Efficient Gradient Updating Strategies with Adaptive Power Allocation for Federated Learning over Wireless Backhaul

. 2021 Oct 13;21(20):6791. doi: 10.3390/s21206791

Algorithm 2. Proposed federated learning train process.

1:
Initialize $θ_{0}$ based on He normal weight initialization method
2:
$g_{0} \leftarrow 0$
3:
for $t \leftarrow 1$ to T do
4:
(Clients) $θ_{t}^{(l)} \leftarrow θ_{t - 1}^{(l)} + g_{t - 1}$
5:
(Clients) Update $g_{t}^{(l)}$ from the datasets $Φ_{t r}^{(l)}$
6:
(Clients) Report $g_{t}^{(l)}$ to the server via the backhaul link
7:
if $\sum_{l = 1}^{L} {∥ {\bar{H}}_{l, t} ∥}_{F}^{2} < ϵ$ then
8:
(Server) $g_{t} \leftarrow \sum_{l = 1}^{L} w_{t}^{(l)} {\hat{\bar{g}}}_{t}^{(l)}$ as in (14)
9:
else
10:
(Server) $g_{t} \leftarrow g_{t - 1}$
11:
end if
12:
(Server) Broadcast $g_{t}$ to multi-clients
13:
end for