Online Program Home
My Program

Abstract Details

Activity Number: 452 - Geometric Statistical and Computational Methods in Imaging
Type: Topic Contributed
Date/Time: Wednesday, July 31, 2019 : 8:30 AM to 10:20 AM
Sponsor: Section on Statistics in Imaging
Abstract #306838 Presentation
Title: Catalyst Acceleration for Non-Convex Optimization on Manifolds
Author(s): Lizhen Lin and Bayan Saparbayeva* and Michael Minyi Zhang and David Dunson
Companies: University of Notre Dame and University of Notre Dame and Princeton University and Duke University
Keywords: Optimization on manifolds; weakly convex functions; acceleration algorithms

In the following work, we propose an extension of a class of optimization algorithms to non-convex problems on manifold spaces. To accomplish this, we leverage ideas from recent work on the ``Catalyst'' algorithm for non convex optimization on Euclidean spaces H.Lin et al. (2017). This method is initially designed to accelerate existing optimization algorithms for convex objective functions but is also amenable to non-convex functions. We apply the Catalyst acceleration scheme to many existing optimization routines, such as gradient descent, Newtonian methods and algorithms that are designed for distributed parallel inference on manifolds such as Sarpabayeva et al. (2018). One can show that the algorithm generally converges to the stationary point, and in the case of strongly convex functions, it converges to the local minimum. We will provide convergence analysis of the algorithms and present applications for a large class of objective functions in various manifold spaces.

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

Back to the full JSM 2019 program