Clarifying experiments in Section 4.1.2 Empirical Deep Dream

[expectopatronum]

Hi everyone, thanks for this great paper, I am really looking forward to using TCAV for my own research. Currently I am trying to make sense of the experiment described in Section 4.1.2 and have a few questions.

1. What does it mean to maximally activate a CAV? Does it mean to maximize the TCAV score for one example? In the context of activation maximization I assume that you start with a randomly generated image, and modify the image until the TCAV score is maximized. 1a. If this is correct, it means that we are computing the TCAV score for one example at a time. Does this also mean we have to do this "CAV maximization" per class (e.g. zebra), because we have to compute the TCAV score for a class, right?

This doesn't really make sense to me so far, so I might be completely wrong.

2. Is there any source code for this experiment to help me understand it better?

Best regards Verena