Generative Adversarial Networks
An architecture where two neural networks contest with each other to generate entirely new, realistic data.
In Plain Language
Two programs play a game. One is a forger trying to paint fake pictures. The other is a detective trying to spot which ones are fake. Every round, both get a little better at their job. Keep going long enough and the forger's pictures become so good that the detective cannot tell — and neither can you.
A GAN trains two networks against each other. The generator turns random noise into an image; the discriminator looks at images and judges real or fake. Each one's failure is the other's training signal — the discriminator improves by catching fakes, and the generator improves by fooling it. Neither needs a human to define what "realistic" means, which is the clever part: the standard is set by an opponent that keeps raising it. The catch is that this balance is delicate, and training can collapse — most notoriously into producing the same convincing image over and over. That instability is a large part of why diffusion models later displaced GANs.
Why It Was Important
Ian Goodfellow introduced GANs, pitting a 'Generator' (creating fake images) against a 'Discriminator' (trying to spot fakes). This adversarial training loop produced stunningly realistic synthetic images (like human faces that don't exist), becoming the absolute bedrock of generative media until Diffusion models emerged.
Who Invented It
Ian Goodfellow et al.
Pioneering deep learning researcher who conceptualized GANs in a Montreal bar.
Applications
- Deepfakes
- Synthetic Data Generation
- Art Generation
- Style Transfer
Key Papers
- Generative Adversarial Networks
Ian J. Goodfellow et al. · NeurIPS 2014
Videos
Generative Adversarial Networks (GANs) - Computerphile
Computerphile
What are GANs (Generative Adversarial Networks)?
IBM Technology