One of the most interesting and most promising areas of scientific research worldwide is related to artificial intelligence. Once more the importance of this topic was stressed by the President of Russia, Vladimir Putin. Holding an open lesson for all of Russia entitled “Russia, Looking to the Future” on the Day of Knowledge he noted:
“He who becomes the leader in the artificial intelligence will be the ruler of the world. To not fall behind, we need to work on this right away.”
SUSU Vice-Rector for Informatization, Doctor of Physical and Mathematical Sciences Professor, Head of the Department of System Programming, Leonid Borisovich Sokolinsky tells us about the promising work being done at South Ural State University in this field.
– Leonid Borisovich, what is artificial intelligence? How can we know for sure if it has already been created?
– Really, there are many definitions. At this time, it is usually understood as mechanical or digital devices able to complete creative work, for example, write poetry, draw pictures, or hold a dialogue. In 1950 famous British mathematician, logician, and cryptographer Alan Turing formed a definition of artificial intelligence in an article entitled “Computing Machinery and Intelligence”, suggesting that it would be realized using computing machinery. This researcher thought up a test. Imagine two rooms isolated from one another. In the first one there is a computer, in the second one – two. Those two are connected to the first one by local area network. A person completing the test sits in one room. In the second room, another person is collecting answers to the questions sent from the first computer. These people do not see each other. There is a program on the third computer for artificial intelligence, which is also answering questions sent from the first computer. If the person from the first room can’t determine where a human is answering and where a computer is answering, that would mean that artificial intelligence has been created. But this formalized definition of artificial intelligence is, unfortunately, very loose and inaccurate.
Modern computers, even the most powerful ones, do not have intelligence. All that they can do is execute programs written by humans – sometimes extremely complex ones. It’s well known that right now computers can play chess very well. For example, in 1997 the computer Deep Blue defeated the World Champion Garry Kasparov. Can we say that this computer had artificial intelligence? Only partially, since there is a very strong program working there, which was worked on by a group of genius programmers. But this program works by different rules – there’s a different mechanism for making decisions which differs from the one that human mind uses when playing chess.
It became clear quite quickly: if we isolate this program, if we don’t give the programmers access to it, then a grand master, master of sport, and even simply a very talented chess player can quickly find its weak spots where it will regularly make mistakes. This is a principal issue: unlike a human, a program will begin to make mistakes in particular moves and lose every time, because without the help of programmers, independently, it is not able to learn. The programmers can find mistakes, make some corrections, finish writing some codes, or improve the program over the span of several years. And this is a huge amount of effort – but this is not artificial intelligence.
– And what is your opinion – what is artificial intelligence?
– I would give the following definition: artificial intelligence is an intellect based on the realization of artificial neural networks and training them. Artificial neural networks are quite complex programs which do the same work as neurons through mathematical models.
At the same time, I would not yet talk about the artificial intelligence which creates poems like Shakespeare, books like Lev Tolstoy, and pictures like Renoir, or plays chess like Alekhine. First let’s talk about artificial intelligence able to complete tasks manageable by an ant or a bee. For example, a robot must get to a certain point after having chosen the optimal path. Directions can be given in different ways: for instance, by magnetic fields, by which, as we know, many living organisms are able to orient themselves, including insects. This robot can be equipped with a relatively small computer the size of a smartphone with low energy use on which an artificial neural network is realized.
One very important thing: such networks must be taught. It is relatively simple to create a neural network. The main difficulty is teaching it to solve tasks, even relatively simple ones that an ant is capable of: for example, getting from point A to point B while avoiding obstacles. And then, in the process of building this artificial neural network (ANN), it’s possible to tackle more difficult tasks.
– How are artificial neural networks built?
– Neurons are connected by bonds, which have weight. Signals enter the neural network – for example, visual images. For output there are a few neurons, among which only one of them will work, which classifies this image. That is to say, at this time, computers model a very small mind.
The bonds have their weight, which is able to strengthen or, conversely, weaken the signal sent to them. A photograph of an animal is input, for example, and the number of output neurons is equal to the number of pixels in this picture. The neurons which receive these pixels signal the next neurons. Some bonds strengthen this signal, others on the other hand, weaken it. Accordingly, neurons having received a signal generate a signal themselves for other neurons. This excitation is transferred along a chain from neuron to neuron, and there are three of them at output: if one is excited, the photo was a cat, if the second one is excited – a dog, if the third was excited – it’s some unrecognizable object. Real working neuralnetworks can solve such tasks today.
– Can you name some examples of application for such networks today?
