Reproduce “Diffusion Models and Concept Formation” 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

Zekun Wang, Karthik Singaravadivelan, Christopher J. MacLellan

The paper interprets diffusion models as concept hierarchies by following score ascent from noisy images: modes at low noise represent instances, while increasing noise merges them into prototypes. It recovers this hierarchy and compares it with Cobweb using distinctiveness. Both MNIST and Fashion-MNIST peak at diffusion timestep 150 and Cobweb depth 3, but evidence is limited to two benchmarks.

Summary by alphaXiv.

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