15 amazing things that have learned to make the neural network

Neural network - an artificial intelligence capable of self-learning. In one form of such a program existedNeurocomputing Technology: Theory and Practice back in the eighties, but particularly the rapid development of this sector has received approximately 2015 m. The possibility of neural networks have begun to actively explore the leading universities such as MIT and Oxford, as well as large corporations, such as Google.

Now this technology is available to anyone who wishes. And mankind has come up with dozens of the most crazy and strange uses for such programs. Here are a few of them.

1. Coming up with the face of non-existent people

The people you see in the picture above, look realistic, but they do not exist. Their images createdProgressive growing of GANs for improved
quality, stability and variation ANN NVIDIA. Program to train on real photos of celebrities, and as a result she has learned to generate accurate images of faces. You can check yourself, how well she obtained.

2. Read aloud

Technologies for speech synthesis with the help of neural networks a lot. Their program is, for example, for this Google and "Yandex». This created thereby obtaining a smooth and realistic, and in applications of this method are numerous: from the audio application for visually impaired to create audiobooks inexpensively.

3. drive cars

Many companies see unmanned vehicles transport future. Their development in this area is at Audi, Uber, Google, Tesla, «Yandex» and many other corporations. Virtually none of these technologies is not without neural networks. They help cars to determine where on the road markings, signs, other vehicles and pedestrians, and make decisions based on these data.

4. Restore color to photos and videos

Researchers from Tokyo's Waseda University have developedLet there be color! a program that makes the black-and-white photographs and video colored. The neural network has learned to determine the images in common motifs (usually sky blue, the trees - green and so on) and paint objects in the appropriate colors.

5. Throughout see dog muzzle

One of the first neural network technologies that have become available to the general audience, was InceptionismInceptionism by Google in 2015. It processes images, adding them silhouettes of dog muzzles, pagodas and arches. Net users have begun to pass through the program my pictures, famous paintings, videos, and movies - and it turned out unusually scary.

6. compose music

The neural network can upload any type of digital information, including music. Some researchers trained their programs on the melodies of famous composers. Meaningful works in computers is not yet obtained, but the styles of the musicians they copy well.

7. Forcing politicians to say anything

One of the most frightening of applications of neural networks - a synthesis of the video, in particular, with public figures. For example, researchers from the University of Washington have developedSynthesizing Obama: Learning Lip Sync from Audio a program that generates a movement of the lips of Barack Obama on the basis of record and inserts it in the video. It turns out very reliably.

8. walk

Google subsidiary company called DeepMind conducted an experiment. Three different virtual figures - humanoid stick with two legs and a ball with four legs - had to learn to walk. They had no information on how this is done - only task is to get from one point to another and sensors that help determine its position in space. After hundreds of hours of practice, all three figures learned to walk, run, jump and walk on uneven surfaces.

9. control robots

based on neural network technology widely used in robotics. For example, a robot created by Disney Research Institute is able to move forward with one, two or three feet. A robot deliveryman company Starship Technologies - to navigate through the streets avoiding obstacles and pedestrians.

10. Detect fraud and corruption

One of the main functions of neural networks - recognition of images, including the correlations between events. This is very useful in the financial sector: it is possible to predict the illegal activity before it happens. For example, in Spain, scientists have createdPredicting Public Corruption with Neural Networks: An Analysis of Spanish Provinces a program that helps to detect acts of corruption in the country's provinces. And some banks are developingCiti Ventures Deploys Machine Learning And Artificial Intelligence With People and use systems which detect credit card fraud.

11. Translate the text in the image in real time

the transfer function of the text appeared in the «Google translator» in real time ago, but few people know that it usesHow Google Translate squeezes deep learning onto a phone Neural networks. With their help, the program recognizes the letters and other characters in the images, even if they are washed, rotated around its axis, stylized or distorted. Then, the application adds them to words and sentences, translates and projects the picture. And all this for a fraction of a second.

12. Wrap the art style from one image to another

In 2016, several companies presented image processing technology in various artistic styles. app appeared like Prisma, DeepArt and Ostagram. Prisma allows you to select from several hundred harvested filters and Ostagram and DeepArt - to upload a picture or photo that will serve as a source of style.

13. Convert rough sketches in realistic paintings

NVIDIA Corporation in early 2019 showedStroke of Genius: GauGAN Turns Doodles into Stunning, Photorealistic Landscapes a program that turns pictures of a few simple figures in detalizovannye beautiful picture. The user makes a couple of strokes, and the neural network is created from this image that distance does not distinguish from the real fabric of a landscape painter. Sea, rocks, city, forest, cloud - at the picture, you can add dozens of different objects. Even the neural network determines where you need a shadow or reflection.

14. Read lips

Google and scientists at Oxford University have createdLipNet LipNet technology that uses a neural network to read lips. And it makes it much more precise person. On average, people with hearing impairments to lip-read with an accuracy of 52%, and LipNet - with an accuracy of 88%.

15. write texts

People are taught to the neural network and work with text. programs writtenDeep-speare: A Joint Neural Model of Poetic Language, Meter and Rhyme poems, stories, feykovye texts to "Wikipedia", the script for the series (for example, for friends").

And in 2016 came the world's first short film Sunspring, who wrote the screenplay for the artificial intelligence. Movie absolutely pointless: work until computers are given with difficulty. But who knows, maybe in a few years the profession of writer is reduced to editing the works created by the machine.