The Birth of the Matrix: artificial neural networks learned to create realistic faces of people and interiors of bedrooms

In the Matrix, in its first part, not only an intelligent computer system could create a virtual world, but the computers of Nebuchadnezzar, a rebel ship, could also generate a miniature virtual world. Many remember the girl in red, who appeared in one of the simulations - she is shown in the picture in the announcement. We (and our computers) are still very far away from creating such a virtual model of the real world. But something is already there now.

A team of researchers from Indico and Facebook has created a neural network that can "come up" with realistic images. By the way, the evolution of neural networks can be found here . Artificial Neural Network ( ANN) - a mathematical model, as well as its software or hardware implementation, built on the principle of organization and functioning of biological neural networks - networks of nerve cells of a living organism. Now ANNs are used in many areas, including forecasting, pattern recognition, working with Big Data. Alec Radford of Indico and his colleagues decided to pay attention to such a kind of NIS as the generative adversarial network.



In such a network, one part of the system tries to create an array of false data in order to “trick” the second part. The idea was that if you constantly repeat this process, the system will learn how to create better images.

The authors of the study trained a system based on an array of bedroom images. Then the system was asked to create their own images. To make sure that the system really creates the original images, and does not copy them from the database, experts gave the command to generate a series of modifications of the same bedroom image - for example, without a window, with a table or TV. This technique made it possible to verify that the system learned to generate interiors on its own.

In addition, experts trained the system to create images, where one important part (for example, a window) is replaced by something. For example, a system replaces a window with a television or fireplace. This indicates that the system has learned to “understand” what constitutes a certain part of the interior.



A similar principle can be used for other shots, such as portrait photographs. In another experiment, the researchers asked the system to highlight a photo of a smiling woman, then add a neutral expression and “add” a man with a neutral expression. The goal is to single out the concept of “smile” and combine it with the concept of “man”.

As a result, the researchers received a set of images of smiling men.

Now, unfortunately, the image size is limited to 32 * 32 pixels, which makes the system faster and more efficient. The next stage of work is the transfer of such a scheme to video and audio.

And after the 3D effect is added, plus dynamics, plus additional environmental details, we get a full-fledged “Matrix”. Or not?