Generative Adversarial Networks
Every generative model so far has needed an explicit loss function measuring "how good is this sample." GANs replace that explicit loss with a second trained network, whose entire job is to learn what "real" looks like — and the generator improves purely by trying to fool it.