To create something new, imagination is needed. If you are human, of course. If you are a machine, you need some algorithms and a lot of training. With this, you can reach create hyperrealistic landscapes from two poorly made lines in Paint. A blue line can automatically become a river or waterfall; a brown blur, in a mountain range; and a purple spot, in a cloudy sky. These are the achievements of an artificial intelligence developed by the company Nvidia, known for the design of computer chips for the development of algorithms (and video games), which translates erratic tracings into dreamy landscapes that do not exist in the real world.
To achieve this, the key lies in developing antagonistic generative networks. Two neural networks (ie, simplified mathematical models of the brain) are taken and confronted in a digital game. Both are trained with the same data set. The generative network has the task of creating variations in the images that you have already seen. The second network, known as the discriminator, must identify if the image it is seeing belongs to the originals if it is a false image produced by its partner. Over time, the generative network is so good at producing new images that it is impossible for its discriminating partner to detect counterfeiting. When this second network is unable to differentiate them, it is assumed that humans will not be able to find the difference either.
The system proposed by Nvidia is called GauGAN, a play on words between the antagonistic generative network -GAN, in English- and the French post-impressionist painter Paul Gaugain. Your goal is to paint an image as a human being would. In addition, the multimodal: if two different users draw the same, the software will create different results in real time. Nvidia has trained him using a million Flickr images. With this wisdom, the program can synthesize hundreds of thousands of natural elements and their relationship with other objects. For example, you can deduce the reflection in a lake if you draw a tree on the shore.
Its creators say it can open a whole new world to the design of video games and even scenarios for movies. However, they highlight some aspects that have been proposed to improve before enabling it for the public. For example, they are aware that the boundaries between objects are not perfect. If you look closely, there is a thin line that separates them.
Antagonistic generative networks have proven to be extremely effective in distorting reality. It is also these networks that are used to create faces of people who do not exist. These types of algorithms are evolving very rapidly with the expansion of deepfakes. A few months ago, Buzzfeed created from scratch a video of Barack Obama talking about fake news using artificial intelligence. The objective was to demonstrate how easy it is to make the former president of the United States say what you want and alert the citizens of the amount of false images that we see at the end of the day and that we take as real. Artificial intelligence, through the use of GAN networks, is increasingly used to falsify images or videos with incredibly realistic results.