In the past, one way AI researchers have tried to overcome this problem is to use a method called data reinforcement. Using the image algorithm again as an example, they try to overcome that problem by creating “distorted” copies of what is available in cases where there are not many materials to work with. Distortion, in this case, means cropping an image, twisting or rotating it. The idea here is that the network does not see the same exact picture twice.
The problem with that approach is that it leads to a situation where GAN learns to emulate those distortions, instead of creating new ones. Nvidia’s new Adaptive Discriminator augmentation (ADA) approach still uses data reinforcement, but is compatible. Instead of distorting the images throughout the entire training process, it is selective and fit, thereby making the GAN more fit.
The potential outcome of the Nvidia approach is more meaningful than you might think. Training AI to write something new Text based adventure game The algorithm is simple because there are so many materials to work with. This is not true for many other things that researchers turn to GANs for help. For example, it is very rare to train an algorithm to detect a rare neurological brain disorder. However, a GAN trained with Nvidia’s ADA policy can overcome that problem. As an added bonus, doctors and researchers will be able to share their results more easily because they work from the base of images created by AI, not patients in the real world. Nvidia will share more information about its new ADA policy in the near future Meet the Neurips, Which starts on December 6th.