Low-Earth-orbit (LEO) satellites are increasingly expected to perform machine learning directly onboard, enabling real-time responses for disaster and environmental applications. But the conditions on a satellite—limited compute, energy, and bandwidth—make standard federated learning impractical.
The paper introduces a resource-aware federated mixture-of-experts framework with adaptive pruning. This design appears to tune model complexity to the available resources during training and inference, a natural fit for satellite hardware.
The abstract is brief and does not include experimental results or architectural details, but the motivation is clear: onboard learning in LEO constellations needs a more resource-conscious alternative to conventional FL, and this work proposes one.