Mathematics Colloquia and Seminars
Multiresolution Matrix FactorizationPDE and Applied Math Seminar
|Speaker:||Risi Kondor, University of Chicago|
|Start time:||Fri, Oct 28 2016, 4:10PM|
The size of today's datasets dictates that machine learning algorithms compress or reduce their input data and/or make use of parallelism. Multiresolution Matrix Factorization (MMF) makes a connection between such computational strategies and some classical themes in Applied Mathematics, namely Multiresolution Analysis and Multigrid Methods. In particular, the similarity (kernel) matrices appearing in data often have multiresolution structure, which can be exploited both for learning and to facilitate computation.
MMF is an algorithm both for finding structure in large matrices (somewhat similar to HSS matrices), and constructing wavelet bases on graphs. I will highlight applications to matrix compression/sketching and graph based semi-supervised learning. I will also present our parallel MMF software library that allows the method to easily scale to sparse matrices with ~10^6 rows/columns.
The work presented in this talk is joint with my students Nedelina Teneva, Pramod Mudrakarta, Yi Ding and Vikas Garg.