What can neural networks do? We tell how AI paints pictures, writes theses, passes exams and reads lips

 

While we are watching the conflict between neural networks and digital artists develop, this technology is also making great strides in other industries: it searches for answers to any question, analyzes credit history, writes witty comments on Twitter, and helps in diagnosing diseases. We have collected the most interesting and important achievements of neural networks in this article.

Let's close the topic of "picture" neurons right away - we released a large material about their capabilities, where we talked about how they work, what they are capable of, and how to configure and use them. We have also prepared various guides on the same topic.

In addition to creating pictures, neural networks are able to cut out the background, adjust the image to different styles, and generate hundreds of non-existent people per second - news about this flashed all past year.

The success of this technology in other areas - science, business, and medicine - is much more important, but the effects are not so obvious. Although the same ChatGPT made even Google worry.

Almighty ChatGPT

ChatGPT is a chatbot based on the GPT-3.5 language model. According to the developers at OpenAI, the model itself can be used to solve "any task in English." And they never exaggerate.

ChatGPT is able to write poems, and stories, have a meaningful dialogue, answer questions better than Google, argue with arguments, and even write code.

Schoolchildren and students have long learned to write off tests, but checking essays and detailed questions, until recently, could still tell the teacher about the level of knowledge of the student. With the advent of ChatGPT, you can no longer be sure of this.

So, a teacher from a Russian lyceum asked a neural network to write a final essay for the Unified State Examination and let several colleagues evaluate the result. If you do not take into account a couple of repetitions and negligence, such an essay would have received a “pass”, it met all the required evaluation criteria: the required volume, the presence of arguments, conclusions, and so on.

Another enthusiast was "driving" a neural network in an introductory microbiology course, and the program did a better job than the average student. Moreover, the neural network understood comparisons, was able to identify a common feature in several concepts, and solve a microbiology problem, justifying each step. According to the author of the material, he would give neural networks 95 points out of 100, which is much more than the average microbiology student receives.

ChatGPT does a great job with large amounts of data. This was taken advantage of by a judge from Colombia, who was faced with a very specific case - it was necessary to find out whether the health insurance of a child with autism covered all the costs of his treatment. The servant of Themis entered all the initial data into a chatbot, which analyzed the data from the Internet and issued a verdict in favor of the child.

Of course, he didn't directly ask what decision needs to be made. The judge entered specific queries, and the chatbot searched for information about judicial precedents. But even the decision-making process has accelerated many times over.

Recently, one of the students of a Moscow university wrote a diploma using ChatGPT in 23 hours, which he then successfully defended. First, he asked to form a diploma plan on a given topic, then - to write down each of the points. Of course, I had to edit the style of the text and edit the translation, but the originality of the work was 83%. And, perhaps, if this case had not received publicity in social networks, no one would have noticed the “artificiality” of the work.

In order to combat this, OpenAI, the company that owns the chatbot, has even released the AI ​​Text Classifier tool, which is quite good at identifying texts written by a neural network. Now the tool is under development, and the creators promise to improve the algorithms for more accurate recognition.

But if the text is rewritten "on paper", certain difficulties may arise with recognition. But this process has also been automated. Blogger Tomary lul on his YouTube channel showed a bunch of ChatGPT and a 3D printer. Instead of a print head on a 3D printer, a regular pen is installed, and the text is generated by a chatbot, and then immediately transferred to the 3D printer software.

It seems that teachers of the future will have to find new criteria for assessing students' knowledge, because neither tests, nor essays, nor detailed answers will no longer guarantee that this is personal knowledge, and not rewriting "from the Internet."

Of course, the fact that the neural network was trained in English imposes some limitations in Russian realities, but even so, the model is capable of a lot.

search engine of the new era

ChatGPT is able to discuss and read works with the user, answer philosophical questions, give clear answers, summarize information (and pour water if necessary), understand strict limits and conditions, rhyme, and search for information better than Google.

Unlike a search engine, it does not provide links to thousands of different pages, but one very specific answer. Moreover, it is formulated in such a way that it can simply be copied into a notebook and submitted for verification.

Therefore, soon after the advent of ChatGPT, other companies that own search engines either began work on introducing a chatbot into their services or announced their own analogues of the technology.

Microsoft just recently showed an updated search engine Bing NEW. An AI model was connected to it, and now the user, when requested, receives two results at once. On the left are “classic” links to pages, and on the right are bot tips.

