DCGAN refers to a model described by Radford, Metz, and Chintala that uses deep convolutional neural networks in a generative adversarial network model.
Generative adversarial networks (GANs) are structured as a competition between two models:
- a generative model that tries to create fake examples of training data interspersed with real training data.
- a discriminative model that tries to classify real examples from fake ones.
DCGAN uses deep convolutional neural networks for both models. Convolutional neural networks (CNNs) are well-known for their performance on image data. DCGAN uses the strong performance of (CNNs) to learn unsupervised representations of the input data.