– Yes, these include recognition of images, road signs, emotions, handwritten text, objects and obstacles in front of a robot, and creation of a path by self-driving cars. For example, an artificial neural network can react to gestures, mime, or poses experienced during psychological testing. In this way it is possible to determine, for example, if a person is smiling or not. To solve such tasks we need artificial neural networks which consist of a few million neural nodes. For comparison: the human brain has around 50-100 billion neurons.
I’d like to note that artificial neural networks have been talked about for a fairly long time already. The first theoretical work in this field was published by American researchers McCulloch and Pitts in 1943. After this, much theoretical work in this field has come to fruition. However, it was impossible to translate theory into practice until recently, since there were no computers strong enough to realize this and, most importantly and most difficultly, train artificial neural networks.
– What are the difficulties?
– To create an artificial neural network, you must think up a proper structure for it, otherwise the task will be solved in a dissatisfactory way or will not be solved at all. Beyond that, there are no algorithms or clear methods for how artificial neural networks should be built to solve one task or another. Researchers are moving along by touch through trial and error. This is the first issue. Nonetheless, today quite a lot of success has been achieved.
Next. A newly created artificial neural network is clean, so to speak, there is no knowledge accumulated in it yet, which means that it cannot solve any tasks. Do you remember how the main character in the film The Fifth Element learns by reading an electronic encyclopedia? After creating an artificial neural network, it needs to be taught solutions to concrete tasks the same way. During this learning the weights of the bonds between neurons which can strengthen or weaken the signal are set. There are a few methods for teaching artificial neural networks. One of the main methods is to form a very large database of specimens – thousands and even millions of them. For each of them we must write the correct answer, which the neural network must generate. The task is physically and technically simple, but it is very resource-intensive. To teach a neural network, for example, to differ a dog from a cat, people must input millions of pictures of dogs and cats and note in each one – this is a dog, this is a cat. This is quite difficult work. This is a method called learning with a teacher. There are neural network learning methods that do not require a teacher, but they don’t fit all tasks. This is the second problem.
Then learning begins: random samples are put in to the networks and they give an answer, for example, an incorrect one, which is compared with the correct answer, and by certain mathematical algorithms the internal bond weights and networks are reset. To teach an artificial neural network, you need a very strong calculator – a supercomputer. It inputs the digitized samples into the ANN, the output answer is compared with the correct answer, and it tells the teaching program how strongly the artificial neural network was mistaken. The teaching program creates the weights settings. Even in modern supercomputers, this process takes days and weeks. On the very strongest supercomputers created 20 years ago teaching an artificial neural network would take a hundred or even two hundred years. This is the third problem. So, systems based on ANN appeared only recently, when humanity learned to create strong enough processors and supercomputers which are used for ANN teaching.
– And could you go a different way: teach a computer to differentiate a dog from a cat without creating such huge databases?
– There are some very complex mathematical algorithms which are able to solve these tasks without such databases – without learning through photos. They can determine the contour of something, compare the geometry of, for example, an animal’s ear – that is they can build some kind of geometric model of a cat. These algorithms are well known and are quite strong – but they can only solve the tasks that are input to them. But artificial neural networks are a universal instrument. Yes, they must be taught, but potentially, they are able to solve a wide range of problems characteristic for artificial intelligence and which are very difficult to solve using computational algorithms. For example, if a cat is standing sideways or is sitting with its back to you – this is already a difficult mathematical task. If you teach an artificial neural network to differentiate dogs, cats, road signs, and handwriting, it would only need a processor like those that are used in modern smartphones.
– And can this artificial neural network learn how to complete several tasks? Or is it only able to do one thing at a time?
– Everything depends on the quantity of artificial neurons, on the sample database, and on the power of the computing technology with which the teaching is performed. At present, artificial neural networks are taught a solution to one particular task, since this is complex and takes a lot of work. But this is a matter of time. It has only been five years since the computing technology appeared with which it is possible to teach artificial neural networks. Before that, teaching a neural network even one task was impossible – it would have taken tens of years. But the process is not stopping, the technology is developing further. We are waiting for the appearance of exaflop supercomputers in a few years (around 2020), which will be able to complete 1018operations per second. Then it will be possible to teach ANNs to solve not just one task, but many.
– What living organism can we compare modern artificial neural networks to?
– The most primitive living things that we see all around us are insects – however they solve more difficult tasks than the ones described above. That is, a brain of an ant or a beetle, right now, is still stronger than what men can create. Artificial neural networks are just getting closer to the potential of insects. Modeling a human brain is still a remote future.
