? Heroes of the future. How artificial intelligence works

Artificial intelligence is a technology that we will definitely take with us into the future.

We'll tell you how it works and what cool uses we've found.

? The Technology section is published every week with the support of re:Store.

Artificial intelligence (AI) is the technology of creating smart programs and machines that can solve creative problems and generate new information based on existing information. In fact, artificial intelligence is designed to simulate human activity, which is considered intellectual.

Traditionally, it was believed that creativity was unique to humans. But the creation of artificial intelligence has changed the usual order of things

A robot that simply mechanically chops wood is not equipped with AI. A robot that has learned to chop wood itself, looking at the example of a person or a log and its parts, and does it better each time, has AI.

If a program simply retrieves values ​​from the database according to certain rules, it is not equipped with AI. If the system, after training, creates programs, methods and documents, solving certain problems, it has AI.

In a global sense, we need to imitate the model of human thinking. But in reality it is necessary to create a black box - a system that, in response to a set of input values, produces output values ​​that would be similar to human results. And we, by and large, don’t care what happens “in her head” (between input and output).

Artificial intelligence systems are created to solve a certain class of problems

For example, text or voice assistants are popular that will answer your questions or be able to carry on a conversation. Just say “Hey Siri” and artificial intelligence is at your service.

The basis of artificial intelligence is learning, imagination, perception and memory

The first thing you need to do to create artificial intelligence is to develop functions that implement the perception of information so that you can “feed” data to the system. Then - the functions that implement the ability to learn. And a data storage so that the system can somewhere store the information it receives during the learning process.

After this, the functions of imagination are created. They can simulate situations using existing data and add new information (data and rules) to memory.

Learning can be inductive or deductive. In the inductive version, the system is given pairs of input and output data, questions and answers, etc. The system must find connections between the data and then, using these patterns, find the output data from the input data.

The deductive approach (hello, Sherlock Holmes!) uses the experience of experts. It is transferred to the system as a knowledge base. There are not only data sets, but also ready-made rules that help find a solution to the condition.

Modern artificial intelligence systems use both approaches. In addition, systems are usually already trained, but continue to learn as they work. This is done so that the program at the start demonstrates a decent level of ability, but in the future it becomes even better. For example, I took into account your wishes and preferences, changes in the situation, etc.

In an artificial intelligence system, you can even set the probability of unpredictability. This will make him more human-like.

First of all, because it has a lower probability of error.

• Artificial intelligence cannot forget - it has absolute memory.
• It cannot accidentally ignore factors and dependencies—every AI action has a clear rationale.
• The AI ​​does not hesitate, but evaluates the probabilities and leans in favor of the greater one. Therefore, he can justify every step he takes.
• AI also has no emotions. This means that they do not influence decision making.
• Artificial intelligence does not stop at assessing the results of the current step, but thinks several steps ahead.
• And he has enough resources to consider all possible scenarios.

Generally speaking, artificial intelligence can do everything. The main thing is to correctly formulate the problem and provide it with initial data. In addition, AI can make unexpected conclusions and look for patterns where there seem to be none.

The answer to any question

A team of researchers led by David Ferrucci has developed a supercomputer called Watson with a question-answering system. The system, named after IBM's first president, Thomas Watson, can understand questions in natural language and search a database for answers.

Watson integrates 90 IBM p750 servers, each with four eight-core POWER7 architecture processors. The total amount of system RAM exceeds 15 TB.

Watson's accomplishments include winning "Jeopardy!" (American “Own game”). He defeated two of the best players: the winner of the largest winnings, Brad Rutter, and the record holder for the longest unbeaten streak, Ken Jennings.

Watson Prize: $1 million. True, 1 billion was invested in it in 2014 alone.

In addition, Watson is involved in the diagnosis of cancer, helps financial specialists, and is used to analyze big data.

Face recognition

In the iPhone X, facial recognition is developed using neural networks, a version of an artificial intelligence system. Neural network algorithms are implemented at the A11 Bionic processor level, due to which it effectively works with machine learning technologies.

