How to generate very large images with GANs (Ep. 85)
It seems that generating images with GAN is a relatively straightforward task.
However, generating large images with GANs, is a completely different story.
In this episode I explain how a research group from the University of Lubeck dominated the curse of dimensionality with imaging and Generative Adversarial Networks. Because they can generate relatively large medical images with GANs. The problem is not as trivial as it seems.
The challenge of generating large images with GANs
It turns out, many researchers have failed in generating large images with GANs before. In fact, one of the main challenges comes from the fact that, in the case of large images, one needs to implement neural networks with an excessive number of parameters. This in turn, makes the entire training procedure slower. In many cases the training procedure becomes prohibitive. Even the best available GPUs at time of writing have a hard time dealing with such excessive memory requirements.
The solution: divide-and-conquer
However, researchers have found that breaking down the problem into smaller components is beneficial. Such an approach is not new. It is usually referred to as divide-and-conquer and it is pretty common in machine learning. However, this not only seems to solve the issue of high memory requirements. But the entire method also performs with very high accuracy and sample diversity. This in turn allows one to generate images that resemble the original ones very well.
More specifically, in order to generate very large images with GANs, researchers focused on one particular application. They have been trying this approach in medical imaging.
In fact, researchers used the method hereby described in the medical imaging field in order to generate CT and X-ray images. Needless to say, such a procedure can be applied to other types of images.
Listen to the full episode. Learn how to generate large images with GANs!
Don’t forget to join the discussion on our Discord server
Multi-scale GANs for Memory-efficient Generation of High Resolution Medical Images https://arxiv.org/abs/1907.01376