CNA -- ICML 2020

Official implementation of "Abstraction Mechanisms Predict Generalization in Deep Neural Networks", published in ICML 2020, an unsupervised generalization gap metric. Paper link: https://proceedings.icml.cc/static/paper_files/icml/2020/6109-Paper.pdf

For correspondence and questions:

• Name: Alex Gain
• Email: [email protected]
• LinkedIn: https://www.linkedin.com/in/alex-gain-184305195/

Fastest Usage

1. Open CNA_Notebook.ipynb in Google Collaboratory (https://colab.research.google.com/drive/15tcbw3KO6B4F-yBO-K5RJtkl9s1KFDgw?usp=sharing).
2. Duplicate copy in your Google Drive File -> Save a copy in Drive.
3. Run all cells Runtime -> Run All Cells.

Installation, Usage, and Minimal Working Example

1. Add cna_utils.py to your root directory.
2. Get dataset loaders and trained network:

train_loader, test_loader = get_dataset_loaders() # <-- Your dataset loader method goes here
net = torch.load(SAVED_NET_PATH_HERE) # <--- Loading your saved network

3. Getting CNA and CNA-Margin values:

from cna_utils import *
cna_val = CNA_all(train_loader, net)
cna_m_val = CNA_M(train_loader, test_loader, net, args.flatten)

For minimal working example in cna-icml2020/usage run:

python get_cna.py --net_path saved_nets/net0.pt --dataset mnist --flatten=1 --cuda=1

Example Results From CNA_Notebook.ipynb

Saved models and Results

All saved models (over 600 saved network states) and results (including margin comparison metrics) can be found here:

• https://drive.google.com/drive/folders/1luB00h5egx-s1OwYav284HSwccfO7H2M?usp=sharing

Legacy code

For posterity, in cna-icml2020/legacy_code, the original code for producing the experimental results from the paper is supplied. This is not updated or maintained. All functionality will be refactored into this new repository for more convenient usage. It is recommended to make use of the non-legacy code for future work.

TO-DO left for refactoring legacy code into cna-icml2020/usage:

• Shuffled label dataset loaders.
• Margin-based generalization metrics (can also be accessed here: https://github.com/bneyshabur/generalization-bounds).