To Talk or to Work: Dynamic Batch Sizes Assisted Time Efficient Federated Learning Over Future Mobile Edge Devices
To Talk or to Work: Dynamic Batch Sizes Assisted Time Efficient Federated Learning Over Future Mobile Edge Devices
复制标题
DOI：
10.1109/twc.2022.3189320
复制
发表时间：
2022-12-01
影响因子：
10.4
通讯作者：
Han, Zhu
中科院分区：
文献类型：
--
作者：
Shi, Dian;Li, Liang;Han, Zhu
联合学习（FL）和多访问边缘计算（MEC）的结合有可能促进许多应用。然而，在移动的边缘设备的有限通信和计算资源的情况下足够快地训练FL带来了巨大的挑战。受超高速无线传输的最新发展和人工智能（AI）计算硬件的移动的设备的有前途的进步，在本文中，我们提出了一个时间有效的FL在未来的移动的边缘设备，称为动态批量大小辅助联邦学习（DBFL）的收敛保证。DBFL允许批量大小在训练期间动态增加，这可以释放GPU的并行性的计算潜力用于设备上的训练，并有效地利用快速无线传输（WiFi-6、5G、6 G等）。移动的边缘设备。此外，基于推导出的DBFL的收敛界，我们开发了一种批量大小控制方案，以最小化FL在移动的边缘设备上的总时间消耗，该方案权衡了“说话“，即，通信时间和“工作“，即，计算时间，通过适当调整增量因子。进行了大量的模拟，以验证我们提出的DBFL算法的有效性，并证明我们的计划优于现有的时间有效的FL方法在各种设置的总时间消耗。
The coupling of federated learning (FL) and multi-access edge computing (MEC) has the potential to foster numerous applications. However, it poses great challenges to train FL fast enough with limited communication and computing resources of mobile edge devices. Motivated by recent development in ultra fast wireless transmissions and promising advances in artificial intelligence (AI) computing hardware of mobile devices, in this paper, we propose a time efficient FL over future mobile edge devices, called dynamic batch sizes assisted federated learning (DBFL) with convergence guarantee. The DBFL allows batch sizes to increase dynamically during training, which can unleash the computing potential of GPU's parallelism for on- device training and effectively leverage the fast wireless transmissions (WiFi-6, 5G, 6G, etc.) of mobile edge devices. Furthermore, based on the derived DBFL's convergence bound, we develop a batch size control scheme to minimize the total time consumption of FL over mobile edge devices, which trade-offs the "talking ", i.e., communication time, and "working ", i.e., computing time, by adjusting the incremental factor appropriately. Extensive simulations are conducted to validate the effectiveness of our proposed DBFL algorithm and demonstrate that our scheme outperforms existing time efficient FL approaches in terms of the total time consumption in various settings.