Reproduce “CountsDiff: A Diffusion Model on the Natural Numbers for Generation and Imputation of Count-Based Data” 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
Renzo G. Soatto, Anders Hoel, Greycen Ren, Shorna Alam, Stephen Bates, Nikolaos P. Daskalakis, Caroline Uhler, Maria Skoularidou
CountsDiff models natural-number counts by removing units and learning to restore them; reverse steps can also remove units. In a synthetic sparse-count test, it better matches the target than Gaussian and masked categorical diffusion. For fetal-cell-atlas imputation at 50% MCAR, its SWD is 0.08 (SE 0.01), versus ReMDM’s 0.12 (SE 0.01); this is not a general robustness claim.
Key insights
- Directly modeling count data on natural numbers with a pure-death process and binomial marginals offers a more appropriate and efficient generative approach compared to continuous or categorical relaxations.
- Integrating advanced diffusion features like attrition (allowing non-monotonic count changes) and classifier-free guidance significantly enhances sample quality, robustness, and control in count-based data generation and imputation.
- A reparameterized p-schedule within a continuous-time framework for count-based diffusion allows for stable training and effective transfer of noise schedule designs from well-established Gaussian diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.