On the Convergence and Complexity of Nonconvex ADMM

June 12, 2018, 1:40 PM - 2:10 PM



Rutgers University

CoRE Building

96 Frelinghuysen Road

Piscataway, NJ 08854

Click here for map.

Shiqian Ma, University of California, Davis

The alternating direction method of multipliers (ADMM) has been successfully used in solving problems arising from different fields such as machine learning, image processing, statistics and so on. In this talk, we discuss several recent results on convergence behavior of ADMM for solving nonconvex problems. We consider two nonconvex models. The first model allows the objective function to be nonconvex and nonsmooth, but the constraints are convex. The second model allows the constraints to be Riemannian manifolds. For both models, we propose ADMM variants for solving them and analyze their iteration complexities for obtaining an $epsilon$-stationary solution. Numerical results on tensor robust PCA, maximum bisection problem and community detection problem are reported to demonstrate the efficiency of the proposed methods.


Slides     Video