– You mean that artificial intelligence like we’ve seen in science fiction will not be created soon, and computers will not be enslaving humanity?
– The kind of artificial intelligence that exists in science fiction novels and films is definitely not appearing in the near future. It is more prudent to worry about a different thing – the spread of computer viruses. Yes, viruses are written by humans, but they are able to function on their own, infect different computers, and this can lead to and bring serious consequences. For example, they could seriously damage things if some object of infrastructure, manufacturing, or bank system is controlled by computers. By the danger they pose and ability to multiply, computer viruses are like ordinary biological ones. But this is a topic for another conversation.
– Could you tell us more about who in SUSU is working on artificial intelligence and in what concrete fields?
– We have begun working in this field relatively recently. This is because artificial neural networks, which are prototypes of true artificial intelligence, began developing very quickly thanks to the successes of computer science. The International Research Council, created for the purpose of enabling the realization of the Russian Academic Excellence Program under Project 5-100 (first meeting of the Council at SUSU was held in autumn of 2016), recommended that we work on this problem in particular. Rector Alexander Shestakov set this task, and we began moving in this direction. In particular, I changed my scientific specialty – all my life I have worked on database systems, and now I deal with artificial intelligence systems. At the Department of System Programming, the Head of which I am, this area has begun to develop rapidly, many theses and dissertations in this topic are being completed. This issue is also being worked on at other SUSU faculties. Work in artificial neural networks is being completed at the Department of Information-Measurement Systems, which is headed by Alexander Leonidovich Shestakov. There, ANN are used for reducing mistakes of dynamic measurement created by sensors on various technical equipment. Work is also being done at the Department of Automation and Control (the Head is Doctor of Technical Sciences, Vladimir Ivanovich Shiryaev): artificial neural networks are used to choose where to land equipment on space bodies. Several artificial neural network projects are being completed at the Department of System Programming.
– Please describe them further.
– With pleasure. The first project is titled “Development of computer vision systems for industrial quality control.” This involves creation and implementation of new classes of industrial equipment and technological systems for automatic detection and classification of defects, diagnosing possible reasons and locations for their appearance on automatic production lines for industrial goods. With the help of web cameras, artificial neural networks can automatically keep track of defective goods using so-called computer vision.
A big team which I am the leader of is working on this project. The project has three fields. The first one is called “Software and hardware support for creating and teaching deep neural networks”, and is supervised by the Head of the SUSU Laboratory of Supercomputer Modeling Pavel Sergeevich Kostenetskiy, and under his guidance, programmers and developers of neural networks Roman Andreevich Chulkevich, Aleksandr Igorevich Rekachinskiy, Anton Olegovich Golovenko, and a student Rustem Alkapov are working. The field called “Client-server software” is supervised by a research engineer from the Department of Information-Measurement Systems, Ivan Sergeevich Nikitin, and a programming student Nikita Vetoshkin. Finally, research in the field “Videoblock” is being completed under the guidance of Deputy Director of the Academic Techniques and Technologies Scientific and Manufacturing Institute (SMI), Rustam Zaynageddinovich Husainov, and a development engineer Sergey Evgenievich Kulikov is the Head of the Laboratory of Automation of the same SMI. As you see, students are also involved in this work. Magnitogorsk Metallurgical Factory acts as a client.
The second project, “Detection, identification, and tracking of targets”, involves development of an automated system for artificial compound (based on insect eyes) computer vision, able to detect, recognize, and track certain kinds of objects, determining their direction and the distance to them in real time. This system can be used to create fully automated robotic equipment able to move across land and air and complete special tasks. The area of application could be extremely broad. Here we are talking about the creation of an artificial compound eye. The natural analog is the compound eyes of insects. Despite their apparent simplicity, they have a lot of capabilities, they allow their holder to complete difficult tasks: for example, fly very fast or maneuver. Flies can see very well: delivery of optical signals is 100 times faster than that of humans. Using such artificial eyes, we can identify objects that enter the field of vision, determine their direction of movement, and the distance to them. For example, such a device can be installed on drones or quadcopters. They could fly along the perimeter of an airport and keep track of other drones, which are flown into the terminal airspace by hooligans or terrorists, and neutralize them. Master’s student Ilya Adigamov and me are working on this project. We are planning on entering into a partnership with the Ural Optical-Mechanical Factory.
