IC Colloquium : Leveraging inexactness for scalability
Event details
| Date | 15.04.2013 |
| Hour | 16:15 › 17:30 |
| Speaker |
Suvrit Sra, Max Planck Institute for Intelligent Systems IC faculty candidate |
| Location | |
| Category | Conferences - Seminars |
Abstract
Mankind generates data at a pace that vastly outstrips processing power, so much so that even "tractable" algorithms may be too slow. Linear, or better yet, sublinear complexity is the most we can afford. These limitations induce a paradigm shift in modern data analysis is approached: lower accuracy, randomization, online processing, distributed computation, inexact methods and inexact models gain the centerstage.
In my talk, I illustrate these views by culling a few examples from my own research. I mention not only progress on key components for large-scale optimization, but also two main items from recent work. The first is a new framework for inexact optimization, which subsumes a slew of known large-scale algorithms; it is also the first of its kind for scalably tackling nonconvex, nonsmooth problems that pervade data driven research. The second is an unexpected instance of inexactness: we replace a difficult convex model with an easier nonconvex one, which amazingly can still be solved to global optimality. I will also briefly mention broader connections of our model to areas within mathematics and information theory. To add perspective, I ground the whole talk in applications from computer vision, image processing, and machine learning that have driven my theoretical and algorithmic work.
Biography
Suvrit Sra is a Sr. Research Scientist at the Max Planck Institute for Intelligent Systems in Tübingen, Germany. He obtained Ph.D. in Computer Science from the University of Texas at Austin in 2007. In Spring 2013 he is visiting UC Berkeley and teaching a graduate course on convex optimization. His research focuses on "large-scale data analysis and optimization"; in particular, he designs, analyzes, and implements algorithms for data intensive problems in scientific computing, statistics, computer vision, and machine learning. His mathematical interests span a variety of areas including matrix analysis, noncommutative algebra, and geometry.
His research has won awards at several international venues; the most recent being the "SIAM Outstanding Paper Prize (2011)" for his work on metric nearness. He regularly organizes the Neural Information Processing Systems (NIPS) workshops on "Optimization for Machine Learning", and and has recently co-edited a book with the same title (MIT Press, 2011).
Mankind generates data at a pace that vastly outstrips processing power, so much so that even "tractable" algorithms may be too slow. Linear, or better yet, sublinear complexity is the most we can afford. These limitations induce a paradigm shift in modern data analysis is approached: lower accuracy, randomization, online processing, distributed computation, inexact methods and inexact models gain the centerstage.
In my talk, I illustrate these views by culling a few examples from my own research. I mention not only progress on key components for large-scale optimization, but also two main items from recent work. The first is a new framework for inexact optimization, which subsumes a slew of known large-scale algorithms; it is also the first of its kind for scalably tackling nonconvex, nonsmooth problems that pervade data driven research. The second is an unexpected instance of inexactness: we replace a difficult convex model with an easier nonconvex one, which amazingly can still be solved to global optimality. I will also briefly mention broader connections of our model to areas within mathematics and information theory. To add perspective, I ground the whole talk in applications from computer vision, image processing, and machine learning that have driven my theoretical and algorithmic work.
Biography
Suvrit Sra is a Sr. Research Scientist at the Max Planck Institute for Intelligent Systems in Tübingen, Germany. He obtained Ph.D. in Computer Science from the University of Texas at Austin in 2007. In Spring 2013 he is visiting UC Berkeley and teaching a graduate course on convex optimization. His research focuses on "large-scale data analysis and optimization"; in particular, he designs, analyzes, and implements algorithms for data intensive problems in scientific computing, statistics, computer vision, and machine learning. His mathematical interests span a variety of areas including matrix analysis, noncommutative algebra, and geometry.
His research has won awards at several international venues; the most recent being the "SIAM Outstanding Paper Prize (2011)" for his work on metric nearness. He regularly organizes the Neural Information Processing Systems (NIPS) workshops on "Optimization for Machine Learning", and and has recently co-edited a book with the same title (MIT Press, 2011).
Links
Practical information
- Informed public
- Free
- This event is internal
Contact
- Christine Moscioni