Federated Online Learning and Optimization: Data Freshness, Model Robustness, and Communication Efficiency (Juncheng Wang et al.)

With the migration of traditional machine learning to federated learning, transmitting model parameters between central servers and local devices can incur a huge amount of communication overhead. This calls for communication-efficient federated learning algorithms that integrate techniques from both machine learning and wireless communications. Most existing works separately optimized machine learning and wireless communications. However, due to the tight coupling between communication efficiency and model transmission, we have observed that the performance of machine learning can be significantly improved by proactively optimizing wireless communications, and vice versa [1], [2]. Meanwhile, most existing works adopted offline optimization based on static computation data and communication environment, ignoring the possibility of adapting model training and transmission to the unknown system dynamics through online optimization [3], [4]. These motivated us to rethink online federated learning through joint optimization of computation and communication in the following three aspects.

Data Freshness: In many machine learning applications, e.g., real-time video analysis, dynamic user profiling, and network traffic classification, new data arrives to local devices in a streaming fashion and thus the local loss functions becomes time-varying. Furthermore, due to the network heterogeneity in computational capacity, the durations required for the local devices to update their local models differ significantly. Without any guarantee on the freshness of the data used in training, the learned model may only fits well to outdated data.

Model Robustness: With streaming data, the distributions of the data used in training may not be independent and identically distributed. Such heterogeneity in data distribution can have a significant impact on the performance of federated learning. Specifically, when the data distributions differ much among the local devices, the trained machine learning model may not fit well to individual data distributions. Furthermore, when the distributions of the local training datasets vary over time, the trained machine learning model can be more fragile in robustness.

Communication Efficiency: In federated learning, a key operation is to aggregate the local model parameters sent from local devices as a global model at the central server. Most of the machine learning methods, e.g., quantization, sparsification, and local updates, considered lossless transmission and ignored the wireless communication layer. Since the central server only needs the aggregated model instead of individual models, over-theair analog aggregation can efficiently reduce the communication overhead. However, most existing over-the-air federated learning approaches separately optimized model update and analog aggregation in offline settings.

Based on the above observations, this project aims at designing new online federated learning algorithms, to improve the data freshness, model robustness, and communication efficiency of federated learning in dynamic environments, via joint online optimization of computation and communication.


References:


For further information on this research topic, please contact Dr. Juncheng Wang.