Decentralised convex optimisation with probability-proportional-to-size quantization

Communication is one of the bottlenecks of distributed optimisation and learning. To overcome this bottleneck, we propose a novel quantization method that transforms a vector into a sample of components' indices drawn from a categorical distribution with probabilities proportional to values at...

Full description

Saved in:
Bibliographic Details
Main Authors: D.A. Pasechnyuk, P. Dvurechensky, C.A. Uribe, A.V. Gasnikov
Format: Article
Language:English
Published: Elsevier 2025-01-01
Series:EURO Journal on Computational Optimization
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2192440625000103
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Communication is one of the bottlenecks of distributed optimisation and learning. To overcome this bottleneck, we propose a novel quantization method that transforms a vector into a sample of components' indices drawn from a categorical distribution with probabilities proportional to values at those components. Then, we propose a primal and a primal-dual accelerated stochastic gradient methods that use our proposed quantization, and derive their convergence rates in terms of probabilities of large deviations. We focus on affine-constrained convex optimisation and its application to decentralised distributed optimisation problems. To illustrate the work of our algorithm, we apply it to the decentralised computation of semi-discrete entropy regularized Wasserstein barycentre's.
ISSN:2192-4406