Academician Alexander Kuleshov told Rusnano about how close humanity is to creating self-improving machines, what their creation will lead to, and whether Stephen Hawking was right when he feared the problems that intelligent machines pose.
Alexander Kuleshov is one of the leading Russian specialists in the creation of neural networks, artificial intelligence and complex information processing systems. Now he heads the Skolkovo Institute of Science and Technology, and until February of this year he headed the Institute for Information Transmission Problems of the Russian Academy of Sciences.
This Friday academician Kuleshov gave a lecture within the walls of the state corporation "Rusnano", at which he told the audience, including Anatoly Borisovich Chubais, about the progress in the field of creating artificial intelligence in recent years and about how AI technologies will change our society already through some years.
"Alien" or Human Intelligence?
“Why are artificial intelligence and smart data processing getting so much attention today? What happened? In fact, data has always been processed. Since the time of Galileo, the results of scientific experiments have been processed (mathematically). What has happened today that has pushed this problem to the fore?”, The Skoltech rector began his story.
As Academician Kuleshov notes, the amount of data that humans and computers work with today has changed - now computer programs collect, store and process terabytes and petabytes of data, the processing of which using traditional information analysis systems is extremely difficult.
People, for example, nuclear power plant operators or airplane pilots, have access to dozens or even hundreds of screens with various diagnostic information, each of which means almost nothing in itself, and will not help to find an error in the operation of equipment, but the combination of which with almost 100 % is likely to allow solving the problem even before it reaches a critical stage.
Naturally, the scientist continues, a person is not able to simultaneously monitor 50 screens, which gives rise to the need to create systems that would analyze this data and display on one screen only what is really important for making decisions and monitoring the situation.
Promotional video:
“Absolutely new mathematical systems that have appeared for the analysis of such 'big data' have grown beyond them, and they are applicable to the analysis of any information using any technical means. In fact, they would have been new in the 17th century and would have been useful to scientists of that time. But I emphasize that all this appeared precisely on the wave of new technologies,”continues Kuleshov.
Most of the discussions around these technologies, as the academician notes, stems from the fact that there is a difference between the Russian word "intellect" and the English word intelligence, which leads many participants in these disputes to believe that artificial intelligence should be some kind of anthropomorphic construction that resembles and imitates properties of human intelligence. In fact, according to Kuleshov, the last 25-30 years of research show that this approach is wrong and does not lead to meaningful results that can be applied in practice.
“Anthropomorphism and likeness of nature are popular terms, but nothing has ever worked out over the past centuries. For example, Leonardo da Vinci drew mechanical horses, Daedalus and Icarus tried to fly like birds, but nothing ever worked out - nowadays there are no mechanical horses running around our streets, and we fly differently. It's the same with the brain - those attempts to understand how the brain works, and to do the same in a computer, have completely failed,”adds the lecturer.
All these unsuccessful attempts to make hand-made analogs of neurons and connect them into a kind of brain, as well as other approaches that imitate the work of the human nervous system and the way we make decisions and analyze information, led to the fact that in the 90s of the last century the phrase "Artificial intelligence" among mathematicians has become a dirty word because of those unreasonable expectations, which carried anthropomorphic ideas about neural networks and artificial intelligence.
Depths of intelligence
In fact, the renaissance of "artificial intelligence" development began very recently, in the late 2000s, when a number of American and Russian mathematicians and programmers proposed and implemented AI algorithms, which later became known as "deep learning" and "diversity-based learning" methods.
“In the end, people began to forget about neural networks, it became clear that nothing was working with them, and everyone somehow missed the publication in 2005 of the article by Hinton and Krizhevsky, which now determines our future. I also participated in these “funerals”, but it turned out that it was not so simple,”explains the scientist.
As it turned out, simple neural networks combined in cascades and complex systems of differently arranged networks do not behave as scientists expected. And, as practice has shown, they are able to solve those tasks that were previously beyond the power of artificial intelligence, including speech recognition, photographs of people, various objects and even predicting breakdowns and disasters.
“A completely unique situation has arisen - no one today can say how deep neural networks work. The American defense agency DARPA is ready to issue a million dollar prize for explaining how they work, but I believe that this prize will remain unclaimed in the next 30-40 years. I know very serious mathematicians who struggle with this problem without the slightest success. We can say that we have returned to the days of natural philosophy - there is a certain method that works fantastically well, but we cannot explain why, says Kuleshov.
Deep neural networks, the scientist says, have long caught up and overtaken humans in many areas of knowledge, being able to identify and distinguish things that an ordinary, untrained person simply cannot do. The most recent versions of such neural networks make fewer mistakes than people trained to solve the tasks that such AI systems will be responsible for in the future.
For example, scientists have already created neural networks that can describe what is happening in photographs and videos no worse than a person does. Such algorithms can help blind or deaf people understand what is happening around them and what they cannot hear or see, and special services can use such networks to search for terrorists or suspects in video surveillance archives or during operational work at airports and other crowded places.
