Conceptual

LEO-Split: Semi-Supervised Split Learning over LEO Satellite Networks

A distributed deep-learning framework for low-earth-orbit satellite constellations that trains models collaboratively between satellites and a ground station under unlabeled data, weak on-board compute, and intermittent connectivity. It couples split learning with semi-supervised learning and adds three mechanisms: an auxiliary model that sustains training through satellite-ground non-contact windows, a pseudo-labeling algorithm that corrects cross-satellite data imbalance, and an adaptive activation-interpolation scheme that curbs server-side sub-model overfitting; it outperforms prior distributed-learning baselines on real Starlink-like orbital traces.