"Learning From Compressed Meaurements of High Dimensional Data"

Location: LTS Auditorium, 8080 Greenmead Drive

Speaker:
Piya Pal
UMD Department of Electrical and Computer Engineering and Institute for Systems Research

Abstract:
Modern signal processing is centered around processing complex high dimensional data which can arise in applications ranging from data collected by sensor arrays to information generated by millions of users in social networks. While classical signal processing techniques often rely upon model-based approaches to signal representation, such techniques lose their applicability in modern data processing tasks.

In this talk, I will focus on two learning tasks. In the first part of the talk, I will consider a linear Bayesian observation model where the hyperparameters of interest exhibit sparsity. Such a model arises in many applications such as imaging, source localization etc.

I will also consider the problem of dictionary learning from compressed measurements in block sparse observation models. It will be shown that under certain practical assumptions on the unknown sparse code, it is possible to uniquely identify the dictionary atoms from far fewer samples than existing literature suggests.

Bio:
Piya Pal is an assistant professor in the UMD Department of Electrical and Computer Engineering and the Institute for Systems Research.

She received her doctorate in electrical engineering from Caltech in 2013, where her thesis won the Charles and Ellen Wilts Prize for Outstanding Doctoral Thesis in Electrical Engineering.

During the course of receiving her doctorate, she developed novel techniques for non-uniform sparse sampling, which received Best Student Paper Awards at the 14th IEEE DSP Workshop and the 45th Asilomar Conference on Signals, Systems and Computers in 2011.

Her research interests span compressive and structured sampling, statistical signal processing with applications in radar and array processing, tensor methods, and statistical learning.