/DeepnetHessian

Measurements of Three-Level Hierarchical Structure in the Outliers in the Spectrum of Deepnet Hessians (ICML 2019)

Primary LanguageJupyter Notebook

Measurements of Three-Level Hierarchical Structure in the Outliers in the Spectrum of Deepnet Hessians (ICML 2019)

For inquiries, please contact papyan@stanford.edu

Features

  • Download 240 models from a GCP bucket:
    • VGG11, ResNet18, DenseNet40
    • MNIST, FashionMNIST, CIFAR10, CIFAR100
    • Various sample sizes
  • Approximate spectrum using the method presented in https://arxiv.org/abs/1811.07062
  • Compute three-level hierarchical structure using the method proposed in this paper

Usage

See Jupyter notebook main.ipynb

Resources

arXiv paper