Corpus ID: 202782699

A Benchmark for Interpretability Methods in Deep Neural Networks

@inproceedings{Hooker2019ABF,
  title={A Benchmark for Interpretability Methods in Deep Neural Networks},
  author={Sara Hooker and Dumitru Erhan and Pieter-Jan Kindermans and Been Kim},
  booktitle={NeurIPS},
  year={2019}
}
  • Sara Hooker, Dumitru Erhan, +1 author Been Kim
  • Published in NeurIPS 2019
  • Computer Science, Mathematics
  • We propose an empirical measure of the approximate accuracy of feature importance estimates in deep neural networks. Our results across several large-scale image classification datasets show that many popular interpretability methods produce estimates of feature importance that are not better than a random designation of feature importance. Only certain ensemble based approaches---VarGrad and SmoothGrad-Squared---outperform such a random assignment of importance. The manner of ensembling… CONTINUE READING

    Create an AI-powered research feed to stay up to date with new papers like this posted to ArXiv

    34
    Twitter Mentions

    Citations

    Publications citing this paper.
    SHOWING 1-8 OF 8 CITATIONS

    Measuring and improving the quality of visual explanations

    VIEW 1 EXCERPT
    CITES METHODS

    Selectivity considered harmful: evaluating the causal impact of class selectivity in DNNs

    VIEW 1 EXCERPT
    CITES BACKGROUND

    References

    Publications referenced by this paper.
    SHOWING 1-10 OF 41 REFERENCES

    Deep Residual Learning for Image Recognition

    VIEW 6 EXCERPTS
    HIGHLY INFLUENTIAL

    Birdsnap: Large-Scale Fine-Grained Visual Categorization of Birds

    VIEW 10 EXCERPTS
    HIGHLY INFLUENTIAL

    ImageNet: A large-scale hierarchical image database

    VIEW 12 EXCERPTS
    HIGHLY INFLUENTIAL

    Axiomatic Attribution for Deep Networks

    VIEW 7 EXCERPTS
    HIGHLY INFLUENTIAL

    Evaluating the Visualization of What a Deep Neural Network Has Learned

    VIEW 4 EXCERPTS
    HIGHLY INFLUENTIAL

    Do Better ImageNet Models Transfer Better?