So Derpina tried to visualize her face that was in the dataset after a few convolutions...

As a disclaimer, Derpina intentionally set the output to two classes: “pretty girl” or “not so pretty girl”. We do not know if she intentionally manipulated the labels that led to her findings.