GAN (Generative Adversarial Network)
A deep learning framework that trains two networks - a generator and a discriminator - in an adversarial manner to produce images indistinguishable from real ones.
A GAN (Generative Adversarial Network) is a generative framework proposed by Goodfellow et al. in 2014. It pits a Generator against a Discriminator, driving both to improve through competition until generated images become indistinguishable from real data.
Training mechanism:
- Generator (G): Takes random noise z and produces synthetic images, aiming to fool the discriminator
- Discriminator (D): Binary classifier judging whether images are real or generated
- Minimax game: G maximizes D's error while D maximizes accuracy, driving improvement
- Nash equilibrium: Ideally converges to D outputting 0.5 for all inputs
Landmark architectures:
DCGAN(2015): GAN with architectural constraints on convolutional networksPix2Pix(2016): Conditional GAN for paired image translationCycleGAN(2017): Unpaired style transfer without matched dataStyleGAN(2018): Style-based generator for high-resolution facesStyleGAN3(2021): Eliminated aliasing in the generator
Applications include super-resolution (SRGAN, ESRGAN), inpainting, and data augmentation. Training instability (mode collapse, vanishing gradients) is addressed by Wasserstein distance and spectral normalization.