New Trick Lets Slow Phones Help Train AI Without Slowing Everyone Down

Distributed optimization in edge networks gets squeezed by clients with different computing capabilities and limited communication. Researchers propose WQ-GADMM — Windowed and Quantized Group-Based ADMM — which groups clients by estimated computation time, activates only a limited number of groups per round, and has the cloud update the global model once every group has updated. It quantizes both downlink and uplink model exchanges, and allows bounded model staleness and inexact proximal local updates. For smooth nonconvex objectives, the authors establish an average squared Karush-Kuhn-Tucker residual bound under stated assumptions and parameter conditions, made up of a term that decreases with iteration count plus a quantization-dependent error term. On MNIST and CIFAR-10, 12-bit communication reduced communication volume and simulated wall-clock time while keeping test accuracy comparable to full precision, and the 12-bit configuration maintained complete group coverage with shorter mean group inter-completion gaps than the evaluated baselines.