Motivation:

Satellite mega constellations in low Earth orbit (LEO) offer the potential to collect vast amounts of data. This data can be used to train deep machine learning models for various tasks. The conventional centralized approach, based on training these models in terrestrial data centers, imposes considerable challenges on the task of data collection, as well as delaying the whole training process. A potential solution is directly training these models in orbit, distributedly on the satellites leveraging the federated learning paradigm. Based on initial feasibility studies, this appears to be a promising approach for the design of future mega constellations that offers a large amount of open research challenges.

Goal:

Several open problems exist in this general research area, for example in the design of

• distributed optimization algorithms for asynchronous federated learning in satellites;         
• training in satellite system architecture;         
• task-specific communication, networking, and routing methods; and         
• energy-efficient resource allocation for federated learning in satellite constellations.         

Specific problems will be defined in discussion with interested students according to their interests.

Requirement:

We are looking for highly motivated students with analytical and mathematical skills, as well as a broad fundamental knowledge in the area of ​​communication systems. Programming skills in Python and/or C++ are beneficial for most tasks, but not mandatory.

Reference:

Bho Matthiesen , Nasrin Razmi, Israel Leyva-Mayorga, Armin Dekorsy, and Petar Popovski. "Federated Learning in Satellite Constellations." arXiv preprint arXiv:2206.00307 (2022).