Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a mixed-timescale differential coding scheme for downlink model broadcasts in wireless federated learning, optimizing resource usage by compressing updates into three distinct timescales. By allowing devices to receive updates more frequently and minimizing the impact of missed transmissions, the approach significantly reduces model staleness and enhances overall learning efficiency. Simulations demonstrate that this method outperforms existing baseline and state-of-the-art techniques, achieving superior performance with comparable communication resources.
Devices can now maintain fresher models even in challenging wireless environments, thanks to a novel mixed-timescale differential coding strategy.
In wireless federated learning, the server needs to periodically broadcast the global model to all devices. To optimize resource usage, this broadcast can be compressed using differential coding, reducing the frequency of full model transmissions and instead sending differential updates. However, devices may occasionally miss a differential update due to fading and interference, preventing them from reconstructing the original model. Consequently, the device must wait for the next full model broadcast, forcing it to work with an outdated model or halt the learning process. To improve efficiency, we propose a mixed-timescale differential coding scheme, where the model is broadcast at three timescales: 1) a full model with a long timescale, 2) a first-level differential model with an intermediate timescale, and 3) a second-level differential model with a fast timescale. This design ensures that if a device misses a second-level update, it only waits for the next first-level differential update, which is more frequent than the full model broadcast. This effectively reduces model staleness. Additionally, we propose an age-aware scheduling policy to address decoding failures, prioritizing devices with fresher models during uplink aggregation. Simulations show that our methods outperform baseline and state-of-the-art approaches, delivering superior learning performance with similar communication resources.