Negative Dependence and Sampling in Machine Learning

Seminar | September 27 | 4-5 p.m. | 1011 Evans Hall

 Stefanie Jegelka, Massachusetts Institute of Technology

 Department of Statistics

Discrete Probability distributions with strong negative dependencies (negative association) occur in a wide range of settings in Machine Learning, from probabilistic modeling to randomized algorithms for accelerating a variety of popular ML models. In addition, these distributions enjoy rich theoretical connections and properties. A prominent example are Determinantal Point Processes.
In this talk, I will survey recent applications and developments, and in particular efficient, fast-mixing Markov Chains for sampling. The sampling results exploit connections with linear algebra and a specific use of classic quadrature, and, importantly, close connections with matroid theory and the theory of real stable polynomials. The resulting algorithms have theoretical convergence guarantees and are easily applicable in practice too.

This talk is based on joint work with Chengtao Li, Zelda Mariet and Suvrit Sra.