Federated Dropout (FedDrop) has each wireless edge device train and transmit only a randomly dropped sub-model of the shared network to cut per-round communication and computation. This establishes its missing convergence theory: a Taylor-expansion analysis shows dropout inflates stochastic-gradient variance by gamma/(1-gamma) with dropout rate gamma, so a higher dropout rate slows the convergence rate, trading shorter per-round latency for more rounds. Building on that bound, it jointly optimizes per-device dropout rates and wireless bandwidth allocation to minimize the final training loss under a per-round latency budget and limited bandwidth.
Federated Dropout (FedDrop) is a communication- and computation-efficient variant of federated learning for wireless edge devices: in each training round every device updates and transmits only a ran…