Quantized Federated Learning Under Transmission Delay and Outage Constraints

Federated learning (FL) has been recognized as a viable distributed paradigm which trains machine model collaboratively with massive mobile devices in the wireless edge while protecting user privacy. Although various communication schemes have proposed to expedite FL process, most of them assumed ideal channels provide reliable and lossless links between server clients. Unfortunately, practical systems limited radio resources such constraint on training latency constraints transmission power bandwidth, large number parameters inevitably suffers from quantization errors (QE) outage (TO). In this paper, we consider non-ideal channels, carry out first analysis showing that convergence can be severely jeopardized by TO QE, but intriguingly alleviated if clients uniform probabilities. These insightful results motivate us propose robust scheme, named FedTOE , performs joint allocation bits across minimize QE making same probability. Extensive experimental are presented show superior performance for deep learning-based classification tasks constraints.