Generative Adversarial Networks

Ian Goodfellow et al. (2014)

Read paper

Why It Matters

Generator vs discriminator adversarial training. Enabled realistic synthesis, style transfer, deepfakes, data augmentation.

Key Ideas

1. Train a generator and discriminator in a minimax game where one produces samples and the other tries to distinguish real from fake.
2. Learn to sample from a complex data distribution without defining an explicit likelihood model.
3. Get sharp realistic outputs, but at the cost of instability, mode collapse, and delicate optimization.
4. Introduce adversarial training as a major new paradigm in generative modeling.

Notes

• GANs were foundational because they changed how generative modeling could be framed, not just because they produced good images.
• A large fraction of later GAN work is about stabilizing the training game.