The new search engine is able to answer subject-specific queries such as “Will an IKEA shelf with article X fit in a minivan”, as well as redirect the user directly to the chatbot interface to refine the query or email the result to friends.

The presentation showed an example with a sofa

The bot is also capable of retelling PDFs, comparing information from multiple sources, displaying results in tables, rewriting code snippets, or posting notes on LinkedIn. The company is trying to introduce this technology into its ecosystem, thereby creating a real virtual assistant.

Information about the Google system is much less. It's called Apprentice Bard. An important feature of technology is the relevance of information. The search bot has an up-to-date database, while ChatGPT is trained on a 2021 dataset. He knows nothing about current events.

The search results for a Google product are similar to those of Bing NEW. All the same two versions - the generated answer and the classic list of links. Unlike the OpenAI product, the system tries to generate responses from multiple sources rather than "copying" the response from one. There are no other details about the technical aspects and capabilities of Apprentice Bard.

Yes, and enthusiasts did not stand aside - for browsers based on Chrome and Mozilla, they created an extension that embeds ChatGPT into search engines such as Google, Yandex, and DuckDuckGo. The mechanism of operation is the same - on the left are requests, on the right is the "response" of the neural network in the user's native language. The capabilities of the chatbot, including its ability to write code, have been fully preserved.

Code that writes itself

Zero-code technology is a promising direction in the development of neural networks, which will allow you to create applications without the need to write code. And these applications will not only run but also work properly.

The ChatGPT mentioned above is already capable of analyzing lines of code, explaining the principles of algorithms, writing programs on demand, and even pointing out errors. The StackOverflow service, where programmers could ask each other questions, had to impose a limit on answers using this program.

The administration took this step due to the fact that the bot could give out reliable at first glance, but erroneous answers and judgments

So, one Reddit enthusiast wrote "neurodebugger". It is able to analyze code written in 22 programming languages ​​and explain where the error is and how to fix it.

The program is based on the OpenAI Codex neural network, which is also part of ChatGPT. With the help of the latter, the user habr wrote a simple application on Android. Well, as I wrote: the enthusiast decided to set up an experiment - he personally will not write a single line of code, but will only copy the answers of the chatbot. At the same time, he carefully followed the advice of the neural network and, when errors and bugs appeared, asked ChatGPT how to solve them.

At some point, the author began to “throw” errors into the chat bot, and he instantly answered with the corrected code

As a result, after a dozen or two requests, he managed to create an application that runs and regularly shows the current stock price of 50 of the most famous companies.

Ultra-precise surgery and instant analysis

We recently wrote in the news that doctors performed the world's first brain surgery using AI. Using Human Connectome technology, doctors scanned the patient's brain, analyzed it, and planned the operation to perform surgery through a hole in the skull less than a centimeter in size. Previously, it was necessary to make incisions of 2.5-3 cm.

One of the models created by Human Connectome

But AI in medicine has been used for a long time. Neural networks, based on an array of patient data, calculate the optimal dosages of drugs, help search for effective treatments, and diagnose various types of cancer using full-format images that can weigh up to 1 GB. The analysis of such images by doctors without AI takes ten times more time.

This is how the data array and various calculations look in the neural network view

To summarize, such technologies are gradually becoming faithful assistants to doctors of all specialties. Even now, AI can make a diagnosis more accurately than most known methods, calculate a treatment program, consult a patient, follow the course of the disease, and answer all the questions that have arisen almost directly.

Of course, legally all this remains only in the framework of laboratory research. The technology is still too young for widespread use. Who will be responsible for the AI ​​error? Is 95% accuracy enough in diagnosis? Will it do things in a bad way, as it happens with autonomous driving technologies?

Scientific discoveries requiring processing of a large amount of data

But, of course, neural networks are fully revealed when it comes to “big data” - areas of science and technology where you need to operate with a huge amount of information, extracting the necessary information according to certain criteria or generating and immediately testing new concepts.

Most drugs are based on an active substance with a certain composition. One of the most promising areas is the creation of new drugs. To do this, the neural network is first asked to "generate" a new molecule, an active substance, and then to model its parameters - to predict how it will behave in a given situation and what properties it will have.

So, option after option, the computer invents new options. Of course, they have yet to be recreated and tested, but the machine does it hundreds and thousands of times faster than a human would.

Interestingly (and scary), not so long ago, enthusiasts did the same thing “in reverse” - they asked specially trained neural networks to model the most toxic substances possible. The neural network produced more than forty thousand options, many of which scientists identified, and some turned out to be much more toxic.

