Awesome

This repository contains the code for the paper "Characterizing the spectrum of the NTK via a power series expansion" (ICLR 2023). If you use this code in a paper, please cite this paper using the bibtex reference below

@inproceedings{murray2023characterizing,
    title={Characterizing the spectrum of the {NTK} via a power series expansion},
    author={Michael Murray and Hui Jin and Benjamin Bowman and Guido Montufar},
    booktitle={The Eleventh International Conference on Learning Representations },
    year={2023},
    url={https://openreview.net/forum?id=Tvms8xrZHyR}
}

This project is organized into two directories.

• spec/
  • Computes the NTK spectrum using Pytorch
• powerseries/
  • Computes the NTK power series

The directories spec/ and powerseries/ both have their own README files to explain the project organization

The dependencies for the project are specified in the environment.yml file. If you use Anaconda you can construct the environment from this file.