Improved Techniques for Training GANs

The paper presents a variety of new architectural features and training procedures for the generative adversarial networks (GANs) framework, focusing on two applications: semi-supervised learning and the generation of images that humans find visually realistic. Unlike most work on generative models, the primary goal is neither training a model that assigns high likelihood to test data nor requiring the model to learn without any labels.

Using the new techniques, the authors achieve state-of-the-art semi-supervised classification results on MNIST, CIFAR-10, and SVHN. Image quality is confirmed by a visual Turing test: the model generates MNIST samples that humans cannot distinguish from real data, CIFAR-10 samples yield a 21.3% human error rate, and ImageNet samples at unprecedented resolution show the model learns recognizable features of ImageNet classes, demonstrating gains on both the semi-supervised and generative fronts.