Key Messages

• Sparse neural pathways encode critical input features. The pruning obective does not identify these critical pathways.
• The paper proposes a feature attribution (saliency map) method, “Pathway Gradient”, by leveraging critical pathways.

Resources

The paper is available on CVPR2021 Open Access and on arXiv.

View the poster

Watch the video on YouTube

Check the Code on GitHub

Citation

Please cite the work using the below BibTeX (also available on the Open Access link above)

@InProceedings{Khakzar_2021_CVPR,
   author    = {Khakzar, Ashkan and Baselizadeh, Soroosh and Khanduja, Saurabh and Rupprecht, Christian and Kim, Seong Tae and Navab, Nassir},
   title     = {Neural Response Interpretation Through the Lens of Critical Pathways},
   booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
   month     = {June},
   year      = {2021},
   pages     = {13528-13538}
   }

or this one (available on arXiv):

@article{Khakzar2021NeuralRI,
  title={Neural Response Interpretation through the Lens of Critical Pathways},
  author={Ashkan Khakzar and Soroosh Baselizadeh and Saurabh Khanduja and C. Rupprecht and Seong Tae Kim and N. Navab},
  journal={ArXiv},
  year={2021},
  volume={abs/2103.16886}
}

Contact

For inquiries and feedback please contact Ashkan Khakzar (ashkan.khakzar@tum.de). We would be happy to help and we appreciate your feedback.