Reproduce “Decomposing Earth Embeddings with Sparse Autoencoders” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vitus Benson, Fanny Yang, Markus Reichstein
The authors train sparse autoencoders (SAEs) on pixelwise satellite embeddings from AlphaEarth and Tessera, representing each pixel with a few active features from a larger dictionary. They evaluate reconstruction, concept alignment, and spatial scale, including a fixed-random-decoder control. At matched sparsity, AlphaEarth reconstruction reaches FVE 0.973 versus 0.775 for the control; specificity measures do not establish learned meaning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.