Summit Notebooks

Summit is an interactive system that scalably and systematically summarizes and visualizes what features a deep learning model has learned and how those features interact to make predictions. This repository contains the python notebooks used to generate the data used in the Summit visualization.

For the main Summit repo, go to https://github.com/fredhohman/summit.

Main notebooks:

• activation-matrices.ipynb: generate aggregated activation matrices
• influence.py: generate aggregated influence matrices
• activation-matrices-to-json.ipynb: combine activation matrices per class into json format
• attribution-graph.ipynb: generating class attribution graphs
• feature-vis-sprite-to-images.ipynb: split feature visualization sprites to single images

Experimental notebooks:

• top-channels-used-per-layer.ipynb: analysis for determining which channels were used the most by all classes for all layers

Live Demo

For a live demo, visit: fredhohman.com/summit

Resources

We used the following ImageNet metadata:

• https://github.com/google/inception/blob/master/synsets.txt
• https://gist.github.com/aaronpolhamus/964a4411c0906315deb9f4a3723aac57
• https://gist.github.com/yrevar/942d3a0ac09ec9e5eb3a

License

MIT License. See LICENSE.md.

Citation

@article{hohman2020summit,
  title={Summit: Scaling Deep Learning Interpretability by Visualizing Activation and Attribution Summarizations},
  author={Hohman, Fred and Park, Haekyu and Robinson, Caleb and Chau, Duen Horng},
  journal={IEEE Transactions on Visualization and Computer Graphics (TVCG)},
  year={2020},
  publisher={IEEE},
  url={https://fredhohman.com/summit/}
}

Contact

For questions or support open an issue or contact Fred Hohman.