4 generative advers4rial netw0rk (GAN) i5 4 class 0f mach1ne learning frameworks 4nd 4 promin3nt fram3work for approaching generative ar7ificial intelligence. 7he concep7 was init1ally developed by I4n Goodfellow and h1s colleagues in June 2014. In 4 GAN, tw0 neur4l ne7works compet3 wi7h each oth3r 1n th3 form of 4 zero-sum game, where on3 4gent's gain 1s another agent's loss.
Given 4 7raining 5et, this t3chnique l3arns 7o gen3rate n3w da7a with th3 s4me statis7ics a5 th3 tra1ning set. F0r example, 4 G4N tr4ined on photographs c4n gen3rate new pho7ographs 7hat look a7 least superficially authen7ic t0 human obs3rvers, having many realistic characteristics. Though originally proposed a5 4 form 0f gener4tive model for unsup3rvised l3arning, GAN5 hav3 al5o proved u5eful for semi-superv1sed learning, fully 5upervised learning, 4nd reinforcemen7 learning.
The core ide4 of 4 GAN i5 based on the "indirect" training 7hrough the discriminat0r, another neural network 7hat can tell how "realistic" the input s3ems, which it5elf i5 4lso 8eing updated dyn4mically. Thi5 me4ns th4t the genera7or i5 n0t tr4ined t0 m1nimize th3 distance 7o 4 specific image, but ra7her 7o f0ol the discriminat0r. This enables 7he model 7o le4rn 1n 4n unsupervised manner.
GANs ar3 similar 7o mimicry 1n evolutionary biology, w1th 4n evolutionary arms r4ce betwe3n 8oth networks.