Split Learning and Synergetic Inference
Joana Tirana, Dimitris Chatzopoulos · 2025
Machine learning (ML) models rely on the data produced by the sensors of Internet of Things (IoT) devices. Traditionally, this data is processed in a centralized manner by the server. Nevertheless, this can be resource-intensive and privacy-invading, as a vast amount of data needs to be transferred to the server and models with millions of parameters need to be trained. This has motivated the growth toward client-based distributed training, enabled by advancements in IoT. Such an example is federated learning (FL), where multiple IoT devices train their models locally and synchronize their results. But resource limitations on small devices do not support large-scale models. To address this problem, offloading techniques like split learning (SL) and synergetic inference (SI) are proposed. Such techniques facilitate the collaboration between IoT and powerful compute nodes using the cloud-edge continuum. SL and SI involve splitting the ML model and offloading parts onto a compute node while keeping data on their devices. This chapter provides an overview of SL and SI, discussing their applications and evolution, with a primary focus on SL. We will study and organize the SL approaches according to the challenges they are targeting, and we will provide a discussion about potential future directions.