SCENE OTA-FD: Self-Centering Noncoherent Estimator for Over-the-Air Federated Distillation
New method eliminates pilot signals, cutting overhead for AI training on mobile devices.
Deep Dive
Researchers Hao Chen and Zavareh Bozorgasl developed SCENE OTA-FD, a new aggregation primitive for over-the-air federated distillation. It uses a pilot-free, phase-invariant design where devices map soft-label vectors to transmit energy. The server's self-centering estimator removes noise, providing an unbiased average with variance decaying as 1/(SM). This targets hardware-constrained, short-coherence wireless regimes where avoiding channel state information overhead is critical.
Why It Matters
Enables more efficient distributed AI training on real-world devices like phones and sensors with limited bandwidth.