The encyclopedia · R&D & Science · Technical decision · 2014–2018
Generative adversarial networks made two models train against each other.
Goodfellow's GANs pit a generator against a discriminator, so each improves by trying to beat the other.
OpenAI · Google Brain
the move
Generative models, which create new data from a distribution, were often hard to train or produced blurry output because they required complicated inference.
Goodfellow and colleagues proposed pairing a generator with a discriminator in a minimax game; the generator learns by trying to fool the discriminator, and the discriminator learns by resisting.
The framework required no Markov chains or approximate inference and could be trained with standard backpropagation, and it produced convincing samples.
why it works
- The discriminator provides a learning signal without a labelled target
- The adversarial objective is a scoreboard that improves automatically
- Backpropagation-only training makes the approach practical
- Competition between the two networks drives both toward better performance
what transfers
You don't always need an objective truth to train against: a well-designed adversary can be the scoreboard, so two imperfect systems push each other toward competence.
what came after
GANs became one of the most influential ideas in machine learning, powering image generation, style transfer and synthetic-media research. They also raised concerns about deepfakes and misuse, and a large body of follow-up work stabilised and extended the original training objective.
references
spotted an error? The archive wants to know.