.png){kind=icon}
.png){kind=icon}
Return
A fully decentralized distributed learning algorithm for latency communication networks
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2024.112829.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The network-induced delay phenomenon has always been one of the bottlenecks in the practical application of distributed learning (DL) algorithms. This paper aims to explore a novel fully decentralized DL algorithm based on the zero-gradient-sum (ZGS) strategy for latency networks, called TDDL (Time Delay Distributed Learning) algorithm. Compared with previous related works, the main challenge is to make the performance of DL algorithm free from the influence of latency in communication networks when applied to real-world scenarios. Specifically, modeled as a distributed optimization problem for latency networks, the distributed learning problem is solved by the improved discrete-time ZGS algorithm. Furthermore, the convergence of the TDDL algorithm is analyzed by constructing a Lyapunov-Krasovskii functional. And we theoretically derive an upper bound on the time delay, which ensures that the TDDL algorithm for latency communication network promotes all agents to cooperatively converge as a sufficient condition. More importantly, the proposed algorithm benefits privacy protection in latency networks in a fully decentralized distributed manner. In other words, each agent only exchanges information with neighboring agents, and model parameters rather than raw data are merely transmitted. Finally, several experiments are given to illustrate the effectiveness of the TDDL algorithm.
Keywords:
Decentralized distributed learning
Zero gradient sum
Latency network
Privacy-preserving
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W