#sanity-checks

Question 1: Single choice

A saliency method produces almost the same maps for a trained CNN and for the same architecture with randomly re-initialised weights. What does this sanity check tell you?

The method ignores what the model learned, so it is useless for debugging1
The model is robust, because its explanations survive changes to its weights2
The method is faithful, since stable explanations are a sign of reliability3
The randomly initialised network must already compute the same features4

Timed test