An Edge-enabled Wireless Split Learning Testbed: Design and Implementation

With advancements in the sixth generation (6G) research, the future communication network is characterized by its integration of artificial intelligence (AI). Among all AI frameworks, split learning (SL) partitions the AI model into device-side and server-side components, distributing them at both wireless device and edge server. Depending on the number of split layers, there exist a trade-off between the computation and communication costs. Until now, the impact of the wireless channel attenuation on the SL performance and impact of SL on the communication cost have been rarely studied, not to mention capturing these impacts in a practical testbed. In this paper, we build the first edge-enabled wireless split learning testbed (EWSLT), where the Intel smart edge open platform is integrated with the software-defined radio (SDR)-based 5G OpenAirInterface (OAI) platform, with split learning implemented between 5G-edge server and 5G user equipment (UE). We then evaluate the trade-off between 5G UE’s computation resource and wireless communication costs by the wireless SL’s effectiveness in terms of training accuracy and overall communication delay across various split layers.