Scalable Aggregated Split Learning for Data-Driven Edge Intelligence on Internet-of-Things
Xinchen Lyu, Shuhan Liu, Junlin Liu, Chenshan Ren · IEEE Internet of Things Magazine · 2023
By combining Al techniques with edge computing, edge intelligence (El) is a promising paradigm for future intelligent Internet-of-Things (IoT). Split learning is one of the underlying technologies for El, where the computation-intensive model portions are offloaded to the edge server and the privacy-sensitive model portions are kept locally. The merits of split learning include data privacy and computation efficiency, which are paramount to intelligent IoT. However, split learning may not be scalable to massive IoT devices due to the excessive training latency, and limited computation and communication resources of the edge server. This article presents a novel scalable aggregated split learning framework that can significantly reduce the server-side computation and communication overhead for intelligent IoT. By exploiting the empirical expectation definitions of loss functions, the edge server is meticulously designed to aggregate the local loss functions of mas-sive devices to the global loss function and perform only one time of aggregated server-side backpropagation. After receiving the multicast global cut-layer gradients from the edge server, the devices can perform local training to generate synchronized device-side models. Case studies are shown to validate the effectiveness of the proposed framework against typical decentralized learning frameworks.