Generative Adversarial Networks
GANs may be seen as providing mechanisms for computational sampling from very high-dimensional distributions. A well-trained GAN is able to draw high-dimensional samples from a distribution that is otherwise only captured by real-world data samples. A GAN is therefore theoretically able to be used in data augmentation. We have achieved some quite spectacular data augmentation using capsule GANs for microscopy image data.
But a more likely end-use for GANs is in creating open datasets of synthetic data that do not reflect anatomy or physiology of any one individual, but represent plausible samples along the high-dimensional manifold corresponding to real patient data.
The adversarial approach can also be applied in the latent space, and we have also developed some techniques for adversarial training of autoencoder networks for representation learning.
Hybrid Generative Models
Modern generative models are commonly assumed to be driven entirely by deep neural networks. But this ignores established fields of data simulation (including CGI) where the distribution is very high dimensional, but constrained by strong physical laws that we understand quite well. This includes imaging systems (such as MRI and ultrasound).