Towards Faster Decentralized Stochastic Optimization with Communication Compression

Authors: Rustem Islamov, Yuan Gao, Sebastian Stich

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide numerical experiments to validate our theoretical findings and confirm the practical superiority of Mo TEF. ... 4 NUMERICAL EXPERIMENTS In this section, we complement the theoretical results on the convergence of Algorithm 1 with numerical evaluations.
Researcher Affiliation Academia Rustem Islamov1 Yuan Gao2,3 Sebastian U. Stich3 1Universität Basel, 2Universität des Saarlandes, 3CISPA Helmholtz Center for Information Security EMAIL, EMAIL
Pseudocode Yes Algorithm 1 Mo TEF ... Algorithm 2 Mo TEF-VR
Open Source Code Yes Our implementation is based on open-source code from (Zhao et al., 2022) https://github.com/ liboyue/beer and is available at https://anonymous.4open.science/r/Mo TEF-0DCF. The code to reproduce our synthetic experiment is available at https:// anonymous.4open.science/r/decentralized-exp-A3C6
Open Datasets Yes We set λ = 0.05, n = 100 and use Lib SVM datasets (Chang & Lin, 2011). ... Finally, we consider training MLP on MNIST dataset (Deng, 2012) with 1 hidden layer of size 32.
Dataset Splits No We do not shuffle datasets to have a more heterogeneous setting. Besides, each dataset is equally distributed among all clients. ... Finally, we consider training MLP on MNIST dataset (Deng, 2012) with 1 hidden layer of size 32. The paper does not provide explicit details about training/test/validation splits (e.g., percentages, sample counts, or citations to specific standard splits used in their experimental setup) for the datasets mentioned.
Hardware Specification Yes We run our experiments on AMD EPYC 9554 64-Core Processor.
Software Dependencies Yes To find a suitable choice of constants we use Symbolic Math Toolbox in MATLAB (Inc., 2023). Our code can be found at https://anonymous.4open.science/r/dec-symb-verification.
Experiment Setup Yes We fix the parameters γ = 0.1, η = 0.0005, λ = 0.005, and n = 16, if the opposite is not stated. ... For Mo TEF we tune stepsize as follows η {0.001, 0.01, 0.05}, γ {0.1, 0.2, 0.5, 0.9}, λ {0.005, 0.01, 0.05, 0.1}.