Salient Imagenet

Explore Salient Imagenet

In the EXPLORE tab, you can view all 5000 annotated class-feature pairs, organized by class.
We present the same visualizations as shown to crowd workers to annotate each pair, and all their responses for transparency.

Navigate to ABOUT to learn more about the visualizations and dataset construction.

Use Salient Imagenet

Obtain the dataset from the DOWNLOAD tab.

We demonstrate usage and summarize findings in the ANALYSIS tab, providing quick access to starter code.

Salient Imagenet can be used to:

Evaluate pre-trained models off the shelf, using a noise analysis and a novel relative sensitivity metric.

Train models to predict for the right reasons, following the new paradigm Core Risk Minimization (CoRM).

Cite Salient Imagenet

Please cite the following papers if Salient Imagenet is of use to you.


  @misc{
    singla2022core
    author = {Singla, Sahil and Moayeri, Mazda and Feizi, Soheil},
    title = {Core Risk Minimization using Salient ImageNet},
    publisher = {arXiv},
    year = {2022},
    url = {https://arxiv.org/abs/2203.15566}
  }
  @inproceedings{
    singla2022salient,
    title={Salient ImageNet: How to discover spurious features in Deep Learning?},
    author={Sahil Singla and Soheil Feizi},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=XVPqLyNxSyh}
  }

Salient Imagenet Dataset

Example Salient Imagenet Features

Explore Salient Imagenet

Use Salient Imagenet

Cite Salient Imagenet

Contact

Sahil Singla

Mazda Moayeri

Soheil Feizi