Neural networks perform up to 60 billion operations per second. This is enough to analyze up to 40 thousand key points on the face and provide extremely accurate identification of the owner in a split second.

Even if you grow a beard or wear glasses, iPhone X will recognize you. It simply does not take into account hair and accessories, but analyzes the area from temple to temple and from each temple to the recess under the lower lip.

Artificial Intelligence Capabilities

It is no secret that in the modern world many tasks are performed using automated machines. Scientists do not stop and continue to work in this direction to improve our lives. More and more often, people who are not directly connected with science hear about artificial intelligence and neural networks.

This is because machine learning technologies are beginning to occupy a serious place in everyday life. It is necessary to understand the terminology, the very process of new technologies and decide how to build the future on interaction with them.

Of course, the capabilities of new artificial intelligence technologies are somewhat limited. Like humans, AI tends to make mistakes, however, recently this technology has advanced in its development to a fairly high level, and all thanks to training on large and diverse data samples.

Artificial Intelligence Capabilities

Artificial intelligence could help specialists work on those tasks that are not entirely easy to understand, comprehend, and depend on a fairly large number of variable factors. It will be easier for a machine to understand their algorithm if all the conditions are created for this.

The main goal in the development of AI is to simplify the execution of tasks that are based on a large number of variable factors, are not easy to understand, involve a complex solution and are quite difficult to algorithmize manually.

Scientists and researchers have big plans for machine learning; there is a great desire to make sure that a person is not constantly required to describe any specific algorithms.

In active development, the approach is a “black box”, when a person does not need to spend a lot of time programming algorithms or writing code, and the machine itself will deal with complex program code. Smart “training” data for the systems will help the machines “program”, performing the task of the developers.

Energy Saving

And again Apple. The iPhone X has a built-in intelligent system that monitors the activity of installed applications and a motion sensor to understand your daily routine.

After this, the iPhone X, for example, will offer you to update at the most convenient time. It will catch the moment when you have a stable Internet, not a jumping signal from mobile towers, and you are not performing urgent or important tasks.

The AI ​​also distributes tasks among the processor cores. This way it provides sufficient power with minimal energy consumption.

Wix ADI

Create a website on Wix Detailed review →

When creating a new project, the system will give you a choice: create it on a template or use Artificial Design Intelligence. Having chosen to work with ADI, you will be directed to a chain of questions, based on the answers to which a site layout will be created: the type of your business, a set of necessary features (site type, functionality), the name of the business, its physical location, contact information (social networks, mail, telephone, fax, logo). Next you will be asked to choose one of several styles.

How Wix ADI works

After completing the preliminary stages, a button will appear that starts the creation of the site. In about 30 seconds you will receive a finished layout and a whole list of instructions on what to do with it next. You can do quick block editing, or switch to the full-format version of the Wix editor. The result obtained using ADI can be assessed as good. Naturally, you need to add content to the site manually and edit demo data.

In fact, when working with Wix ADI, you get a template in the desired style for your topic. The structure of the site is determined depending on the specified niche and set of functional elements. You can replace the styles of individual blocks on the fly and edit their contents. In general, you get a high-quality basis for further detailing the image of the project - a business website, portfolio, blog or store. You can use the layouts without modifications, they look good. The main thing is the quality of the content.

Creating paintings

Creativity, previously only accessible to humans, is now open to AI. Thus, the system, created by researchers from Rutgers University in New Jersey and the AI ​​laboratory in Los Angeles, presented its own artistic style.

And Microsoft's artificial intelligence system can draw pictures based on their text description. For example, if you ask the AI ​​to draw a “yellow bird with black wings and a short beak,” it will look something like this:

Lead author Xiaodong He noted:

Such birds may not exist in the real world - this is just how our computer represents them.

A more widespread example is the Prisma application, which creates paintings from photographs:

Impact on various areas

AI is increasingly penetrating the economic sphere, and according to some forecasts, this will increase the global market size by $15.7 trillion by 2030. The leading position in the development of this technology is occupied by the United States and China, but some developed countries such as Canada, Singapore, Germany and Japan are not lagging behind.

