15 amazing things neural networks have learned to do

2024 Author: Malcolm Clapton | [email protected]. Last modified: 2023-12-17 03:44

From driving a car to creating masterpieces.

A neural network is an artificial intelligence capable of self-learning. In some form, similar programs existed. Neurocomputer technology: theory and practice back in the eighties, but this area received especially rapid development around 2015. Leading universities such as Massachusetts and Oxford, as well as large corporations, such as Google, began to actively explore the possibilities of neural networks.

Now these technologies are available to anyone. And humanity has already come up with dozens of the most insane and strange applications for such programs. Here are a few of them.

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

The people you see in the picture above look realistic, but they don't exist. Their images created Progressive growing of GANs for improved

quality, stability and variation neural network from NVIDIA. The program was trained on real photographs of celebrities, and as a result, it learned how to generate reliable images of faces. You can check for yourself how well she does it.

2. Read aloud

There are many technologies for synthesizing speech using neural networks. For this, there are programs for this, for example, and "". Speech created in this way is fluid and realistic, and there are many uses for this method, from dubbing applications for the visually impaired to creating audiobooks at low cost.

3. Drive cars

Many companies see self-driving cars as the future of transportation. Audi, Uber, Google, Tesla, Yandex and many other corporations have their own developments in this area. Almost none of these technologies is complete without neural networks. They help vehicles determine where markings, signs, other vehicles and pedestrians are on the road and make decisions based on this data.

4. Recover color of photos and videos

Scientists from Waseda University in Tokyo have developed Let there be color! a program that makes black and white photographs and videos in color. The neural network has learned to identify common motives in images (the sky is usually blue, trees are green, and so on) and to paint objects in the appropriate colors.

5. See dog faces everywhere

One of the first neural network technologies to become available to a wide audience was Google's Inceptionism Inceptionism in 2015. She processed the images, adding silhouettes of dog faces, pagodas and arches to them. Netizens began to pass their photos, famous paintings, videos and films through the program - it turned out unusual and creepy.

6. Write music

Any kind of digital information can be loaded into neural networks, including music. Some researchers teach their programs on the tunes of famous composers. Computers have not yet produced meaningful compositions, but they copy the styles of musicians quite well.

7. Make politicians say anything

One of the most frightening uses of neural networks is video synthesis, in particular with public figures. For example, scientists at the University of Washington have developed Synthesizing Obama: Learning Lip Sync from Audio, a program that generates Barack Obama's lip movements based on audio recordings and substitutes them in video. It turns out very reliably.

8. Walk

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

9. Control robots

Technologies based on neural networks are widely used in robotics. For example, a robot created by the Disney Research Institute can move forward with one, two, and three legs. And the delivery robot from Starship Technologies is to navigate the streets, avoiding obstacles and pedestrians.

10. Recognize fraud and corruption

One of the main functions of neural networks is pattern recognition, including correlations between events. This is very useful in the financial arena: you can predict illegal activity before it happens. For example, in Spain, scientists have created Predicting Public Corruption with Neural Networks: An Analysis of Spanish Provinces, a program that helps detect corruption in the provinces of the country. And some banks are developing Citi Ventures Deploys Machine Learning And Artificial Intelligence With People and are using systems that recognize credit card fraud.

11. Translate text on an image in real time

The real-time text translation feature appeared in Google Translate for a long time, but few people know that it uses How Google Translate squeezes deep learning onto a phone neural networks. With their help, the program recognizes letters and other symbols in images, even if they are blurry, rotated around their axis, stylized or distorted. Then the application puts them into words and sentences, translates and projects them onto the picture. And all this in a split second.

12. Transfer art style from one image to another

In 2016, several companies presented technologies for image processing in different artistic styles. Apps like Prisma, DeepArt and Ostagram have appeared. Prisma allows you to choose from several hundred pre-made filters, and Ostagram and DeepArt - you can upload a picture or photo yourself, which will serve as a source of style.

13. Turn rough sketches into realistic paintings

In early 2019, NVIDIA showed Stroke of Genius: GauGAN Turns Doodles into Stunning, a Photorealistic Landscapes program that turns pictures from a few simple shapes into beautiful detailed pictures. The user makes a couple of strokes, and the neural network creates an image from this, which from afar cannot be distinguished from a real canvas of some landscape painter. Sea, rocks, city, forest, clouds - dozens of different objects can be added to the picture. The neural network even itself determines where shadows or reflections are needed.

14. Read lips

Scientists at Google and Oxford University have created LipNet technology LipNet, which uses neural networks to read lips. And she does it much more accurately than a person. On average, hearing impaired people read lips with 52% accuracy, and LipNet with 88% accuracy.

15. Write texts

People taught neural networks and how to work with text. Programs are written by Deep-speare: A Joint Neural Model of Poetic Language, Meter and Rhyme poems, short stories, fake texts for Wikipedia, scripts for serials (for example, for Friends).

And in 2016, the world's first short film, Sunspring, was released, the script for which was written by artificial intelligence. Cinema is absolutely meaningless: computers are still struggling to create. But who knows, maybe after a few years the profession of a screenwriter will be reduced to editing works created by a machine.