An Experimental Study of DNN Online Split Learning on Edge Devices

Split learning (SL), as a novel distributed learning approach, significantly alleviates the computational burden on each individual client, making it feasible to train complex artificial intelligence(AI) models on resource-constrained edge devices. This paper conducts a pilot experimental study that integrates SL with online learning, aiming to explore the feasibility and performance of training complex deep neural networks (DNNs) under constantly arriving data samples on practical edge devices. We implement the online SL of some representative DNNs such as ResNet-110 and VGG16 on typical edge devices (i.e., NVIDIA Jetson devices), and evaluate their performances when the data samples of each device are constantly updated during the training process. Experimental results demonstrate that online SL successfully achieves efficient and high-precision DNN model training on edge devices. The incorporation of online learning in SL improves the real-time performance and practicality of the trained DNNs. Both ResNet-110 and VGG16 exhibit a steady increase in accuracy in the test set, eventually converging and stabilizing at a higher level of accuracy. Moreover, the accuracy improves by nearly 10% after incorporating online learning. These findings not only validate the effectiveness of online SL in edge computing environments, but also provide valuable experimental data and theoretical support for future deployment and application of complex DNNs in edge computing scenarios, like unmanned aerial vehicle (UAV) swarms, indicates promising prospects and profound implications in practical applications.