Companion code for the paper Theoretical characterization of uncertainty in high-dimensional linear classification

Usage

The required packages are listed in requirements.txt

The folder Code/experiments contains two jupyter notebook that reproduce each one figure of the paper :

• lambda-fixed-experiments.ipynb plots the joint density of the uncertainties of Logistic regression / Bayes, and Logistic regression / Teacher (both theoretical prediction and experimental estimation)
• lambda-min-error-experiments.ipynb plots the calibration of ERM at a fixed level, as a function of the sampling ratio.

Acknowledgement

The folder Libraries/GCMProject contains a modified version of the repository https://github.com/IdePHICS/GCMProject, and the folder Libraries/state-evolution-erm-logistic contains code provided by github.com/benjaminaubin