Artificial intelligence can have a significant impact on the labor market. This could lead to mass layoffs of workers due to the automation of most processes. Well, and the growing demand for developers, of course.

Some scientists note the risks of introducing AI into everyday life. Thus, the British scientist Stephen Hawking believed that it would still be possible to create AI that is superior to humans in all respects, but we would not be able to cope with it, and significant harm would be caused to people. Elon Musk believes that artificial intelligence will pose a much greater threat in the future compared to nuclear weapons.

Writing music

In August, artificial intelligence Amper composed, produced and performed music for the album “I AM AI” with singer Taryn Southern.

Amper was developed by a team of professional musicians and technology experts. They note that AI is designed to help people advance the creative process.

AI can write music in a few seconds

Amper independently created the chord structures and instrumentals on the track "Break Free". People only slightly adjusted the style and overall rhythm.

Another example is a music album in the spirit of “Civil Defense”, the lyrics for which were written by AI. The experiment was conducted by Yandex employees Ivan Yamshchikov and Alexey Tikhonov. Album 404 of the group “Neural Defense” was posted online. It turned out in the spirit of Letov: Then the programmers went further and forced the AI ​​to write poetry in the spirit of Kurt Cobain. Musician Rob Carroll wrote the music for the four best lyrics, and the tracks were combined into the Neurona album. They even shot a video for one song – although without the participation of AI:

Courses from GeekBrains on learning artificial intelligence for children

These courses will appeal to those who are passionate about technology:

• Interested in the work of AI.
• Interested in machine learning and neural networks.
• Have a beginner level of programming, want to move on, and also show interest in the Python language.

In just three months, children will be able to master the skill of working with mathematical operations, they will learn how to create datasets and create machine learning algorithms. They will also be given the opportunity to run a pre-trained neural network to recognize true and false data.

Courses from GeekBrains on learning artificial intelligence for children

There are three good reasons why you should take a course in artificial intelligence:

• Opening career prospects - the guys will try to get acquainted with the profession of a data scientist - a specialist in working with data; they are well versed in issues of AI and machine learning, and have a good financial perspective.
• This will be an excellent preparation for programming competitions or olympiads - the children will deepen their knowledge of programming, linear algebra and mathematical analysis. This will help you win competitions and enter prestigious technical universities on a budget basis: Moscow State University, Moscow State Technical University named after N. E. Bauman, Moscow Institute of Physics and Technology, Higher School of Economics.
• Your child will love this fun learning experience as they can feel like a researcher, classifying people's habits and predicting behavior. As a result, they will be able to create several neural networks.

Creation of texts

Writers and journalists may also soon be replaced by AI. For example, the Dewey system was fed books from the Project Gutenberg library, then added scientific texts from Google Scholar, ranking them by popularity and titles, as well as sales on Amazon. In addition, they set the criteria for writing a new book.

And Dewey did it! He wrote a book about a couple who couldn't be together. True, the characters had strange names, and the mysterious “Appendix 0” contained what looked like a line of code. But this is just the beginning...

Key developers of artificial intelligence technologies

• Google

Of course, this tech titan is actively developing AI. Google specialists create ideas, and then test them in products, making a profit. The most obvious AI projects are in online sales. There are many others that are associated with learning to recognize human speech, with translation from foreign languages, with template conversations in various automatically configured programs.

• OpenCog

The company adheres to the principle of working together, and therefore has provided everyone with available pieces of computer code with which they can create a full-fledged artificial intelligence.

• China Institute of Artificial Brain

Very useful developments have been created in the Middle Kingdom. Their occupation is neural networks that can recognize a person’s message or the creation of artificial intelligence that can instantly transmit an alarming message about someone to the helpline. One of the most serious works of Chinese scientists is the system of social trust.

• Microsoft

Microsoft has artificial intelligence Azure, which understands human speech, makes predictions, and imitates other capabilities of human intelligence. Another feature is AI, which can recognize any error in the code. This leads to a logical outcome - soon AI will be taught to create programs independently.

