Reproduce “Quantifying Statistical Significance of Neural Network-based Image Segmentation by Selective Inference” 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

Vo Nguyen Le Duy, Shogo Iwazaki, Ichiro Takeuchi

Quantifying the statistical reliability of Deep Neural Network (DNN)-based image segmentation, this research from Nagoya Institute of Technology, Nagoya University, and RIKEN introduces a Selective Inference framework. A novel homotopy method characterizes the complex DNN selection event, enabling the computation of exact p-values that accurately control the False Positive Rate and outperform conventional statistical tests.

Key insights

  • The 'selection event' of piecewise-linear DNNs can be rigorously characterized for Selective Inference by identifying breakpoints, extending exact statistical guarantees to deep learning models for the first time.
  • Correctly accounting for the data-driven nature of DNN-based hypotheses via conditional inference is crucial, as traditional statistical tests fail to control the False Positive Rate.
  • The computational efficiency of the homotopy method, with breakpoint calculations increasing almost linearly with problem size, makes the selective inference approach practically feasible for DNNs.

Summary by alphaXiv.

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