Reproduce “Density Deconvolution with Normalizing Flows” 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
Tim Dockhorn, James A. Ritchie, Yaoliang Yu, Iain Murray
The authors model the density of clean values from noisy observations using a normalizing flow and approximate variational inference with a known noise likelihood. Although designed for varying noise, the method is tested with fixed Gaussian noise. On two small UCI wine datasets, flows beat XD-GMM; for white wine, Flow L_IW(50) reaches 8.685 ± 0.082 versus 9.903 ± 0.112, without establishing broader generality.
Key insights
- Normalizing flows provide a more expressive and flexible framework for modeling complex, potentially multimodal or non-Gaussian, underlying clean densities in deconvolution problems than Gaussian Mixture Models.
- The study identified that while normalizing flows possess the representational capacity for both prior and posterior, the standard variational inference objective can lead to a loose bound, biasing the learned prior and approximate posterior.
- The deconvolution framework is adaptable to arbitrary and observation-specific noise distributions, extending beyond the Gaussian noise assumption often seen in traditional methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.