Figures for "Near-Interpolators"

Official code repository for "Near-Interpolators: Rapid Norm Growth and the Trade-Off between Interpolation and Generalization.

See below for instructions on reproducing the figures.

Note: this repository uses code from the neural-tangent-kernel-UCI repository.

Instructions: Figure 1 to 4

For each figure panel in the paper, there are three associated directories inside the directory figure_and_experiment_code:

• NAME-plot/
• NAME-run1/
• NAME-run2/

For instance, NAME can be one of figure1-left, figure1-right, figure4-bottom-left, and so on.

To recreate the figures, for say figure1-left do the following:

cd figure1-left-run1
source run_experiment.sh
cd ..
cd figure1-left-run2
source run_experiment.sh
cd ..
cd figure1-left-plot
jupyter run plot_figure.ipynb

The figure will be in figure1-left-plot/outputs/

Instructions: Figure 5 and 6

Navigate to the respective folder figure5 or figure6.

Run the first block

!pip install git+https://github.com/treforevans/uci_datasets.git

if running for the first time.

Run all cells.