Discussion of: Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications

David:

• Equation 1: This assumes there is a root and that it is not too far from the point of interest. Why not choose a point that represents some average or trusted prediction value? e.g. use a nearby point that was part of the training set and for which the model predicts an accurate value? partially addressed at boom of pg. 249
• For voice interpretation, they switched to frequency basis. Something one would have guessed ahead of time. Is there a way you could know to make such a transformation based on their initial attempt using just samples?