3/9/2016- Stefanie Jegelka (MIT)- Algorithms and new applications for determinantal point processes

Presentation Date: 

Wednesday, March 9, 2016

Title: Algorithms and new applications for determinantal point processes

Abstract:  Many real-world inference problems are, at their core, subset selection problems. Probabilistic models for such scenarios rely on having distributions over discrete sets that are sufficiently accurate yet computationally efficient to work with. We focus on sub-families of such distributions whose special mathematical properties are the basis for fast algorithms. As a specific example, Determinantal Point Processes (DPPs) have recently become popular in machine learning, as elegant and tractable probabilistic models of diversity. We explore new applications of DPPs for variational inference over combinatorial objects, such as coupled cascades in a collection of networks, where we are able to leverage combinatorial and convex structure in the problem. 

In the second part of the talk, I will outline ideas for speeding up sampling from DPPs. These ideas build on new insights for algorithms that compute bilinear inverse forms. These results have applications beyond DPPs, including sensing with Gaussian Processes and submodular maximization.

This is joint work with Chengtao Li, Josip Djolonga, Suvrit Sra and Andreas Krause.

See also: 2016