Wikipedia's GAN primer traces origin to Goodfellow, 2014

The Wikipedia entry on generative adversarial networks locates the technique's starting point with Ian Goodfellow and his colleagues at Université de Montréal in June 2014. Two neural networks compete, the article explains, 'in the form of a zero-sum game, where one agent's gain is another agent's loss.' A generator maps from a latent space toward the real data distribution; a discriminator tries to tell real samples from synthetic ones.

The entry is blunt about what breaks. GANs 'often suffer from mode collapse where they fail to generalize properly, missing entire modes from the input data,' and even BigGAN, which the entry dates to 2019, 'could not avoid mode collapse.' Vanishing gradients and unstable convergence round out the three documented failure modes.

Around that core sit Conditional GANs, DCGAN, Self-Attention GAN, Wasserstein GAN, StyleGAN and BigGAN, with Inception Score, Fréchet Inception Distance and LPIPS as the standard evaluation metrics. Two of the AI researchers we follow passed the entry around recently, which reads less as news than as a note on which foundations stay on the reading list.