Distributed Accelerated Proximal Coordinate Gradient Methods

Yong Ren,Jun Zhu
DOI: https://doi.org/10.24963/ijcai.2017/370
2017-01-01
Abstract:We develop a general accelerated proximal coordinate descent algorithm in distributed settings (Dis-APCG) for the optimization problem that minimizes the sum of two convex functions: the first part f is smooth with a gradient oracle, and the other one Ψ is separable with respect to blocks of coordinate and has a simple known structure (e.g., L 1 norm). Our algorithm gets new accelerated convergence rate in the case that f is strongly convex by making use of modern parallel structures, and includes previous non-strongly case as a special case. We further present efficient implementations to avoid full-dimensional operations in each step, significantly reducing the computation cost. Experiments on the regularized empirical risk minimization problem demonstrate the effectiveness of our algorithm and match our theoretical findings.
What problem does this paper attempt to address?