Optimal Cut Layer Bounds for Split Learning

Abstract

Split learning (SL) is a distributed learning method where a deep learning model is partitioned between the client and server, aiming to optimize the training process. A key challenge in split learning is selecting the cut layer to minimize energy consumption while considering both computational and communication overheads. In this letter, we address this challenge within the context of a wireless system with multiple clients and a central server. We introduce a pruning-based cut layer selection scheme that effectively reduces the energy consumption for each client. Our approach leverages analytical bounds for optimal cut layer location, which we derive and validate against state-of-the-art SL benchmark schemes, demonstrating the high efficiency of our proposed method.