Repository for the paper Gaussian Universality of Linear Classifiers with Random Labels in High-Dimension.
An illustration of the universality in high-dimension: The training loss is shown as function of the number of samples n per input dimension p at vanishing regularization. In the left panel the square loss, in the right panel the hinge loss. The black solid line represents the outcome of the theoretical calculation for iid Gaussian inputs, while dots refer to numerical simulations on different full-rank datasets.
In this repository we provide the code and some guided example to help the reader to reproduce the figures of the paper [1]. The repository is structured as follows.
| File | Description |
|---|---|
/theory |
Solver for the fixed point equations . Add description of files. |
/simulations |
Finite instance simulations. Add description of files. |
The notebooks are self-explanatory.
[1] Gaussian Universality of Linear Classifiers with Random Labels in High-Dimension, Federica Gerace, Florent Krzakala, Bruno Loureiro, Ludovic Stephan, Lenka Zdeborová, arXiv:2205.13303[stat.ML]