“There are about 70 million design engineers in the world today, and statistics show that only 20% of their products are some kind of new development. The remaining 80% were either already created by other engineers, or are minor modifications to existing models. Building an AI system that can find what you need will drastically reduce the time and resources that are usually spent developing them. There are no such systems yet, but in 1-2 years they will appear,”the academician continues.
According to him, another example of such systems is a program developed by graduate students of Kuleshov, which makes it possible to determine whether a person has Alzheimer's disease or not by studying photographs of his brain obtained using a magnetic resonance imaging scanner.
Only 200 MRI images of people suffering from this disease were enough for Russian scientists to "teach" artificial intelligence to distinguish between healthy and diseased brains with 90% accuracy. In a similar way, Russian mathematicians have learned to find ulcers in a person's stomach by his electrocardiogram.
In cooperation and on order from RSC Energia, Kuleshov and his colleagues have created a new revolutionary algorithm for controlling the ISS engines, which will reduce fuel costs for maintaining the station's altitude by about 40 times compared to the current program created by American scientists to replace the old Russian system, and five times better than NASA's upcoming program.
The new system, based on Diversity Learning technologies, will be tested on board the station next year. Another AI system, created by Russian mathematicians and programmers, is already working at Russian Railways and helps determine which breakdowns should be repaired in the first place to minimize resource costs.
Similar programs, according to the scientist, are sometimes used for the most unexpected purposes - for example, AI, created to render aircraft wings, is used by Louis Vuitton to create skin whitening creams.
“Further development of these technologies will radically change human life. Imagine, you are leaving a foreign hotel, you are accidentally photographed by tourists, this picture gets into a search engine, it "calculates" you on these pictures and in five minutes your boss will find out about it. As a result, it will be very difficult for you to convince him that you went on a 'local' business trip,”explains Kuleshov.
Augmented archaic reality
The first examples of this "new, wonderful world" exist today - it is the AI system AlphaGo, which beat the world champion in Go this year. As Kuleshov explains, it is the first example of a unique class of machines capable of solving incalculable problems and improving themselves.
“Go differs from chess in that this game is simply impossible to calculate mathematically. The number of possible moves in Go exceeds the number of atoms in the Universe, it is impossible to stupidly count the moves in it. In chess, if you have a powerful computer, then you will beat anyone, both Kasparov and Karjakin. This is impossible in Go, because no computer can do it. And the neural network was able to solve this problem,”says the scientist.
The main distinguishing feature of AlphaGo from all other AI systems is that this program can play with itself and improve itself, adapting to the opponent and finding absolutely non-trivial and unexpected ways for a person to beat him.
“Why am I stopping at this - this is the first step into a completely mysterious future. How was AlphaGo born? First, its creators collected a database of 30 million different game positions, and trained the primary neural network on it. Then they duplicated it, and the second network started playing from the first. And as a result, after several billion iterations, something third arose that a person no longer controls. It is not clear where it came from - this is the result of some self-construction. Nobody knows how it happens,”Kuleshov emphasizes.
The birth of AlphaGo and its victory, according to the academician, opens the door to a completely new space, into which humanity will enter very quickly. And not everything in this world will be useful and pleasant for humanity in general and individuals in particular.
“It is clear that the social shifts from this will be enormous. The number of semi-skilled workers is already dwindling like shagreen leather, and the emergence of AI capable of solving these problems will deprive them of their jobs. All these engineers, taxi drivers, pilots, nurses, workers - millions of people - will have to disappear, and only 1%, as current studies show, can adapt to new realities and retrain, says the scientist.
According to him, “we are on the verge of absolutely monstrous social consequences from the development of artificial intelligence systems. We cannot now assess their scale, like people in the middle of a hurricane or at the height of a revolution. Money now needs to be urgently invested in education, as people of average qualification are becoming completely unnecessary."
As the rector of Skoltech notes, the world today is able to feed all of humanity, but it is not able to occupy it. This unemployment and lack of purpose in life could already affect the life of Europe and other developed countries and give rise to various radical movements like IS and other banned extremist and religious groups.
“This is a conscious archaization of life, the creation of a situation in which I will feel needed. To hell with him that I live worse, but I do not live like everyone else. The feeling that you are constantly being sent fast food for free and given sneakers every six months, but at the same time you are not needed for anything, is actually terrible. And this feeling will only grow with the development of AI and robotics,”continues Kuleshov.
A noticeable part of this problem is related to the fact that a person simply does not have time to "evolve" after AI - generations of people change every 25 years, and technological revolutions occur with an interval of 5-6 years. Therefore, as the rector notes, the number of "unnecessary" people will constantly grow, and only mass education can help avoid a social explosion and the emergence of a new wave of Luddites.
“What we are on the verge of has no name yet, and I don’t even know what to call it. Perhaps they can be called "unmanaged intelligent systems". These are fundamentally new systems that generate themselves, and we are not far from the time when they begin to penetrate into our lives,”the scientist concludes.