The original GAN (Goodfellow, 2014) (https://arxiv.org/abs/1406.2661) is a generative model, where a neural-network is trained to generate realistic images from random noisy input data. GANs generate predicted data by exploiting a competition between two neural networks, a generator (G) and a discriminator (D), where both networks are engaged in prediction tasks. G generates “fake” images from the input data, and D compares the predicted data (output from G) to the real data with results fed back to G. The cyclical loop between G and D is repeated several times to minimize the difference between predicted and ground truth data sets and improve the performance of G, i.e., D is used to improve the performance of G.
The code discussed here is Semi-supervised learning with Generative Adversarial Networks (https://arxiv.org/abs/1606.01583), utilizes a GAN architecture for multi-label classification.