Part IV · Ch. 05 — GAN (Generative Adversarial Network)

draft

Connectionist · Generation · Generator vs. discriminator. Two models compete — one generates, the other judges — until realistic samples are produced. Card: ../02-types-of-ai/02-connectionist.kmd.

🎨 Figure F-IV.5.0 — The generator-discriminator duel. Brief: two blocks facing each other: a generator producing images, a discriminator giving a "real/fake" verdict; a gradient arrow feeding back to the generator.

GAN — generator vs. discriminator

1. Definition and short history

A generative model trained by an adversarial game (Goodfellow, 2014). It dominated image generation until diffusion took over (~2021). Lineage: DCGAN → StyleGAN (faces) → pix2pix/CycleGAN (image translation). See Part III, era 5.

2. Foundations

  • Game theory — Nash equilibrium of a minimax game.
  • Probability — approximates the data distribution without explicit density.
  • Optimization — unstable adversarial training (equilibrium, not minimum).
  • Information theory — divergences (JS, Wasserstein) as the objective.

3. Algorithms and architectures

  • Generator — latent noise → sample.
  • Discriminator — distinguishes real from generated.
  • Objective — minimax; WGAN (Wasserstein) stabilizes; gradient penalty.
  • Variants — conditional (cGAN), StyleGAN (style control), CycleGAN

    (unpaired).

4. Inputs

  • Hardware: GPU; fast inference (1 step, unlike diffusion).
  • Data: images of the target domain (faces, landscapes).
  • Data structures: tensors; structured latent space (StyleGAN).
  • Systems: PyTorch; training-stabilization tricks.

5. Specialized life cycle

Stage Specialization
0 Problem Fast 1-step generation, style editing, domain translation
1 Data Domain images, alignment/curation
2 EDA Dataset diversity (risk of mode collapse)
3 Modeling Architecture (StyleGAN/WGAN), objective, conditioning
4 Training Adversarial; monitor mode collapse and instability
5 Evaluation FID, generation precision/recall, human evaluation
5.5 Acceptance Artifact detection, deepfake safeguards
6 Production 1-step inference (fast); editing in latent space
7 Monitoring Quality, abuse (deepfakes)
8 Retraining New styles/domains
9 Governance Deepfakes, consent, watermark

6. Capabilities, modes and modalities

Artistic/visual: fast generation, fine editing of attributes (StyleGAN), super-resolution, image→image translation. Also audio (HiFi-GAN vocoders).

7. Limits, risks and ethics

Unstable training; mode collapse (low diversity); deepfakes and fraud; largely surpassed by diffusion in quality/diversity, but wins in speed (1 step) and as a vocoder.

8. State of the art and examples

StyleGAN3, GigaGAN; HiFi-GAN (TTS vocoder); niches where single-step inference matters.