I found it easier to understand the idea by skimming the introduction in the paper [1]: > TCAV uses directional derivatives to quantify the model prediction’s

I found it easier to understand the idea by skimming the introduction in the paper [1]: > TCAV uses directional derivatives to quantify the model prediction’s sensitivity to an underlying high-level concept [...] For instance, given an ML image model recognizing zebras, and a new, user-defined set of examples defining ‘striped’, TCAV can quantify the influence of striped concept to the ‘zebra’ prediction as a single number. In addition, we conduct statistical tests where CAVs are randomly re-learned and rejected unless they show a significant and stable correlation with a model output class or state value. [1]: https://arxiv.org/abs/1711.11279 https://arxiv.org/abs/1711.11279