Other meanings of Generative adversarial network
Artificial intelligence
A generative adversarial network (GAN) is a class of machine-learning framework in which two neural networks — a generator and a discriminator — compete in a zero-sum game to produce realistic synthetic data. The generator learns to create samples that mimic a training distribution, while the discriminator tries to distinguish real from generated samples. This adversarial process drives both networks to improve, enabling GANs to generate high-quality images, audio, and text. Introduced by Ian Goodfellow and colleagues in 2014, GANs have become a cornerstone of generative modeling and found applications in image synthesis, style transfer, data augmentation, and medical imaging.
The generator G maps a random noise vector z to a synthetic sample, while the discriminator D outputs a probability that its input is real rather than generated. During training, G and D are updated alternately: D maximizes the log-likelihood of correctly classifying real and fake samples, and G minimizes the log-likelihood of D being correct — effectively playing a minimax game1. Standard GANs use fully connected layers, but later variants incorporate convolutional (DCGAN) or residual architectures. Training is notoriously unstable due to issues such as mode collapse, where the generator produces limited varieties of outputs, and vanishing gradients, often addressed by alternative loss functions (e.g., Wasserstein distance, least squares) or regularization techniques2.
Ian Goodfellow conceived GANs in 2014 while at the University of Montreal, publishing the seminal paper with Yoshua Bengio and others1. The idea grew out of a discussion about generating images using adversarial examples. Early adopters rapidly improved the framework: in 2015, Alec Radford introduced DCGAN, which stabilized training with convolutional layers7; in 2017, Wasserstein GAN (WGAN) used the Earth mover’s distance to mitigate mode collapse2. Progress in high-resolution synthesis came with progressive growing of GANs (PGGAN) in 2017, followed by StyleGAN (2018) and StyleGAN2 (2020), which achieved unprecedented photorealism3. Conditional GANs (cGANs) introduced label conditioning, enabling controlled generation4.
GANs are widely used in image-to-image translation (e.g., CycleGAN for unpaired domain transfer), super-resolution (SRGAN), and text-to-image synthesis (StackGAN). In medical imaging, they generate synthetic MRI or CT scans to augment training data, improving diagnostic models6. GANs also power deepfakes, raising ethical concerns about misinformation. In drug discovery, they generate molecular structures with desired properties. Other applications include video frame prediction, speech enhancement, and adversarial defense. The GAN Zoo compiles hundreds of variants, each tailored to specific tasks5.
Beyond image generation, GANs have been used to create photorealistic 3D models (e.g., HoloGAN) and to generate tabular data for privacy-preserving datasets. A niche but influential variant, the Coupled GAN (CoGAN), learns joint distributions across domains without paired data. Mode collapse, though well-known, is often mitigated by minibatch discrimination or unrolled GANs. The GAN framework has also inspired the development of adversarial attacks on classifiers, showing that the same adversarial training can improve robustness. Another obscure fact: the original GAN paper used a generator that produced 32×32 pixel images, and the first training sessions took days on a single GPU1.
GANs are also studied in the context of adversarial robustness and fairness, but these remain active research areas.
Help improve the encyclopedia. Reports go straight to the site manager.