TY - CONF
T1 - Wireless Time-Triggered Federated Learning with Adaptive Local Training Optimization
AU - Zhou, Xiaokang
AU - Deng, Yansha
AU - Xia, Huiyun
AU - Wu, Shaochuan
N1 - Funding Information:
This work was supported in part by the National Natural Science Foundation of China under Grant 62271167, 62171163 and 61831002; and in part by UK Research and Innovation (UKRI) under the UK government’s Horizon Europe funding guarantee (grant number 10061781), as part of the European Commission-funded collaborative project VERGE, under Smart Networks and Services Joint Undertaking (SNS JU) program (grant number 101096034).
Publisher Copyright:
© 2023 IEEE.
PY - 2023
Y1 - 2023
N2 - Traditional synchronous and asynchronous federated learning suffer from limitations, such as straggler, high communication overhead, and model staleness problems. As a generalization of the former two, time-triggered federated learning (TT-Fed) offers a new perspective to alleviate these problems and provides better trade-off between communication and training efficiencies. In this paper, we consider the dynamic aggregation frequency optimization problem under constrained system resources for TT-Fed. We analyze how the number of local training epochs affects the performance of TT-Fed and quantify the impacts of distributed data heterogeneity and model staleness heterogeneity to its convergence upper bound. Based on the derived upper bound, we propose an adaptive local training optimization algorithm to minimize the system loss under constrained resource budget. Numerical simulations show that compared to TT-Fed with fixed number of local training epochs, our proposed adaptive optimization algorithm can provide near optimal results under different degrees of non-IID data distributions.
AB - Traditional synchronous and asynchronous federated learning suffer from limitations, such as straggler, high communication overhead, and model staleness problems. As a generalization of the former two, time-triggered federated learning (TT-Fed) offers a new perspective to alleviate these problems and provides better trade-off between communication and training efficiencies. In this paper, we consider the dynamic aggregation frequency optimization problem under constrained system resources for TT-Fed. We analyze how the number of local training epochs affects the performance of TT-Fed and quantify the impacts of distributed data heterogeneity and model staleness heterogeneity to its convergence upper bound. Based on the derived upper bound, we propose an adaptive local training optimization algorithm to minimize the system loss under constrained resource budget. Numerical simulations show that compared to TT-Fed with fixed number of local training epochs, our proposed adaptive optimization algorithm can provide near optimal results under different degrees of non-IID data distributions.
UR - http://www.scopus.com/inward/record.url?scp=85171628715&partnerID=8YFLogxK
U2 - 10.1109/INFOCOMWKSHPS57453.2023.10226087
DO - 10.1109/INFOCOMWKSHPS57453.2023.10226087
M3 - Paper
AN - SCOPUS:85171628715
T2 - 2023 IEEE INFOCOM Conference on Computer Communications Workshops, INFOCOM WKSHPS 2023
Y2 - 20 May 2023
ER -