Code repository for the paper "The Nuclear Route: Sharp Asymptotics of ERM in Overparameterized Quadratic Networks"
Overview of the main files:
visualise.ipynb
: Notebook that produces the figures in the paper starging from the data indata/
and saving them infigs/
.theory.py
: Main file containing all the implementation of the theory except the limit of small rank. An example of how to use it is given in theexample_theory.py
notebook.{ALGO}_code.py
: Scripts that run the experiments for different algorithms. We have implementations with GD, LBFGS (both using PyTorch) and CVXPY.small_rank_Mathematica.nb
: Mathematica notebook that produces data for the limit of small rank. We recommend to run it in a Mathematica 13.0
Dependencies: We used Python 3.10.15 and the following packages:
- cvxpy==1.3.0
- numpy==2.2.6
- pandas==2.2.3
- scipy==1.15.3
- torch==1.12.1
- tqdm==4.67.1