Redes generativas de confrontación

Ian Goodfellow and colleagues introduced the GAN framework in 2014. In the original unconditional formulation, a generator maps sampled latent noise into synthetic data. A discriminator receives real training examples and generated samples, learning to distinguish their origins. The generator receives gradients through the discriminator and aims to make its outputs harder to distinguish. This original setup should not be generalized into a claim that every GAN generator is forbidden from receiving real images. Conditional and image-to-image formulations can provide other inputs: the pix2pix research, for example, learns mappings from input images to target images with paired training data. Practical GAN training can be difficult. Mode collapse describes inadequate variety when a generator concentrates on a limited range of outputs; realistic individual samples do not establish that the intended distribution is represented. The original theoretical analysis describes an ideal distribution-matching equilibrium. When the generator distribution matches the data distribution, the optimal discriminator assigns probability one-half to either origin. That result depends on the theoretical setup and does not guarantee convergence of a finite trained system. Conditioning adds another distinction. In the conditional GAN formulation by Mirza and Osindero, information such as a class label can be supplied to both generator and discriminator. It specifies the condition being modeled, while the generated content can still vary. A conventional feed-forward generator can produce a sample in one forward evaluation, but comparisons of latency or quality with another model must identify the actual architecture and workload.