arXiv:1703.01365

[refraction-ray]

• Link
• Title: Axiomatic Attribution for Deep Networks
• Keywords (optional):
• Authors (optional):
• Reason (optional):
• Summary (optional):
• Ideas (optional): Just integrate (by discrete sum in application) the gradient in each direction of deviation from a baseline input giving zero score. Then each component represent how much it contributes to the final prediction. And we can locate the important feature here.
  Implementation by tf: https://www.tensorflow.org/tutorials/interpretability/integrated_gradients