Reproduce “Valid P-Value for Deep Learning-Driven Salient Region” on OpenResearch

About this paper

Daiki Miwa, Vo Nguyen Le Duy, Ichiro Takeuchi

Researchers from Nagoya Institute of Technology, RIKEN, and Nagoya University developed a method that provides statistically valid p-values for deep learning-identified salient regions, rigorously controlling the false positive rate. This approach demonstrates superior statistical power compared to conservative alternatives while accurately discerning genuinely significant explanations in both synthetic and real-world scenarios.

Key insights

• Saliency map selection inherently introduces bias, causing naive p-values to incorrectly indicate significance and fail to control false positive rates.
• The piecewise linear nature of common CNN layers allows the complex conditional data space for selective inference to be efficiently characterized as a one-dimensional line.
• A balanced conditional selective inference approach provides valid p-values with higher statistical power compared to conservative multiple testing corrections or over-conditioning strategies.

Summary by alphaXiv.