Time-sensitive Learning for Heterogeneous Federated Edge Intelligence

Real-time machine learning (ML) has recently attracted significant interest due to its potential support instantaneous learning, adaptation, and decision making in a wide range of application domains, including self-driving vehicles, intelligent transportation, industry automation. In this paper, we investigate real-time ML federated edge intelligence (FEI) system, an computing system that implements (FL) solutions based on data samples collected uploaded from decentralized networks, e.g., Internet-of-Things (IoT) and/or wireless sensor networks. FEI systems often exhibit heterogenous communication computational resource distribution, as well non-i.i.d. arrived at different servers, resulting long model training time inefficient utilization. Motivated by fact, propose time-sensitive (TS-FL) framework minimize the overall run-time for collaboratively shared with desirable accuracy. Training acceleration both TS-FL synchronous coordination (TS-FL-SC) asynchronous (TS-FL-ASC) are investigated. To address straggler effect TS-FL-SC, develop analytical solution characterize impact selecting subsets servers time. A server dropping-based is proposed allow some slow-performance be removed participating if their accuracy limited. joint optimization algorithm consumption local epoch number (the iterations per coordination), batch size each iteration). fact slowest may special characteristics cannot training, expression staleness FL TS-FL-ASC. We load forwarding-based allows slow offload part trusted higher processing capability. hardware prototype evaluate heterogeneous system. Experimental results show our TS-FL-SC TS-FL-ASC can provide up 63% 28% reduction, time, respectively, compared traditional solutions.