1: \begin{abstract}

2: In this paper, we study a federated learning system at the wireless edge

3: that

4: uses over-the-air computation (AirComp).

5: In such a system, users transmit their messages over a multi-access channel

6: concurrently to achieve fast model aggregation.

7: Recently, an AirComp scheme based on digital modulation

8: has been proposed featuring one-bit

9: gradient quantization and truncated channel inversion at users

10: and a majority-voting based decoder at the

11: fusion center (FC).

12: We propose an improved digital AirComp scheme to relax its requirements on the transmitters,

13: where users perform phase correction and transmit with full power.

14: To characterize the

15: decoding failure probability at the FC,

16: we introduce

17: the normalized detection signal-to-noise ratio (SNR), which can be interpreted as the

18: effective participation rate

19: of users.

20: To mitigate wireless fading, we further propose a cluster-based

21: system and design the relay selection scheme based on the normalized

22: detection SNR. By local data fusion within each cluster and relay selection,

23: our scheme can fully exploit spatial diversity to increase the effective

24: number of voting users and accelerate model convergence.

25: \end{abstract}

26: