JSM 2015 Preliminary Program

Online Program Home
My Program

Abstract Details

Activity Number: 258
Type: Contributed
Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
Sponsor: Section on Statistical Learning and Data Mining
Abstract #317440
Title: Blessing of Massive Scale: Spatial Graphical Model Inference with a Total Cardinality Constraint
Author(s): Ethan X. Fang* and Han Liu and Mengdi Wang
Companies: Princeton University and Princeton University and Princeton University
Keywords: Graphical Model ; $\ell_0$-constraint method ; fMRI data ; nonconvex combinatorial optimization
Abstract:

We propose a novel inferential framework for estimating and uncertainty assessment of large-scale spatial graphical models with a total cardinality constraint. This work has two major contributions. (i) From a computational perspective, we need to minimize the summation of massive amounts of localized loss functions coupled by a global cardinality constraint. Though this optimization problem is highly nonconvex, we propose a dual ascent algorithm which runs in an iterative fashion. Within each iteration, the algorithm divides the global graph estimation problem into many nodewise neighborhood pursuit problems which can be trivially solved by any parallel computing architectures (e.g., GPU or cloud). Between iterations, the algorithm only needs to update a scalar dual variable and is communication-efficient. Theoretically, we show that the computational accuracy (measure by averaged duality gap) increases when the problem scale (measured by the number of nodes) increases. Thus we see a bless of massive scale. We provide a convex geometry justification of this seemingly surprising phenomenon, along with a rigorous characterization of the diminishing rate of the duality gap.


Authors who are presenting talks have a * after their name.

Back to the full JSM 2015 program





For program information, contact the JSM Registration Department or phone (888) 231-3473.

For Professional Development information, contact the Education Department.

The views expressed here are those of the individual authors and not necessarily those of the JSM sponsors, their officers, or their staff.

2015 JSM Online Program Home