Abstract :
[en] Federated Learning (FL), as an effective decentral-
ized approach, has attracted considerable attention in privacy-
preserving applications for wireless edge networks. In practice,
edge devices are typically limited by energy, memory, and
computation capabilities. In addition, the communications be-
tween the central server and edge devices are with constrained
resources, e.g., power or bandwidth. In this paper, we propose
a joint sparsification and optimization scheme to reduce the
energy consumption in local training and data transmission.
On the one hand, we introduce sparsification, leading to a
large number of zero weights in sparse neural networks, to
alleviate devices’ computational burden and mitigate the data
volume to be uploaded. To handle the non-smoothness incurred
by sparsification, we develop an enhanced stochastic gradient
descent algorithm to improve the learning performance. On
the other hand, we optimize power, bandwidth, and learning
parameters to avoid communication congestion and enable an
energy-efficient transmission between the central server and edge
devices. By collaboratively deploying the above two components,
the numerical results show that the overall energy consumption
in FL can be significantly reduced, compared to benchmark FL
with fully-connected neural networks.
Scopus citations®
without self-citations
0