October 26, 2021

12:00 pm / 1:00 pm

Venue

https://wse.zoom.us/j/99567504456?pwd=WkI2UlpGT3p6MldLS05VNkdmcGxiZz09Held Virtually; Zoom Link TBA

Recorded seminar:

https://wse.zoom.us/rec/play/vqRsxNkUgTbu7K7rjCW3HNGmjDrfFgrPEdwbZGRbX_WmGEo02PT_axGcWLmdrppihNLqlL-BfM8fqVAb.Qc4tCWgPvEIXqVRw?autoplay=true&startTime=1635263949000

?Finding low-dimensional structure in messydata?

 Laura Balzano, PhD

AssociateProfessor

Electrical Engineering and Computer Science

Universityof Michigan

 Abstract: Inorder to draw inferences from large, high-dimensional datasets, we often seeksimple structure that models the phenomena represented in those data. Low-ranklinear structure is one of the most flexibleand efficient such models,allowing efficient prediction, inference, andanomaly detection. However,classical techniques for learning low-rank models assume your data have onlyminor corruptions that are uniform over samples. Modern research inoptimization has begun to develop new techniques to handle realistic messy data? where data are missing, have wide variations in quality, and/or are observedthrough nonlinear measurement systems.

In this talk I will give a high-level overview of recent research in this area.Then I will focus on the problem of learning linear subspace structurefrommultiple data sources of varying quality. This is common in problems likesensor networks or medical imaging, where different measurements of thesamephenomenon are taken with different quality sensing (eg high or low radiation).In this context, learning the low-rank structure via PCA suffers from treatingall data samples as if they are equally informative. I will discuss ourtheoretical results on weighted PCA. I will then present new algorithms for thenon-convex probabilistic PCA formulation of this problem and anovel SDPrelaxation.

Biography: LauraBalzano is an associateprofessor of Electrical Engineering and ComputerScience, and of Statistics by courtesy, at the University of Michigan. She isrecipient of the NSF Career Award, ARO Young Investigator Award, AFOSR YoungInvestigator Award, and faculty fellowships from Intel and 3M. She received theVulcans Education Excellence Award at the University of Michigan. Her mainresearch focus is on modeling with big, messy data ? highly incomplete orcorrupted data,uncalibrated data, and heterogeneous data ? and itsapplications in a widerange of scientific problems. Her expertise is instatistical signal processing, matrix factorization, and optimization. Laura receiveda BS from Rice University, MS from the UCLA, and PhD from the University ofWisconsin in Electrical and Computer Engineering.