GANs consist of a generator and discriminator trained in a minimax game. The generator learns to produce samples from the real data distribution.
Key Uses:
Data augmentation for imbalanced medical datasets and privacy-preserving synthetic data.
Hacks:
Use Wasserstein GAN with Gradient Penalty (WGAN-GP) for training stability. Apply StyleGAN’s style-based generator for fine-grained control. Conditional GANs (cGANs) for class-specific generation. Mixup and CutMix techniques to further augment synthetic data quality.
References:
Good fellow, I., et al. (2014). Generative Adversarial Nets. NeurIPS.