1: \begin{abstract}

2: In this paper, we propose a framework where \ac{OAC} occurs in both \ac{UL} and \ac{DL}, sequentially, in a multi-cell environment to address the latency and the scalability issues of \ac{FEEL}.  To eliminate the \ac{CSI} at the \acp{ED} and \acp{ES} and relax the time-synchronization requirement  for the \ac{OAC}, we use a non-coherent computation scheme, i.e., \ac{FSK-MV}. With the proposed framework,  multiple \acp{ES} function as the aggregation nodes in the \ac{UL} and each \ac{ES} determines the \acp{MV} independently. After the \acp{ES} broadcast the detected \acp{MV}, the \acp{ED} determine the sign of the gradient through another \ac{OAC} in the \ac{DL}. Hence, inter-cell interference is exploited for the \ac{OAC}. In this study, we prove the convergence of the non-convex optimization problem for the \ac{FEEL} with the proposed \ac{OAC} framework. We also numerically evaluate the efficacy of the proposed method by comparing the test accuracy in both multi-cell and single-cell scenarios for both homogeneous and heterogeneous data distributions. %We demonstrate  improvements of test accuracy in our \ac{OAC} method in the multi-cell scenario. Moreover, we demonstrate the superiority of our approach by performing training based on only local data in terms of accuracy.

3: \end{abstract}

4: