Optimal Algorithms for Decentralized Stochastic Variational Inequalities

Authors: Dmitry Kovalev, Aleksandr Beznosikov, Abdurakhmon Sadiev, Michael Persiianov, Peter Richtarik, Alexander Gasnikov

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments on bilinear problems and robust regression problems confirm the practical efficiency of our methods, both in the non-distributed stochastic setup and in the decentralized deterministic one.
Researcher Affiliation Collaboration Dmitry Kovalev KAUST , Saudi Arabia dakovalev1@gmail.com Aleksandr Beznosikov MIPT , HSE University and Yandex, Russia anbeznosikov@gmail.com Abdurakhmon Sadiev MIPT, Russia sadiev.aa@phystech.edu Michael Persiianov MIPT, Russia persiianov.mi@phystech.edu Peter Richtárik KAUST, Saudi Arabia peter.richtarik@kaust.edu.sa Alexander Gasnikov MIPT, HSE University and IITP RAS , Russia gasnikov@yandex.ru
Pseudocode Yes Algorithm 1
Open Source Code No we run simple experiments just for theoretical purpose, it is easy to rerun it
Open Datasets Yes We take datasets from Li BSVM [19] and divided unevenly across M = 25 workers. For communication networks we chose the star, the ring and the grid topologies.
Dataset Splits No No specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) is provided for train/validation/test sets.
Hardware Specification No we run simple experiments just for theoretical purpose, they require no computing power and can be run on any laptop
Software Dependencies No The paper mentions 'LibSVM' as a source for datasets but does not provide specific version numbers for any software dependencies used in the experiments.
Experiment Setup Yes The parameters of all methods are selected in two ways: 1) as described in the theory of the corresponding papers, and 2) tuned for the best convergence. We run all methods with different batch sizes. The comparison criterion is the number of epochs (one full gradient = epoch).