Thanks to AI, discoveries have already been made in the field of chemistry, astronomy, mathematics, physics, as well as in many other areas that require thoughtful and painstaking analysis.

The fact is that humanity has come close to technologies that generate too much data. For example, they will soon launch the Square Kilometer Array, a space exploration device that will generate as much traffic as the entire Internet. Without “virtual assistants”, there is no way to process so much data, even if all the astronomers of the world work on them.

Routine affairs and business planning

Back in 2018, Sber laid off 14 thousand employees, some of which were sent for retraining to other departments. According to the bank's management, they performed routine work, which they managed to shift to the shoulders of artificial intelligence. And by 2025, it is planned to “optimize” half of the working staff in this way.

This news is another reason to think about the possibilities of neural networks. Increasingly, they are used for everyday things: AI issues fines, finds missing cars using CCTV cameras, analyzes handwritten text and document scans, calculates wear and tear on equipment, and even develops business strategies.

For example, specially equipped cars are now driving around Moscow, which automatically fix the "flaws" of the city: potholes on the roads, rickety road signs, and broken city property. In total, six such cars were able to record more than 23,000 violations in a couple of months. What if there are more cars?

Or, for example, another direction. If AI can read lips? Upload the recordings of surveillance cameras, make a database with biometric data of people - and the same “Big Brother”, which is so often drawn in old action films, is ready. No need to introduce though - Meta Research has unveiled the AV-HuBERT audiovisual latent BERT model, which, if the recording is clear enough, delivers excellent near-real-time lip reading.

Another project, Dbrain, deals with handwriting recognition and checking documents for authenticity. It compares the image format, looks for artifacts and “overwriting”, checks the informational “signature” of the photo, thereby allowing you to recognize a fake document or its photo in seconds.

Neural network-generated language

And recently, Google introduced its neural network for video editing. It is able to create video from a text description and change objects in the frame. For example, you can take a picture of a car in a parking lot, and the neural network will create a video where this car fords a river.

This technology originated from another - some "picture" neural networks have long been able to complete images, adding new details or replacing some objects with others according to a text description. One such program is Text to Image Editor. But even in the same Stable Diffusion there is a special inPaint module. Draw glasses to Van Gogh? Easily! Paint a red car green? No problem! All in all, video generation is just the next step in a technology that could change cinema and the CGI effects industry.

After such a full-length Mona Lisa no longer seems to be something special.

Although much more often we are faced with "advertising AI". Most large companies like Google, Amazon, or Alibaba use them to optimize their banner ads, better and better guessing the needs and demands of citizens. So if you uttered the word “mattress” somewhere, and after a couple of minutes you saw its advertisement, you should know that this is the work of a neural network.

About the future of neural networks and risks

Before our eyes, the world is beginning to use neural networks with excitement in all spheres of life. In science, they process petabytes of data in a matter of days; in business, they take on routine work and suggest how processes can be optimized. In medicine, they monitor the patient and develop treatment strategies no worse than professional doctors.

Perhaps very soon we will not be able to imagine a world without neural networks

But it is worth remembering that neural networks are just a tool. And this tool uses the knowledge accumulated by the Internet. And here a problem arises - a conditional programmer or blogger leaves a much clearer information trail on the network than a locksmith or lumberjack. As a result, the latter are valued much lower for neural networks, because there are fewer references to them.

Neronets often form one image and try to stick to it: if it’s a lumberjack, then it’s definitely a brutal man with a beard, and plumbers, for example, are completely similar to Mario

Due to such incidents, for example, in the United States, the project for the automatic issuance of loans was curtailed. The created neural network rejected 80% of applications from black clients or people with certain surnames or professions because in the data that was “fed” to it, the status of people with similar “tags” was very low.

Yes, and the Internet knows many cases when neural networks or chatbots quickly became villains who poured out racist and anti-social statements. Recently, Neuro-Sama, a vituber neural network, was banned from Twitch for this reason. Then, however, they unbanned.

The streamer went to watch Nothing, Forever, an endless cartoon sitcom in which the script, picture, and voice acting are created by a “collective” of neural networks.

But Neuro-Sama did not last long - soon she again began to think about unacceptable topics and her channel was blocked. However, the Twitch channel of the sitcom went into the ban, the main character of which also began to think about the “diseases” of transgender people.

Of course, all these scandals and high-profile cases will not stop progress. But we should remember that neural networks are a tool of limitless power that is directed by people. The possibilities of technology are difficult to grasp and, quite possibly, in the next decade it will change the way people think about both science and technology in general.