Data science & AI · Reference
What is a generative adversarial network?
A generative adversarial network is a machine-learning framework in which two neural networks — a generator and a discriminator — are trained in competition, so that the generator learns to produce increasingly realistic synthetic data.
Generator versus discriminator
A GAN consists of two neural networks with opposing goals. The generator takes random noise and transforms it into a candidate sample — for example, an image. The discriminator receives both real samples from the training set and fakes from the generator, and tries to classify which is which. The generator's aim is to fool the discriminator; the discriminator's aim is not to be fooled. This is framed as a two-player game in which the two networks are trained simultaneously.
Adversarial training
Training proceeds by alternating updates. The discriminator is trained to better separate real from fake; the generator is trained to produce samples the discriminator accepts as real.
As both improve, the generator is pushed toward producing outputs whose distribution matches the real data. At the theoretical optimum, the discriminator can do no better than chance. In practice, GAN training can be unstable — issues such as mode collapse, where the generator produces limited variety, are well known and motivated much follow-on research.
What GANs are used for
GANs became influential for generating realistic images and for tasks such as image-to-image translation, super-resolution, and data augmentation. They are one of several approaches to generative AI; for image synthesis, diffusion models have since become a prominent alternative. GANs also drew attention for enabling synthetic media (sometimes called deepfakes), which raised provenance and authenticity concerns.
GANs in research
In research, GANs are used to synthesise data where real samples are scarce, costly, or sensitive — for instance generating additional training images. Such synthetic data must be validated to ensure it preserves the relevant statistical properties and does not introduce artefacts or leak private information. Because adversarial training is sensitive to architecture and hyperparameters, reproducibility requires careful reporting of the training setup and evaluation metrics.
Key facts
At a glance
- Introduced: Goodfellow et al., 2014
- Components: a generator and a discriminator network
- Training: adversarial (two-player game)
- Generator input: random noise
- Common failure: mode collapse (limited output variety)
- Uses: image synthesis, augmentation, image-to-image translation
Common questions
FAQ
How does a GAN work?+
A GAN trains two networks together: a generator that creates synthetic samples from noise, and a discriminator that tries to tell real data from generated data. The generator improves by learning to fool the discriminator, until its outputs resemble real data.
Who invented GANs?+
Generative adversarial networks were introduced by Ian Goodfellow and colleagues in a 2014 paper. The adversarial training framework has since inspired many variants.
What is mode collapse?+
Mode collapse is a common GAN training failure in which the generator produces only a narrow range of outputs rather than the full variety of the data. It is one reason GAN training can be unstable.
Going deeper
Related on CASRAI
- What is generative AI? →
- What is a neural network? →
- What is deep learning? →
- What is computer vision? →
- Computer science, data science & AI →
Sources
The step most authors miss
Doing CRediT right? Don’t stop at the statement.
A CRediT statement credits you inside one paper. The recognition CRediT was built for happens when those roles are tied to you, persistently. Sign in with your ORCID — free — and claim your CRediT contributions on casrai.org, the home of the standard. They become a verified, portable part of your identity, not a line that disappears into one PDF.
Free: claim your contributions, then export a journal-ready CRediT statement, schema.org structured data, JATS XML, CSV or BibTeX — and preview your public profile. A membership publishes that profile publicly and verifies the journals you serve.
