Generative Adversarial Network
A Generative Adversarial Network (GAN) is a generative modeling framework in which two neural networks — a generator and a discriminator — are trained in competition, with the generator learning to produce increasingly realistic synthetic…
Definition
A Generative Adversarial Network (GAN) is a generative modeling framework in which two neural networks — a generator and a discriminator — are trained in competition, with the generator learning to produce increasingly realistic synthetic data as the discriminator learns to detect it.
Overview
Introduced by Ian Goodfellow and colleagues in 2014, a GAN consists of two networks trained simultaneously in a minimax game. The generator takes random noise as input and tries to produce synthetic data — commonly images — realistic enough to fool the discriminator. The discriminator is trained as a binary classifier, given a mix of real training data and the generator's synthetic output, and tries to correctly distinguish real from fake. As training progresses, the generator improves at producing convincing fakes, and the discriminator improves at catching them, ideally converging to a point where the generator produces samples indistinguishable from real data. This adversarial setup made GANs notoriously difficult to train, prone to instability and a failure mode called mode collapse, where the generator learns to produce only a limited variety of outputs that reliably fool the discriminator rather than capturing the full diversity of the training data. Numerous architectural and training refinements followed the original paper — including DCGAN, which established stable convolutional architectures for image generation, and StyleGAN, which introduced style-based generation enabling fine control over image attributes like pose, texture, and expression, achieving some of the most photorealistic synthetic faces of its era. GANs were, for several years, the dominant approach for high-fidelity image generation, powering applications from deepfakes to art generation to data augmentation, and popularizing techniques like image-to-image translation (as in CycleGAN, which translates images between domains without paired training examples). More recently, diffusion models have overtaken GANs as the leading approach for state-of-the-art image and video generation, offering more stable training and better sample diversity, though GANs remain valuable for applications requiring very fast, single-pass generation, since a trained GAN generator produces an output in one forward pass, unlike the iterative denoising process diffusion models require.
Key Concepts
- Two competing networks: a generator and a discriminator, trained adversarially
- Generator learns to produce synthetic data that fools the discriminator
- Discriminator learns to distinguish real training data from generated fakes
- Trained as a minimax game, converging toward realistic generated output
- Prone to training instability and mode collapse
- DCGAN and StyleGAN introduced major architectural improvements for images
- Generates output in a single fast forward pass, unlike iterative diffusion models
- Introduced by Ian Goodfellow and colleagues in 2014