NVIDIA Says that They have Found New Way to Train AI

NVIDIA has developed a new approach for training generative adversarial networks (GAN). This new approach has the capacity to make AI suitable for a greater variety of tasks.

Before understanding NVIDA’s work let’s understand a little about GAN. It consists of two competing neural networks i.e., a generator and a discriminator. The goal of the first algorithm is to create new images and the second is to examine thousands of sample images.  

Then it uses data to coach its counterpart and create believable results consistently. Traditional GANs tend to overfit and in such instances, the discriminator doesn’t have enough of a base to effectively coach the generator.      

In the past, one way AI researchers have tried to get around this problem is by data augmentation or using an image algorithm. In such a case, distorting is a simple reason which means cropping an image, rotating it, or flipping it. Here, the network users see the exact same image twice.  

The problem with this is, instead of creating something new, it would lead to a situation where the GAN would mimic those distortions.  

NVIDIA’s new adaptive discriminator augmentation (ADA) approach uses data augmentation but also uses it adaptively. For instance, instead of distorting an image throughout the entire training process selectively. It will be just enough so that the GAN avoids overfitting.    

NVIDIA’s approach is very meaningful than one might think as it can train an AI to write a new text-based adventure game. This is possible as there is very little material for the algorithm to work with.  

However, we can’t say the same for a lot of other tasks where researchers could not turn to GANs for help. For example, training an algorithm to spot a rare neurological brain disorder precisely because of its rarity.  

However, a GAN that has been trained with NVIDIA’s ADA approach can easily get around that problem. Researchers will be able to easily share their findings as an added bonus since they’re working from a base of images which is created by an AI.  

Starts on December 6th, the company will share more information about ADA at the upcoming NeurIPS conference.