Earlier quoted context omitted.
Why? Just use two separate datasets, one of images of adults and children, to learn age, the other of porn and non-porn, to learn what's "nudity" or "sex".
Knowledge transfer training is really hard. If you can do this (with different categories, for instance, determining what is nudity, then determining black people, to automatically classify porn with black actors automatically), it would be quite impressive. Results have nearly 5% error on CIFAR-10, which is full of clearly distinct categories - planes versus boats.
Can you explain (just a broad overview) why that wouldn't work? I mean, it's not like black porn actors are categorically different from black people in general... in particular, if you have an algorithm that can locate the face, you could just pass the face of the porn actor to the other algorithm to determine age/race/whatever...