ASU's Wireless AI Training Trick Hits 2x Faster Convergence by Embracing Bias
A new framework lets wireless federated learning deliberately embrace bias to escape the straggler bottleneck when devices have wildly different signal quality.
PRO- New wireless federated learning framework introduces a structured, time-invariant bias to cut update variance under heterogeneous channels.
- Unified convergence theory covers both OTA and digital FL, spanning strongly convex and non-convex objectives.
- Successive convex approximation optimizes participation, pre-scalers, and quantization bits offline using only statistical CSI.
- Achieves roughly 2x faster convergence versus SOTA digital FL baselines on MNIST softmax regression.
- Matches genie-aided OPC OTA-FL on accuracy while requiring no future or global instantaneous channel info.
- Validated on ResNet-18 with CIFAR-10, an 11.17M-parameter non-convex model, under extreme non-i.i.d. splits.
Federated learning has a stubborn wireless problem. In the real world, one phone sits next to the base station with a pristine channel, while another is at the cell edge with a signal barely above the noise floor. Most existing schemes either pretend this heterogeneity does not exist, or they force every aggregation to be statistically unbiased, which drags the whole system down to the pace of the worst device. A new paper from Arizona State University proposes a different bargain: let the model be a little biased on purpose, and use that freedom to slash the variance of updates.
The work, titled Biased Federated Learning under Wireless Heterogeneity and slated for IEEE Transactions on Wireless Communications, tackles both major flavors of wireless FL under one theoretical roof and delivers roughly 2x faster convergence than state-of-the-art baselines on standard image tasks.
The straggler tax nobody wants to pay
To understand why this matters, it helps to know how gradients travel in wireless FL. Two main strategies dominate. Over-the-air (OTA) computation exploits the natural superposition property of wireless multiple access channels, so all devices transmit simultaneously and the base station receives a single aggregated gradient. Digital transmission instead gives each device its own orthogonal time or frequency slot.
Both approaches usually assume every device has the same average path loss, which lets designers enforce unbiased updates cleanly. In practical heterogeneous networks, weaker channel devices act as stragglers, and enforcing a zero-bias design introduces high variance in the updates. For OTA, the classic fix is channel inversion, which forces transmit power to be dictated by the worst device. For digital FL, weak devices hog time slots to hit a required rate. Either way, the tail wags the dog.
Some prior work allowed non-zero bias but treated it as an uncontrolled side effect, with no theoretical hook to quantify or optimize it. That gap is what this paper closes.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
