Generative adversarial networks (GANs) are a class of networks that were introduced by Ian Goodfellow in 2014. In GANs, two neural networks play off against one another as adversaries in an actor-critic model, where one is the creator and the other is the scrutinizer. The creator, referred to as the generator network, tries to create samples that will fool the scrutinizer, the discriminator network. These two increasingly play off against one another, with the generator network creating increasingly believable samples and the discriminator network getting increasingly good at spotting the samples. In summary:
- The generator tries to maximize the probability of the discriminator passing its outputs as real, not generated
- The discriminator guides the generator to create ever more realistic samples
All in all, this process is represented as follows...
