Hands-On Generative Adversarial Networks with Keras
上QQ阅读APP看书,第一时间看更新

Real and fake data

Real and fake are the terms used to refer to the data used in GANs. Real refers to the data that comes from the distribution we want to learn, and fake refers to the data produced by the generator. In most GAN use cases, the training procedure builds a decision boundary between all real data and all fake data.

In cases where there is not enough data, triplet loss learning can be used so that the discriminator also learns to differentiate between the target real data and non-target real data. This approach has been used in the paper Attacking Speaker Recognition with Deep Generative Models.