Decentralized learning has emerged as an influential paradigm for facilitating collaborative learning without a central server. With the Device-to-Device (D2D) communication mechanism, the collaborative topology in decentralized learning dictates the pathways of information flow, significantly affecting learning performance. Furthermore, latency becomes a critical factor in the architectural design of the decentralized learning system. Motivated by these considerations, this letter addresses the convergence prerequisites of decentralized learning by proposing two distinct topological configurations: ring topology and multiple ring topology. We transform these topology design challenges into optimization problems with the goal of minimizing global latency. Finally, we propose corresponding solution strategies and design time-efficient topologies for the optimization problems. Numerical results demonstrate that the proposed topological structures and solution methods can effectively balance latency and test accuracy performance in the D2D communication environment for decentralized learning.

中文翻译：

---

通过设备到设备通信实现分布式学习的省时环形拓扑设计


分散式学习已成为一种有影响力的范式，可以在没有中央服务器的情况下促进协作学习。借助设备到设备 （D2D） 通信机制，去中心化学习中的协作拓扑决定了信息流的路径，从而显着影响学习性能。此外，延迟成为去中心化学习系统架构设计中的一个关键因素。在这些考虑的推动下，这封信通过提出两种不同的拓扑配置来解决去中心化学习的收敛先决条件：环形拓扑和多环形拓扑。我们将这些拓扑设计挑战转化为优化问题，目标是最大限度地减少全局延迟。最后，我们提出了相应的求解策略，并为优化问题设计了省时的拓扑结构。数值结果表明，所提出的拓扑结构和求解方法可以有效地平衡 D2D 通信环境中的延迟和测试精度性能，以实现去中心化学习。