Réseaux antagonistes génératifs (GAN)
Définition
GANs entraînent un générateur et un discriminateur dans un jeu: the generator produces samples; the discriminator tries to distinguish them from real data. Training pushes the generator toward realistic outputs.
They étaient l'approche générative dominante avant diffusion models. Compared to VAEs, GANs often produce sharper images but training can be unstable (mode collapse, discriminator/generator balance). Still used for style transfer, data augmentation, and some image editing.
Comment ça fonctionne
Generator: Takes noise (vecteur aléatoire) and outputs a fake sample (par ex. image). Discriminator: Receives real data and fake sample, outputs real or fake (or a score). Training is a min-max game: the generator tries to maximize the discriminator’s loss (fool it), the discriminator tries to minimize it (tell real from fake). In practice you alternate gradient steps. Variants (DCGAN, StyleGAN, etc.) use better architectures and training tricks (par ex. spectral norm, progressive growing) for stability and quality.
Cas d'utilisation
GANs are used for generative and discriminative tasks when you want adversarial training and sharp samples (images, audio, data aug).
- Image generation and editing (par ex. StyleGAN, face synthesis)
- Data augmentation and synthetic data for training
- Domain adaptation and style transfer