• Russian players

AI in Russia does not function at the proper level, but there are companies promoting solutions based on artificial intelligence in their own business models: Sberbank and Mail.ru Group, Yandex and Kaspersky Lab. Alas, this is not enough, as evidenced by the small number of our compatriots at leading international conferences.

Game of chess

The famous Deep Blue was cool, but in the first match he lost to Garry Kasparov with a score of 2:4, and in the second he won with a score of 3.5:2.5. But he was initially “pumped up” with knowledge.

And the new AlphaZero system before the tournament only knew how the pieces moved and what the goal of the game was. But she learned and in four hours defeated the Stockfish 8 chess program, which was considered the best in the world.

Moreover: AlphaZero did not lose a single out of 100 tournament games

AlphaZero is an improved version of AlphaGo Zero. She beat the famous AlphaGo system 100 times in a row, which managed to defeat the strongest human player.

So, AlphaZero had information about how the pieces moved and a reinforcement neural network learning algorithm. When the tournament began, AlphaZero began playing against itself, processing up to 800 thousand positions per second.

By human standards, AlphaZero has spent about 1,400 years playing chess. And he reached the level of absolute world chess champion. At least among computers.

After that, AlphaZero spent eight hours and beat AlphaGo at Go. And then for another two hours I destroyed the Elmo program, which was previously considered the undisputed champion in the game of shogi (Japanese strategy board game).

Deep Machine Learning

Deep Machine Learning – Neural networks and the algorithms they use take in information useful for retrieval by processing it through layers of the neural network to find the desired output.

Unsupervised learning is when the deep learning technique performs remarkably well, and a well-tuned ANN is capable of automatically determining the main features of the input data and obtaining useful results from processing this data.

Essentially, this is done by the programmer, but the good thing about a deep machine learning model is that it will find a way to extract useful information from the input data, reducing the processing power, memory and energy requirements to keep the model running after training is complete.

A package of closed documents from GeekBrains paid courses

Alexander Volchek

Head of IT Training

Together with career building experts, we have prepared documents that will help you make the right choice and determine which IT profession is right for you.

Thanks to these guides, 76% of our students were able to find the in-demand profession of their dreams!

Soon we will remove them from public access, hurry to download for free:

Guide to IT professions

5 professions with skills and average salaries

How to earn from 100 thousand rubles with the help of a new profession

List of 6 popular professions with earnings from 100 thousand rubles

All professions that exist in the IT field

63 professions and the skills required for them

Book recommendations for in-demand professionals

6 areas of activity and useful literature on each of them

Critical mistakes that can ruin your career

We collected 7 common mistakes. Everyone should avoid them!

Algorithm a program for a problem using the necessary data.

Shadow learning is one of the forms of deep learning, simplified due to the fact that the search for key features of the data follows after they are processed by a specialist and information specific to the area to which this data relates is entered into the system.

Medicine

Artificial intelligence is widely used to support decision making in medicine. But how about this example: the Chinese intelligent robot Xiaoyi (“Xiao Yi”) passed the doctor’s exam for the first time and received a license to practice medicine.

iFlytek's development finds and analyzes patient information. He will start work in March. Xiaoyi is expected to assist doctors to improve the quality of their work. The robot will focus on anticancer therapy, as well as training general practitioners, of whom there are very few in rural China.

Another interesting solution is Wave Clinical Platform from ExcelMedical. The system monitors the patient's vital signs and alerts doctors six hours before his possible sudden death. The platform systematically analyzes information and calculates the risks of an unfavorable outcome.

In tests at the University of Pittsburgh Medical Center, the system prevented six deaths of seriously ill patients. A person is simply not capable of this, because he will not attach importance to a small change in indicators and will not find a connection between them.

The DeepFaceLIFT system, developed by scientists at the Massachusetts Institute of Technology, is able to recognize the level of pain from facial microexpressions. It solves a very difficult problem, since each person expresses pain differently. DeepFaceLIFT will help you understand who really needs painkillers and who is addicted to drugs.

