Home

Awesome

Welcome to Barlow

Barlow is a tool for identifying the failure modes for a given neural network. To achieve this, Barlow first creates a group of images such that all images in the given group have the same predicted class or the label. Then, it uses a (possibly separate) robust model to extract the embedding for all images in the group. The embedding encodes the human-interpretable visual attributes present in the image. It then learns a decision tree using this embedding that predicts failure on the neural network being inspected for failure.

Prerequisites

Setup

Running on ImageNet classes

Running on custom dataset

images heatmaps attacks

Citation

@inproceedings{singlaCVPR2021,
  title     = {Understanding Failures of Deep Networks via Robust Feature Extraction},
  author    = {Sahil Singla and Besmira Nushi and Shital Shah and Ece Kamar and Eric Horvitz},
  booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition, {CVPR} 2021},
  publisher = {Computer Vision Foundation / {IEEE}},
  year      = {2021},
  url       = {https://openaccess.thecvf.com/content/CVPR2021/papers/Singla_Understanding_Failures_of_Deep_Networks_via_Robust_Feature_Extraction_CVPR_2021_paper.pdf},
}