Benign Overfitting in Linear Regression with Separable Sample Covariance Matrices

This repo contains simulation files for an ongoing project on Benign Overfitting in Linear Regression with Separable Sample Covariance Matrices supervised by Professor Yanrong Yang and Professor Hanlin Shang.

There are delays in the documentation of the project due to active development. We will update the documentation once experiments are finalized.

Environment Setup

To set up the environment, run

conda create -n bo python==3.9.16

and install the packages in requirements.txt using

pip install -r requirements.txt

File Structure

Simulation Results

The simulation results are in the folder results. For large simulation records with size over 100MB, we include a zip file for the results. To unzip the file, run

Interative Plots

To see the plots interactively, please run the plot.ipynb in visualization folder using jupyter notebook or jupyter lab.

Jupyter Widgets need to be activated in the jupyter lab environment.

Simulation Code

Simulations are written in R and Python. The Python version uses Just-In-Time Compilor to speed up the simulation. The R version is also available. Both code are in the folder simulation.

To replicate our simulation, run

sh run_simulation.sh

Parameters can be changed in the file run_simulation.sh.

fastmath is used in the simulation code. To benefit from fastmath, please run

conda install -c numba icc_rt

(Read more here about SVML)

Tests

A small number of tests are written to test the simulation code. To run the tests, run

python -m unittest simulation/tests/test_benign_overfitting.py