Generative adversarial network

A generative adversarial network (GAN) is a type of generative model in machine learning that uses two neural networks contesting against each other to generate new data. Key characteristics:

The generator network creates synthetic data from noise.
The discriminator network evaluates if data is real or fake.
The two networks train via a minimax game to fool one another.
Can generate images, audio, video, and other multimedia.
Powerful for accurately modeling target distributions.

GAN architectures include:

Vanilla GAN with fully connected networks.
Deep convolutional GANs for image synthesis.
Conditional GANs that incorporate auxiliary information.
Cyclic GANs with constraint and reconstruction losses.

Applications of GANs include creating photorealistic media, data augmentation, image-to-image translation, and domain transfer learning.

GAN training can be unstable and prone to issues like mode collapse. But overall, GANs represent a major advance in generative modeling and unsupervised learning.