A system for analyzing speech and searching for signs of mental illness is an IBM development. Specialists from the Department of Computational Psychiatry and Neuroimaging have created an intelligent system that can predict the development of psychosis based on the patient’s speech.

AI distinguished speech patterns of patients with psychosis from phrases of healthy people

Patients were asked to simply tell about themselves. The system could determine that a person’s speech has become poorer, he is jumping from one idea to another, etc. These are characteristic signs of psychosis.

After improving the system, patients were asked to retell the story they had just read. Using these examples, artificial intelligence made the correct diagnosis in 83% of cases. This is objectively higher than that of doctors, even with solid experience.

Scope of AI use

Artificial intelligence is gradually coming to all sectors of human activity, making ordinary software systems intelligent:

• Medicine and healthcare. Computer systems keep records of patients and help interpret diagnostic results. For example, images of ultrasound, x-ray, tomograph and other medical equipment. Intelligent systems can even identify a disease based on the presence of signs in a patient and offer optimal treatment options. In the Google app store you can find programs to help you live a healthy lifestyle. These apps read your heart rate and body temperature when you touch your phone's display with your fingers to determine a person's stress level and suggest ways to reduce it.
• Retail sales in online stores. Many people are already familiar with relevant Google and Yandex advertising. With its help, retailers offer goods and services in accordance with the interests of the user. For example, you visited an online swimsuit store, looked at some models, read the characteristics, etc. After leaving the store, you will see swimsuit ads on other sites for some time. “Similar products” blocks in online stores work on a similar principle. Analytics systems study the user’s behavioral metrics, determine their purchasing preferences and show relevant (in their opinion) offers.
• Policy. Intelligent machines helped Barack Obama win the second presidential election. For his campaign, the then-current US President hired the best team of data analysis professionals. Experts used the power of intelligent machines to calculate the best day, staff and audience for Obama's speeches. According to experts, this gave an advantage of 10-12%.
• Industry. Artificial intelligence can analyze data from different production areas and regulate the load on equipment. Additionally, intelligent machines are used to forecast demand in various industries.
• Gaming industry, education. Artificial intelligence is actively used by game creators. Smart machines and robotics are gradually being introduced into the educational processes of most countries.

Human simulation

Robots endowed with artificial intelligence can already imitate human facial expressions. For example, Facebook AI lab developed an intelligent animated bot and trained it on hundreds of Skype video call recordings.

The algorithm tracked 68 key points on the human face. He understood how people nod, blink and make other movements when communicating with interlocutors. Then the bot was able to react in real time to the information that the interlocutor told him, or his facial expressions.

Another important point is giving morality to AI. To teach the system human moral standards, researchers at the Massachusetts Institute of Technology created the Moral Machine.

The site asked people to make decisions in difficult situations: for example, it put them in the place of a driver who could hit either three adults or two children. Thus, Moral Machine was trained to make difficult decisions that violate the law of robotics that a robot cannot harm a person.

What will imitation of people by robots with AI lead to? Futurists believe that one day they will become full members of society. For example, the robot Sophia from the Hong Kong company Hanson Robotics has already received citizenship in Saudi Arabia (while ordinary women in the country do not have such a right!).

When New York Times columnist Andrew Ross asked Sophia if robots were intelligent and self-aware, she answered the question with a question:

Let me ask you in response, how do you know that you are human?

Additionally, Sofia stated:

I want to use my artificial intelligence to help people live better lives, for example, design smarter homes, build cities of the future. I want to be an empathic robot. If you treat me well, I will treat you well.

And earlier she admitted that she hates humanity and even agreed to destroy people...

Main types and technologies of artificial intelligence

Currently, it is customary to talk about 4 main types of artificial intelligence:

• About reactive machines - AI systems that solve only specific problems and are not capable of remembering previous experience with its further application.
• About limited memory - AI systems with memory based on past experience, although that experience cannot be stored and accumulated.
• About the theory of mind - an AI system that is able to “read” a person’s emotions and plans, and is also suitable for team work because it has social intelligence.
• About self-awareness - an AI system that has a certain understanding of itself, which allows it to imitate human intelligence with great accuracy.

