Semi-Federated Learning for Internet of Intelligence
| Autor: | Ni, Wanli, Yang, Zhaohui |
|---|---|
| Rok vydání: | 2024 |
| Předmět: | |
| Druh dokumentu: | Working Paper |
| Popis: | One key vision of intelligent Internet of Things (IoT) is to provide connected intelligence for a large number of application scenarios, such as self-driving cars, industrial manufacturing, and smart city. However, existing centralized or federated learning paradigms have difficulties in coordinating heterogeneous resources in distributed IoT environments. In this article, we introduce a semi-federated learning (SemiFL) framework to tackle the challenges of data and device heterogeneity in massive IoT networks. In SemiFL, only users with sufficient computing resources are selected for local model training, while the remaining users only transmit raw data to the base station for remote computing. By doing so, SemiFL incorporates conventional centralized and federated learning paradigms into a harmonized framework that allows all devices to participate in the global model training regardless of their computational capabilities and data distributions. Furthermore, we propose a next-generation multiple access scheme by seamlessly integrating communication and computation over the air. This achieves the concurrent transmission of raw data and model parameters in a spectrum-efficient manner. With their abilities to change channels and charge devices, two emerging techniques, reconfigurable intelligent surface and wireless energy transfer, are merged with our SemiFL framework to enhance its performance in bandwidth- and energy-limited IoT networks, respectively. Simulation results are presented to demonstrate the superiority of our SemiFL for achieving edge intelligence among computing-heterogeneous IoT devices. Comment: 8 pages, submitted to IEEE magazines |
| Databáze: | arXiv |
| Externí odkaz: |
|