Explaining black-box deep neural models' predictions, behaviors, and performances through the unsupervised mining of their inner knowledge