查看文章

Machine learning is one of the key building blocks in 5G and beyond [1–3], spanning a broad range of applications and use cases. In the context of mission-critical applications [2, 4], machine learning models should be trained with fresh data samples that are generated by and dispersed across edge devices (eg, phones, cars, access points, etc.). Collecting these raw data incurs significant communication overhead, which may violate data privacy. In this regard, federated learning (FL)[5–8] is a promising communication-efficient and privacy-preserving solution that periodically exchanges local model parameters, without sharing raw data. However, exchanging model parameters is extremely costly under modern deep neural network (NN) architectures that often have a huge number of model parameters. For instance, MobileBERT is a stateof-the-art NN architecture for on-device natural language processing (NLP …