OFedQIT: Communication-Efficient Online Federated Learning via Quantization and Intermittent Transmission

Online federated learning (OFL) is a promising framework to collaboratively learn a sequence of non-linear functions (or models) from distributed streaming data incoming to multiple clients while keeping the privacy of their local data. In this framework, we first construct a vanilla method (named OFedAvg) by incorporating online gradient descent (OGD) into the de facto aggregation method (named FedAvg). Despite its optimal asymptotic performance, OFedAvg suffers from heavy communication overhead and long learning delay. To tackle these shortcomings, we propose a communication-efficient OFL algorithm (named OFedQIT) by means of a stochastic quantization and an intermittent transmission. Our major contribution is to theoretically prove that OFedQIT over $T$ time slots can achieve an optimal sublinear regret bound $\mathcal{O}(\sqrt{T})$ for any real data (including non-IID data) while significantly reducing the communication overhead. Furthermore, this optimality is still guaranteed even when a small fraction of clients (having faster processing time and high-quality communication channel) in a network are participated at once. Our analysis reveals that OFedQIT successfully addresses the drawbacks of OFedAvg while maintaining superior learning accuracy. Experiments with real datasets demonstrate the effectiveness of our algorithm on various online classification and regression tasks.