Reproduce “Neural Embedding Compression For Efficient Multi-Task Earth Observation Modelling” on OpenResearch

About this paper

Carlos Gomes, Thomas Brunschwiler

A framework called Neural Embedding Compression (NEC) enables highly efficient data transfer and storage for Earth Observation Foundation Model embeddings. The method reduces data volumes by one to three orders of magnitude while preserving high accuracy on tasks such as scene classification and semantic segmentation.

Key insights

• A self-supervised learning loss, specifically MAE reconstruction, can effectively serve as a proxy for multi-task utility, enabling highly compressed embeddings without explicit downstream task tuning.
• Pre-trained Foundation Models can be efficiently adapted for compression by updating only a small subset of their parameters, significantly reducing computational cost and training time.
• Neural Embedding Compression provides one to three orders of magnitude data reduction while retaining high performance, especially excelling at extreme compression levels compared to traditional methods.

Summary by alphaXiv.