Reproduce “SCAN: Learning Hierarchical Compositional Visual Concepts” 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

Irina Higgins, Nicolas Sonnerat, Loic Matthey, Arka Pal, Christopher P Burgess, Matko Bosnjak, Murray Shanahan, Matthew Botvinick and 2 more

SCAN learns reusable visual factors from images, then uses image–symbol pairs to ground symbols in concepts whose specified properties stay fixed while unspecified properties can vary. An operator combines concepts using AND, IN COMMON, and IGNORE. In DeepMind Lab, SCAN achieves 0.79 accuracy and 1.05 diversity on 50 operator-test concepts, outperforming SCAN_R, but this result is limited to a room-image task.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.