Skip to content

Generative Adversarial Networks (GANs)

Authors: Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, et al. Year: 2014 Venue: NeurIPS 2014 Citations: 40,000+

Summary

Introduces Generative Adversarial Networks (GANs), two competing neural networks: generator creates fake data, discriminator distinguishes real vs. fake. Enables training of deep generative models without explicit likelihood.

Key Concepts

  • Generator Network: Creates synthetic data from random noise
  • Discriminator Network: Distinguishes real vs. generated data
  • Adversarial Training: Alternates between generator and discriminator updates
  • Minimax Game: Zero-sum game between generator and discriminator
  • Implicit Distributions: Generates data without modeling explicit probability
  • Mode Coverage: Generator learns full data distribution

Impact

  • 40,000+ citations
  • Revolutionary framework for generative modeling
  • Inspired StyleGAN, BigGAN, Pix2Pix, CycleGAN, and 1000s of variants
  • Foundation of modern generative AI before diffusion models
  • Enabled image synthesis, style transfer, data augmentation
  • Influenced Transformers and attention mechanisms
  • Changed landscape of generative modeling research

Key Results

  • Generated realistic synthetic images
  • Learned to synthesize complex visual data
  • Novel training approach without explicit likelihood
  • Inspired many architectural improvements
  • Variational Autoencoders (Kingma & Welling, 2014)
  • StyleGAN (Karras et al., 2018)
  • Conditional GANs (Mirza & Osindski, 2014)
  • Diffusion Models (Ho et al., 2020)