🔍 CNN explainability

Why explainability?

Ways of explainability

• AM: Global Activation Map. Average feature map on the channel axis.
  • This shows a map all classes detected.
  • Video by Jeremy Howard
• CAM: Class Activation Map. Final feature map followed by a 1x1 convolution with the weights of the final FullyConnected.
  • This shows activation map per class.
  • Paper: Learning Deep Features for Discriminative Localization
  • Paper: Fully Convolutional Networks for Semantic Segmentation
  • Notebook deepcourse
• Grad-CAM: Final features multiplied by class gradients and the averaged.
  • paper
  • Keras example
• SmoothGrad
  • paper
• Integrated gradients
  • Keras example
• LIME
  • Paper “Why Should I Trust You?”: Explaining the Predictions of Any Classifier
• Extra: Distill: feature visualization
• Extra: Distill: building blocks

References

• link 1
• link 2
• Libraries
  • Captum by Pytorch
  • Lucid by Tensorflow