GANs: when neural networks learned to invent images (2014)
In June 2014 Ian Goodfellow and colleagues introduced generative adversarial networks: two neural networks trained against each other. The idea made realistic image synthesis possible and raised the question of fakes early on.
In 2014, at the Montreal pub Les 3 Brasseurs, Ian Goodfellow sketched an idea and coded it into the early hours of the morning. It worked on the first try. On June 10, 2014, a paper titled Generative Adversarial Networks appeared on arXiv, written by a team of eight researchers at the Université de Montréal, including Yoshua Bengio. They introduced generative adversarial networks (GAN): two neural networks trained against each other so one learns to produce data the other cannot tell from real examples.
Counterfeiters versus police
The setup pairs two models. A generator G captures the data distribution and produces samples. A discriminator D estimates the probability that a sample came from the training data rather than from G. G is trained to maximize the probability that D makes a mistake. The authors framed this as a minimax two-player game; in the space of arbitrary functions, they showed there is a unique solution where G recovers the training distribution and D equals 1/2 everywhere.
The paper’s analogy is vivid: G is like a team of counterfeiters trying to pass fake money, and D is the police trying to catch them. The contest pushes both sides until the fakes are indistinguishable from the real thing. When G and D are multilayer perceptrons, the whole system can be trained with backpropagation, without Markov chains or unrolled inference networks. Early experiments used MNIST handwritten digits, the Toronto Face Database, and CIFAR-10.
From blurry faces to photorealism
Earlier attempts to generate images with neural nets often produced blurry results. Follow-on work made GANs far more convincing. In November 2015, Alec Radford, Luke Metz, and Soumith Chintala published a deep convolutional GAN for unsupervised learning of image features. Nvidia later generated images of invented celebrities. In December 2018, NVIDIA’s StyleGAN project showed generated faces with the note “These people are not real.”
Yann LeCun called GANs “the coolest idea in deep learning in the last 20 years.” The line from a pub-night prototype to photorealistic faces runs through deep learning in computer vision.
The flip side: fakes
Synthetic media also raised alarms early. MIT Technology Review quoted Hany Farid of Dartmouth: “We’re fundamentally in a weak position.” The article warned that “GANs didn’t create this problem, but they’ll make it worse.”
Regulators took notice. The EU AI Act defines deepfakes (Article 3(60)) and requires deployers of systems that generate or manipulate them to disclose that content is artificially created or manipulated (Article 50(4)). See the EU AI Act article for how that fits into the wider rules.
Why it matters
The ACM cited Bengio’s GAN work with Goodfellow when awarding the 2018 Turing Award, calling it a revolution in computer vision and computer graphics. In 2021, Dhariwal and Nichol showed that diffusion models could beat GANs on image synthesis, yet the GAN idea remains the starting point of generative AI for images. At NeurIPS 2024, the original paper received a Test of Time award alongside Seq2Seq; by then it had been cited more than 85,000 times.
Two years before, AlexNet had taught networks to recognise images; GANs taught them to invent new ones. The contest between generator and discriminator opened a path that still shapes how we think about real and synthetic media online.