Generative Adversarial Networks

GAN Fundamentals

GANs generate synthetic data that looks real.

Generator and Discriminator

Generator: creates fake samples. Discriminator: distinguishes real from fake. Train simultaneously.

Adversarial training: min-max game. Generator wants to fool discriminator.

Training

Generator loss: fool discriminator. Discriminator loss: correctly classify. Balance training important.

Mode collapse: generator produces limited variety. WGAN, spectral normalization address this.

Applications

Image generation. Data augmentation. Style transfer. Image-to-image translation.

Key Takeaways

1. Generator and discriminator compete
2. Training is adversarial game
3. Various GAN architectures exist