Over-the-Air Federated Learning in Heterogeneous Mobile Wireless Networks
FedOAG enables unbiased over-the-air federated learning across heterogeneous fading without requiring every device to transmit.
The paper addresses over-the-air federated learning when wireless fading is heterogeneous. Existing designs either force unbiased updates from every device, amplifying noise from the weakest channel, or accept partial participation that converges to a biased objective and depends on the fading model. FedOAG uses gradient normalization to meet energy constraints and implicit gossiping to mix device updates, without requiring all devices to transmit or knowledge of time-varying channel statistics. The authors prove convergence to a stationary point of an unbiased non-convex objective at rate O(1/sqrt(T)) and report numerical results on real-world datasets under dynamic wireless conditions.
- Heterogeneous fading either amplifies noise or biases partial aggregation.
- FedOAG normalizes gradients and mixes updates through implicit gossiping.
- It requires neither universal transmission nor a known fading model.
- Analysis proves unbiased non-convex convergence at rate O(1/sqrt(T)).
From the original article
1 min read at arxiv.orgOver-the-air computation has emerged as a scalable and efficient solution for deploying federated learning algorithms in wireless networks by exploiting waveform superposition for simultaneous model aggregation. Most existing work struggles with heterogeneous fading channels.
By Ming Xiang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.