The third project, “Isolating human voices in noisy environments”, involves creation of complex systems, including stereo microphones, cellphones, and a cloud-based data processing system, which will allow us to selectively filter out the speech of a particular person in a noisy environment. There is one well-known phenomenon at parties: a person is able to keep track of the speech of his/her conversation partner, isolating it from many different sounds – for example, in a restaurant, when many people are speaking, music is playing, and so on. The goal of the project is to create an artificial neural network which will be combined with two stereo microphones and will selectively filter the speech of a particular person in a noisy environment. Why do we need this? This development can have various applications. For example, we have a digital sample of a person’s voice without external noises. Then, if this person is speaking with a different person, for example, on a street where cars are making noise, a taught artificial neural network can filter these street noises out and give us a clean recording of the voice. In my opinion, other applications are extremely interesting and very important – for hearing aids. It is well known that many older people have poor hearing. Modern technologies allow us to create miniature hearing aids – but these devices equally amplify both necessary and unnecessary noises (cars, construction, speakers, and so on). And a brain, when a hearing aid is used, loses its ability to isolate the speaking partner’s voice from background noise. As a result, a person with damaged hearing, standing, for example, on the street, can’t figure out what people are telling him, and turning up the volume in the hearing aid doesn’t help. There are two ways out of this: either create a hearing aid that completely imitates human hearing organ – this is a difficult task, more difficult than transplanting a lens of an eye; or, teach hearing aids to filter out background noises, which is what we are working on. This development is helpful also for filtering out background noises on the telephone. Or, for example, it can be used for lifeguards, when they need to isolate a drowning person’s voice among the sounds of the ocean waves. Maksim Vladimirovich Gubin, Professor of the SUSU Branch in Zlatoust, is involved in this project, which is headed by me.
The fourth project – “Intellectual systems for protection against DDoS attacks” involves analyzing requests sent to an internet site in real time, determining anomalies characteristic of DDoS attacks, and automatically implementing adequate means of protection. This is an intelligent system for protecting websites against attacks from hackers (DDoS attacks). Here we are speaking of creation of an artificial neural network, which will determine suspiciously large activity or a jump in the number of requests, and automatically block these requests coming from appropriate IP-addresses, so that the site remains in working condition. It is well known, that hackers can initiate a large number of requests to a site from several computers and paralyze it. To avoid this, these requests must be identified and turned off. The head of this project is Senior Professor of our Department, Kseniya Yurevna Nikolskaya.
Finally, the fifth project, “Diagnosing individual Jung psychological types”, involves developing and realizing methods and algorithms for automatically determining a person’s psychotype based on analyzing video clips of psychological testing of a subject. Imagine: standard tests are being held while choosing candidates for an open work space, an interviewee is given questions, a video is being taken, and on this video you can see the gestures and body language of the individual. These recordings are input into an artificial neural network, which knows which questions were asked, and what the reaction was to them, and it can say, for example, if the candidate is disposed to lying, if he is disciplined or not, etc. Associate Professors of our Department, Valentin Aleksandrovich Golodov and Sergey Aleksandrovich Ivanov, are working on this project.
– Are there other clients ordering research in artificial intelligence?
– Yes, of course. These include the Ministry of Education and Science and the Ministry of Industry and Trade, the National Technological Initiative (the NeuroNet platform), the Skolkovo Fund, the Center for Strategic Developments Fund, and the Rostech State Corporation. In the future, I’d like to ensure partnership in the previously mentioned projects with airports in Russia and abroad, telephone and hearing aid manufacturers, as well as recruiting and HR agencies and services.
– Are there any scientific or other organizations partnering with SUSU in working on artificial intelligence? Who would you like to partner with?
– Yes, there are such partners. The most important of them is the Korean Institute of Advanced Study (the Director is Professor Jaewan Kim, member of the International Research Council of our University). We have been working for a long time, and with great success, with such large partners as Intel and Nvidia, who are the developers of hardware and software solutions for creation and teaching of artificial neural networks.
In the future, I’d like to work in this area with such centers of science as the Massachusetts Institute of Technology (the President is Professor Rafael Reif), the University of Tokyo (the President is Professor Makoto Gonokami, Technical University of Munich (the President is Professor Wolfgang Herrmann), and Harbin Institute of Technology (the President is Professor Wang Zhaoguo).
– What would such partnerships give?
– A lot, including the ability to actively push our ideas forward in highly-rated scientific journals in the Scopus and Web of Science bases. This allows us to show our authority and improve SUSU’s position in various ratings, to increase awareness of the university. We will exchange ideas and experiences. Of course, we won’t share our secrets, and the ideas and developments which can be implemented into real manufacturing will be protected by patents.