Training More Robust Models through Creation and Correction of Novel Model Errors

This repository demonstrates how to run and use Defuse - a method to train more robust models through creation and correction of novel model errors. Given a trained image classification model (classifier), Defuse generates realistically looking images which are incorrectly predicted by the classifier. Further,Defuse groups images into high-level model bugs and efficiently corrects them. With Defuse users are able to identify scenarios under which their model would fail and subsequently train a more robust model. Defuse works in three steps: identification, distillation, and correction.

Security

See CONTRIBUTING for more information.

License

This library is licensed under the CC BY-NC 4.0 License. See the LICENSE file.

Citation

Please use the following citation when publishing material that uses our code:

@article{slack2021defuse,
      title={Defuse: Training More Robust Models through Creation and Correction of Novel Model Errors},
      author={Dylan Slack and Nathalie Rauschmayr and Krishnaram Kenthapadi},
      year={2021},
      url={https://xai4debugging.github.io/files/papers/defuse_training_more_robust_mo.pdf},
      journal={XAI4Debugging Workshop @ NeurIPS},

}