[4eyes] FW: [Seminar]: REMINDER, TODAY 3:30pm: "Nearly Optimal Robust Subspace Tracking, " Namrata Vaswani, Professor, ECE, Iowa State University

[Matthew Turk]

This is an ECE talk today at 3:30pm that some may be interested in.
 
From: Melinda Sherwood <cao-admin at ece.ucsb.edu> 
Sent: Tuesday, February 19, 2019 10:32 AM
To: seminar at ece.ucsb.edu
Subject: [Seminar]: REMINDER, TODAY 3:30pm: "Nearly Optimal Robust Subspace Tracking," Namrata Vaswani, Professor, ECE, Iowa State University
 
Nearly Optimal Robust Subspace Tracking 
Namrata Vaswani, Professor
ECE, Iowa State University
Feb 19 (Tue) 3:30pm
Harold Frank Hall (HFH), Rm 4164
ABSTRACT:
Principal Components Analysis (PCA), a.k.a. subspace learning, is one of the most widely used dimension reduction techniques. It tries to find a low-dimensional subspace approximation of a given dataset. PCA is a solved problem when the observed data is relatively clean. Modern datasets are often corrupted by outliers though. Robust PCA refers to this harder problem of PCA in the presence of entry-wise outliers (sparse corruptions). An important example application for robust PCA is video analytics when slow-changing videos are corrupted by occlusions, e.g., by moving vehicles or persons. For long data sequences, e.g., long surveillance videos, if one tries to use a single subspace to represent the entire sequence, the required subspace dimension may be too large. For such data, a better model is to assume that the data subspace changes with time, albeit gradually. This problem of tracking data lying in a changing subspace, while being robust to additive sparse outliers is referred to as Robust Subspace Tracking (RST). While robust PCA has received a lot of attention in the last decade, its dynamic version was largely open until recently. In a recent body of work, we have introduced the first provably correct and practically usable online solution framework for RST that we call Recursive Projected Compressive Sensing (ReProCS). Our most recent work from ICML 2018 shows that a simple ReProCS-based algorithm provides a provably fast and nearly (delay and memory) optimal RST solution under mild assumptions: weakened standard robust PCA assumptions and subspace change that is slow enough compared to the smallest magnitude outlier entry. Our theoretical claims are also backed by extensive experimental evidence for two video applications.

We will end the talk with describing our new work on subspace tracking and signal sequence recovery from phaseless (magnitude-only) linear projections of each signal. This finds applications in Fourier ptychographic imaging of dynamic (time-varying) scenes.
 
BIO:
Namrata Vaswani is a Professor of Electrical and Computer Engineering at Iowa State University. She received the Ph.D. in 2004 from the University of Maryland, College Park and a B.Tech. in 1999 from IIT-Delhi. Her research interests are in data science and include statistical machine learning, signal processing, and computer vision. She received the 2014 Iowa State Early Career Engineering Faculty Research Award as well as the 2014 IEEE Signal Processing Society Best Paper Award (for work co-authored with her graduate student Lu in the IEEE Transactions on Signal Processing). Vaswani is a Fellow of the IEEE for contributions to dynamic high-dimensional structured data recovery.
 
Hosted by: Professor B.S. Manjunath
 
 
    
 


 
-- 
Melinda Sherwood
Administrative Office Coordinator, ECE
University of California, HFH Room 4155
Santa Barbara, CA 93106
Phone: (805) 893-5364  
 
     





-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cs.ucsb.edu/pipermail/ilab-users/attachments/20190219/a212d431/attachment-0001.html>