Physical backpropagation improves by waiting over aging channels
Wait to Learn: Lobes and Dead Zones in Physical Backpropagation over an Aging Single-Antenna Channel
Information Theory
Summary
Sending feedback signals through a wireless channel can get tricky when the channel changes over time. The paper finds that sometimes the best way to get a useful signal back is to wait until the channel conditions improve. The authors show that in certain moments when the connection is weak, trying to correct the information right away doesn't help, but holding onto it and waiting for better moments works better. This helps improve training in systems that learn by sending gradients back through the channel.
What this means in practice
- •For wireless communication engineers: Improve gradient-based training accuracy over single-antenna channels by timing feedback returns to favorable channel conditions.
- •For signal processing teams: Design feedback protocols that delay transmission of gradient signals until channel conditions yield more accurate corrections.
Authors
Xianzhong Li, Yuzhi Yang, Tianhao Guo
Abstract
Sometimes the best way to correct an aged gradient is to wait. In physical backpropagation, a receiver returns its loss gradient through the reciprocal channel. With a single moving antenna, the mean gradient returned at the antenna is the true one counter-rotated by the phase of the temporal correlation. In simulation, uncorrected training fails in negative correlation lobes and revives in positive ones. At correlation zeros the return is independent of the forward channel, so no linear correction has a mean to recover; near them, more calibration barely helps, but holding the return until the next correlation extremum does.