STATISTICS COLLOQUIUM
Garvesh Raskutti
Associate Professor
Department of Statistics
University of Wisconsin-Madison
Sketching meets tensor estimation: Recent developments in large-scale low-rank problems
Abstract
Tensors or higher-order arrays are playing an ever-increasing role in data science and algorithms, both because tensor representation is common in many datasets and because tensor encodings of data frequently arise in many algorithms. Developing computationally efficient and statistically reliable algorithms for tensors remains a significant open challenge due to the unique challenges that tensors present. In this talk, I discuss some of my work on algorithms for large-scale high-dimensional tensor estimation. I focus on a recent method which exploits low-rank structure using the idea of importance sketching to yield both efficient algorithms and optimal statistical performance. I demonstrate through theory, simulation studies and a real-data example that our method can tackle problems far beyond the scale of existing methods and lead to statistically reliable solutions.
Bio: Dr. Garvesh Raskutti is an Associate Professor at the University of Wisconsin-Madison in the Department of Statistics. He is also an affiliate for the Departments of Computer Science, Electrical and Computer Engineering and the Wisconsin Institute of Discovery Optimization Group. Prior to starting at UW, he completed a Masters of Engineering at the University of Melbourne in 2008 under the joint supervision of Rodney S. Tucker and Kerry Hinton, and a PhD at UC Berkeley in 2012 under the joint supervision of Martin Wainwright and Bin Yu. His research interests include statistical machine learning, optimization, graphical and network modeling and information theory with applications to systems biology and neuroscience. In particular, his research broadly focuses on challenges that arise in large-scale statistical inference problems. These challenges are especially motivated by problems in systems biology and neuroscience. Existing methods typically suffer from both statistical and computational limitations. Dr. Raskutti focuses on developing methods that address these statistical and computational challenges which leads to: (1) novel algorithms and theoretical insights; and (2) potentially novel insights in both systems biology and neuroscience.
For more information, contact: Tracy Burke at tracy.burke@uconn.edu