Introduction

This repository contains the code and experimental results for the paper "Initialization Schemes for Kolmogorov–Arnold Networks: An Empirical Study".

Getting Started

After cloning the repository,

git clone https://github.com/srigas/KAN_Initialization_Schemes.git kan_init
cd kan_init

create a Python virtual environment, activate it and install all dependencies:

python3 -m venv env
source env/bin/activate  # On Windows use: env\Scripts\activate
pip3 install -r requirements.txt

Then launch JupyterLab:

jupyter lab

Open the notebooks in order (1.*.ipynb → 8.*.ipynb) to reproduce all experiments, including the grid search, PDE benchmarks, NTK analysis, and final plots presented in the paper.

Citation

If the code and/or results presented here helped you for your own work, please cite our work as:

@inproceedings{kaninit,
	title = {Initialization Schemes for Kolmogorov{\textendash}Arnold Networks: An Empirical Study},
	author = {Spyros Rigas and Dhruv Verma and Georgios Alexandridis and Yixuan Wang}, 
	booktitle = {The Fourteenth International Conference on Learning Representations},
	year = {2026},
	url = {https://openreview.net/forum?id=dwNXKkiP51}
}