Options
MatchCurv: communication-efficient decentralized federated learning in heterogeneous environments
Publikationstyp
Conference Paper
Date Issued
2024
Sprache
English
Author(s)
Citation
1st Workshop on Enabling Machine Learning Operations for next-Gen Embedded Wireless Networked Devices (2024) at the 22nd International Conference on Embedded Wireless Systems and Networks, (EWSN 2025)
Scopus ID
Federated learning offers a privacy-preserving method for training machine learning models. Yet, traditional centralized federated learning has drawbacks like single points of failure and communication bottlenecks. While decentralized federated learning has been proposed to overcome these limitations, challenges such as statistical and system heterogeneity remain. Whereas most works focus on solving only one of these challenges, this paper introduces MatchCurv, a decentralized federated learning framework designed to handle statistical and system heterogeneity while improving communication efficiency. In our evaluation, for example, a multi-layer perceptron with two hidden layers of 128 units each shows a significant accuracy increase when using our framework compared to PD-SGD, achieving up to 17 percentage points more accuracy. The source code is available at https://github.com/ds-kiel/matchcurv.
Subjects
Decentralized Federated Learning
Deep Learning
FedCurv
Federated Learning
MATCHA
MLE@TUHH
DDC Class
600: Technology