Subgradient Descent Learning over Fading Multiple Access Channels with Over-the-Air Computation

We focus on a distributed learning problem in communication network, consisting of N nodes and central parameter server (PS). The PS is responsible for performing the computation based data received from nodes, which are transmitted over multiple access channel (MAC). objective function this sum local loss functions nodes. This has gained attention field sensing systems, as well area federated (FL) recently. However, current approaches to solving rely assumption that continuously differentiable. In paper, we first address case where does not hold. develop novel algorithm called Sub-Gradient descent Multiple Access (SGMA) solve MAC. SGMA involves each node transmitting an analog shaped waveform its subgradient MAC, receiving superposition noisy signals, resulting bandwidth-efficient over-the-air (OTA) used update learned model. analyze performance prove it convergence rate centralized large networks. Simulation results using real datasets show effectiveness SGMA.