Sparse Federated Representation Learning for Planetary Geology Survey Missions

This article explores a novel approach called Sparse Federated Representation Learning (SFRL) to address the challenges of autonomous planetary geology survey missions with limited computational and energy resources.

Details

Planetary geology survey missions face severe constraints in terms of bandwidth, power, and onboard computing capabilities. Traditional approaches of either pre-programmed behaviors or downlinking all sensor data are inadequate. The author explored federated learning (FL) as a potential solution, but found that vanilla FL was still too costly for transmitting full model updates. The key breakthrough was the combination of three principles: Federated Learning to learn a shared model across agents without centralized data, Sparse Learning to maintain an extremely compact model with only critical parameters updated, and Representation Learning to focus on learning a general-purpose feature embedding that can be fine-tuned locally. This Sparse Federated Representation Learning (SFRL) approach creates a collaborative 'geological intuition' that can be efficiently adapted by individual agents to their unique environments, revolutionizing autonomous planetary exploration.