Generative Adversarial Networks (GANs)

Deep learning has launched a profound reformation and has even been applied to many real-world tasks such as image classification (He et al. 2016), object detection (Ren et al. 2015), and image segmentation (Long et al. 2015). These tasks all fall into the scope of supervised learning, which means that large amounts of labeled data are provided for the learning processes. Compared with supervised learning, however, unsupervised learning shows little effect from deep learning. Generative modeling is a typical problem in unsupervised learning, the goal of which is to learn the distribution over training data and then to generate new data by sampling from the learned distribution. Generative modeling is usually more difficult than supervised learning tasks because the learning criteria of generative modeling are intractable (Goodfellow et al. 2016). For supervised learning tasks, the corresponding mapping information between the inputs and the outputs is given, and the supervised learning models need only learn how to encode the mapping information into the neural networks. In contrast, for generative modeling, the correspondence between the inputs (usually a noise vector) and the outputs (the training data) is unknown, and the generative models must learn how to arrange the mapping between the inputs and the outputs efficiently.