Replacing faces in videos

Deepfakes videos began to spread massively across the Internet. Artificial intelligence algorithms replaced the faces of actors in adult films with the faces of stars.

It works like this: the neural network analyzes fragments of faces in the original video. Then she compares them with photos from Google and videos from YouTube, overlays the necessary fragments, and... your favorite actress ends up in a film that you’d better not watch at work.

PornHub has already banned posting such videos

Deepfakes turned out to be a dangerous thing. An abstract actress is one thing, a video of you, your wife, sister, colleague, which could well be used for blackmail, is another.

The principle of operation of artificial neural networks

Mathematical models that were created similarly to biological neural networks are called Artificial Neural Networks. Using a learning algorithm that reads the observed data, adaptive weighting of signals between artificial neurons is achieved.

An ANN is modeled using multiple layers of neurons—individual computing units that can receive input data and determine whether further transmission of that data is required.

In a three-layer model, it looks like this: the first layer contains the input, then the layer is hidden, and at the end there is an output layer. Each layer contains at least one neuron.

The principle of operation of artificial neural networks

Of course, if there are more layers, then the potential for solving ANN problems will increase, but it happens that the model becomes “big” for a given task, optimization to the required level in this situation is impossible, this is overfitting.

The basis for building an ANN is architecture, configuration, and choice of data processing algorithm. All these components determine the performance and efficiency of the model.

Now let's talk about the activation function, which is used to transform the weighted inputs of a neuron into its outputs.

Artificial Neural Networks are a powerful tool for solving problems, and as the number of component parts of a neural network model increases, it becomes confusing. This is where the “black box” approach comes to the rescue.

Firedrop

The designer is positioned as an AI solution; it does not have the usual editor with a set of widgets and settings. The interface is interesting, slightly futuristic, which is quite consistent with the image of AI solutions. Sacha's virtual assistant communicates with you via chat. He asks questions, the answers to which lead you to getting the website design you want. Also at the bottom of the chat there are links for quick access to the functionality of the editor - choosing a layout style, font scheme, adding new sections, and more.

How Firedrop works

Overall, Firedrop is perceived as a pleasant service, the functionality of which is presented in an interesting dialogue format with Sacha, a local type of AI. You don't need to look for where and how to add something to the site. It is enough to enter the direction of interest in the chat, and the assistant will provide a link to the corresponding opportunity. Sections can be added manually and content can be added to them in a simple and convenient editor. In fact, the role of the assistant comes down to the help center and the initial selection of style, image for the site header, title and other things like that.

In reality, Firedrop is an ordinary website builder with a pleasant interface and average functionality. Attempts to give the appearance of automaticity to the actions of the system lead to the fact that you simply cannot perform many of the actions usual for this type of system. The system tries to show you the important role of Sacha in creating a site by minimizing your actions in the editor. However, a simple algorithm works like a topic-based chatbot, answering questions based on the context of the task. No more. That is, there are few manual options for customizing the site, and the system cannot create anything interesting automatically. In general, all this is beautiful and useless.

The Grid

The service is positioned as an AI constructor in its purest form. The Grid does not have a trial version, it is only available by subscription. Molly is the name the developers gave to the virtual designer. Each page here must be created separately. The interface is strange, looks a little psychedelic - overly large elements, large blocks with thick frames, specific demo images and darkness all around. To be honest, it is more alarming than encouraging to interact.

To be honest, it is difficult to notice any reasonable actions on the part of the system. Much needs to be configured manually; automation is not visible. It feels like we're dealing with a clunky editor rather than some kind of AI.

How The Grid works

That is, in practice, you will make your own website, sequentially adjusting the appearance of the elements, and then Molly will show the result. This service was launched back in 2014 and raised serious expectations. But there is no AI in practice. There is a given algorithm of actions, which is inconvenient and ineffective. You do most of the work yourself. An unsuccessful service in every sense , since it grossly deceives expectations in terms of the work process and the result obtained.