Take a look at the “celebrities” in the photos below and try to recognize them. You should probably notice some similarities with Oscar winners and other famous names who have stepped on the Hollywood red carpet. Only it will not identify anyone with absolute certainty: because these people are not real, they were created by a computer.
Every great company has been carrying out his experiments with digital neural network to train a learning machine of artificial intelligence (or artificial intelligence – AI) in support of its activities. In the case of the Nvidia graphics card manufacturer, the team has been testing a system capable of generating content very similar to the actual one, from a series of examples, while the results are evaluated by the machine itself, which needs to be defined if they are natural or artificial.
The project is called the Generative Adversarial Network (GAN) and the purpose of the studies is to make the AI less and less need humans to produce schemas, environments and virtual structures. “The main idea is to evolve both the generator and the discriminator in a progressive way: we start with low-resolution models and add layers that enhance detail as the training progresses. This streamlines and stabilizes the experiment, enabling us to produce unprecedented quality images, ” the researchers said .
How the experiment works
Nvidia took advantage of the CelebFaces Attributes library – which contains no less than 200,000 images with 40 different attribute classes – and used the combined power of CUDA graphics modeling software, with the powerful artificial intelligence Tesla P100 and the cuDNN learning machine for process 30,000 images in 20 days. Inside the Subject: See with Hekima the Evolution of Artificial Intelligence from the 14th Century to the Big Data Era Sponsored
First, they started with JPEG photos and optimized the quality of each sample. Then they scanned and filtered the models and then dense scans on each element of the celebrity faces. Afterwards, they cut out unnecessary components and set out for the “new drawing” with a resolution of 1024 x 1024 pixels.
It is certain that many of the results and some of the figures under construction could well be part of any good citizen’s nightmare, but the promising conclusion may lead to more ambitious projects in a future in which AI is already expected to generate digital worlds complete without the help of humans.
Check out the video below:
See More At TechLinkn.