Orbit-Aware Split Learning: Optimizing LEO Satellite Networks for Distributed Online Learning

This paper proposes a novel split learning architecture designed to exploit the cyclical movement of Low Earth Orbit (LEO) satellites in non-terrestrial networks (NTNs). Although existing research focuses on offloading tasks to the NTN infrastructure, these approaches overlook the dynamic movement patterns of LEO satellites that can be used to efficiently distribute the learning task. In this work, we analyze how LEO satellites, from the perspective of ground terminals, can participate in a time-window-based model training. By splitting the model between a LEO and a ground terminal, the computational burden on the satellite segment is reduced, while each LEO satellite offloads the partially trained model to the next satellite in the constellation. This cyclical training process allows larger and more energy-intensive models to be deployed and trained across multiple LEO satellites, despite their limited energy resources. We formulate an optimization problem that manages radio and processing resources, ensuring the entire data is processed during each satellite pass while minimizing the energy consumption. Our results demonstrate that this approach offers a more scalable and energy-efficient way to train complex models, enhancing the capabilities of LEO satellite constellations in the context of Artificial Intelligence-driven applications.