Reply to post:

Couldn't you use a variation of a GAN setup where you get the adversarial network to generate examples to fool the discriminator, while having a human supervisor ensure the GAN still produces images clearly recognizable as a cat?