present their ideas and/or work. Talks are generally targeted to electrical engineering and computer science graduate students. This colloquium series is free and open to everyone.

Upcoming Colloquia

Multi-view Representation Learning with Canonical Correlation Analysis

Monday, October 8, 2018 - 4:00pm to 4:50pm
LINC 200

Speaker Information

Weiran Wang
Amazon

Abstract

Canonical correlation analysis (CCA) has been the main workhorse for multi-view feature learning, where we have access to multiple ''views'' of data at training time while only one primary view is available at test time. The idea of CCA is to project the views to a common space such that the projections of different views are maximally correlated.

In the first part of the talk, we compare different nonlinear extensions of CCA, and find that the deep neural network extension of CCA, termed deep CCA (DCCA), has consistently good performance while being computationally efficient for large datasets. We further compare DCCA with deep autoencoder-based approaches, as well as new variants. We find an advantage for correlation-based representation learning.
In the second part of the talk, we study the stochastic optimization of canonical correlation analysis, whose objective is nonconvex and does not decouple over training samples. Although several stochastic optimization algorithms have been previously proposed to solve this problem, no global convergence guarantee was provided by any of them. Based on the alternating least squares formulation of CCA, we propose a globally convergent stochastic algorithm, which solves the resulting least squares problems approximately to sufficient accuracy with state-of-the-art stochastic gradient methods. We provide the overall time complexity of our algorithm which improves upon that of previous work.

This talk summarizes primarily my postdoc research at TTI-Chicago, and I will give pointers to more recent development. The talk includes joint work with Raman Arora (JHU), Jeff Bilmes (UW), Jialei Wang (U Chicago), Dan Garber (Technion), Nathan Srebro (TTIC), and Karen Livescu (TTIC).

Speaker Bio


Weiran Wang is currently an applied scientist at Amazon Alexa, trying to build intelligent personal assistant. From 2014 to 2017, he was a postdoctoral scholar at Toyota Technological Institute at Chicago. He obtained his PhD degree from the EECS Department at University of California, Merced in 2013. His research includes algorithms for deep learning, multi-view representation learning, sequence prediction, manifold learning, optimization for machine learning, and applications to speech and audio processing.

STRUCTURAL SUBBAND DECOMPOSITION: A NEW CONCEPT IN DIGITAL SIGNAL PROCESSING

Monday, October 15, 2018 - 4:00pm to 4:50pm
LINC 200

Speaker Information

Sanjit K. Mitra
University of California, Santa Barbara, California
University of Southern California, Los Angeles, California

Abstract

Polyphase decomposition of a sequence was advanced to develop computationally efficient interpolators and decimators, and has also been used to design computationally efficient quadrature-mirror filter banks.  The polyphase decomposition represents a sequence into a set of sub-sequences, called polyphase components.  However, the polyphase components do not exhibit any spectral separation.  In this talk, we first review the concept of structural subband decomposition, a generalization of the polyphase decomposition, which decomposes a sequence into a set of sub-sequences with some spectral separation that can be exploited advantageously in many digital signal processing applications. We then outline some of the applications of the structural subband decomposition, such as, efficient design and implementation of FIR digital filters, development of computationally efficient decimators and interpolators, subband adaptive filtering, and fast computation of discrete transforms.

Speaker Bio

Sanjit K. Mitra is a Research Professor in the Department of Electrical & Computer Engineering, University of California, Santa Barbara and Professor Emeritus, Ming Hsieh Department of Electrical Engineering, University of Southern California, Los Angeles.

Dr. Mitra has served IEEE in various capacities including service as the President of the IEEE Circuits & Systems Society in 1986. He is a member of the U.S. National Academy of Engineering, a member of the Norwegian Academy of Technological Sciences, an Academician of the Academy of Finland, a foreign member of the Finnish Academy of Sciences and Arts, a foreign member of the Croatian Academy of Sciences and Arts, Croatian Academy of Engineering, and the Academy of Engineering, Mexico, and a Foreign Fellow of the National Academy of Sciences, India and the Indian National Academy of Engineering. Dr. Mitra is a Life Fellow of the IEEE.

Reinforcement Learning for Healthcare

Monday, October 22, 2018 - 4:00pm to 4:50pm
LINC 200

Speaker Information

Finale Doshi-Velez
Assistant Professor
Computer Science
Harvard Paulson School of Engineering and Applied Sciences

Abstract

Many healthcare problems require thinking not only about the immediate effect of a treatment, but possible long-term ramifications.  For example, a certain drug cocktail may cause an immediate drop in viral load in HIV, but also cause the presence of resistance mutations that will reduce the number of viable treatment options in the future.  Within machine learning, the reinforcement learning framework is designed to think about decision-making under uncertainty when decisions may have long-lasting effects. 

In this talk, I will talk about a number of directions we are developing in my lab to identify personalized treatment policies from electronic health and registry records.  Our approaches achieve state-of-the-art results on HIV management and initial promising results for sepsis management.  Next, I'll dive into how we evaluated these algorithms when we could not test on new patients and had to rely only on the observational data -- highlighting both current work in our lab on off-policy evaluation as well as more general gotchas that remind us to all be careful scientists.
 
This is joint work with: Sonali Parbhoo, Maurizio Zazzi, Volker Roth, Xuefeng Peng, David Wihl, Yi Ding, Omer Gottesman, Liwei Lehman, Matthieu Komorowski, Aldo Faisal, David Sontag, Fredrik Johansson, Leo Celi, Aniruddh Raghu, Yao Liu, Emma Brunskill, and the CS282 2017 Topics in Machine Learning Course.

Speaker Bio

Finale Doshi-Velez is an Assistant Professor in Computer Science at the Harvard Paulson School of Engineering and Applied Sciences.  She completed MSc at the University of Cambridge as a Marshall Scholar, her PhD at MIT, and her postdoc at Harvard Medical School.  She has won an NSF CAREER Award, is a Sloan fellow, and co-organizes the Machine Learning for Healthcare Conference.

Past Colloquia

Hans-Benjamin Braun
Monday, October 20, 2014 - 1:00pm to 2:00pm
Anurag K Srivastava
Thursday, August 14, 2014 - 2:00pm to 3:00pm
Hermann Schumacher
Wednesday, July 30, 2014 - 2:00pm to 3:00pm
Ron Jansen
Monday, June 2, 2014 - 3:00pm to 4:00pm
Paul Bennett
Monday, May 19, 2014 - 4:00pm to 4:50pm
Xiaojin (Jerry) Zhu
Monday, May 12, 2014 - 4:00pm to 4:50pm
Nader Bagherzadeh
Monday, May 5, 2014 - 4:00pm to 4:50pm

Pages