Federated Learning via Local Update with Uploading Zone
Seong Hoon Jeon, Dong In Kim · 2022 13th International Conference on Information and Communication Technology Convergence (ICTC) · 2022
In this paper, we propose a novel idea of Uploading Zone for Federated Learning (FL) which largely affects transmission time and the number of participating devices. In order for sufficient devices to join FL, we assume heterogeneous quantization (HQ) where each device adapts its quantization level according to its signal-to-noise ratio (SNR). The devices located in the inner uploading zone utilize more bits for quantization because they have enough power to transmit gradients. However, the devices located in the outer uploading zone use fewer bits for quantization because of their limited transmit power. Since the quantization heterogeneity affects learning performance, we propose how to optimize the number of local updates under this HQ condition. Further, we consider the mobility condition where devices move freely and change their locations, which renders to increase traffic diversity. To look into these effects, we compare two types of uploading zone, small and large zones, without and with mobility. Finally, we optimize the size of the uploading zone and compare the two zones with the optimal uploading zone without mobility.