Wait to Learn: Timing the Gradient Return in Over-the-Air Split Learning over an Aging Single-Antenna Channel
The paper investigates how to correct an aged channel gradient when using a single antenna receiver in over-the-air split learning. It found that uncorrected training fails near correlation zeros and recovers in positive lobes by delaying the gradient return until the next correlation extremum. This suggests that patience—waiting for the next significant correlation point—is a necessary strategy for accurate learning with aging channels.