Head of the laboratory of big Data Analysis methods, faculty of Computer Science , Higher School of economics denis aleksandrovich derkach denis aleksandrovich hello and thank you very much for taking the time to come to the studio. Hello, thank you very much for the invitation. Where to begin . Lets start with , um, what can now be imitated using the means that, well, you and your colleagues are conventional. Do you have a voice, an image of a person , facial expressions, gestures, gait, uh, in general, thats how far Literary Works are styles, what how would you describe it . Here is an achievement over the last i dont know a year, a couple of years, six months. What, where the computer catches up with us and ceases to differ from what they do from what is, as if in reality we will return, then to what it really is, well, in fact, we are already, in principle, surrounded by such things, which have already been generated. E neurons. That is, when you call somewhere on the phone, quite often you hear a robot and this robot has a voice that was somehow recorded and generated like this usually a robot. Well, uh, it depends, its possible that what youre saying is, uh, the technology is two years old or Something Like that, and the level is but basically, uh, the further we develop, the greater the chances that you will encounter in life with a robot, not a person. Well, thats what im facing. Eh, wait means eh, lets face it this way. Uh, using vision and hearing. Uh, reading uh. And burn like that. Smart collision is correct. I understand you. Yes, of course, because, in principle, many technologies are now aimed at something to trick the human brain. So we want to make sure that the human brain does not distinguish. Uh, something that doesnt actually exist from uh something that exists and its actually quite simple because our brains arent designed to. He is not designed to distinguish between these and this. What does our brain do . He usually looks with uh, where is some kind of fruit or where is the predator . Well, as usual, yes, how all this developed for us. But we, in principle, cannot, in reality, do what answer e, what was not generated by the Russian Network is almost similar to reality from reality. Well, please take a look, here is a brief excursion into the history of what Generative Networks looked like in 2014. And what they generated and uh, for example, what they could generate in 2018, it is clear that they are already generating, not just some blurry picture, but a completely normal person. Eh, its quite difficult to distinguish from reality. And in principle, if you look closely, of course, for the 2018 picture. Uh, you may notice some features that will allow you to perhaps, uh, tell that this is not a real person. Well, there are blurry parts somewhere. Uh, blurry parts of the background that blends into the hair, for example. And somewhere there may be, uh, an extra amount, and the eye or there or a slightly irregularly shaped pupil, this happens to a person, but very rarely. But for a Neural Network this is a completely normal situation. But this is, uh, a thing of the past now. Thats what its been like here for 18 years, yes, yes. Of course, eighteen years old is already a fairly old image. But its already typical characteristically we see that we need to peer at the picture for a long time in order to understand that there is something wrong in the picture. Well, please tell me, how about this video of the movement of changing poses, this is the dynamics. You said that our brain is not trained , some things are not trained to fix, and some are trained just for movement and gait. We determine this in people and, uh, we read a persons mood, for example, because he walks, as with this now. A quite good. That is, how would Neural Networks already learn or already they have learned to completely deceive us there too, and the main idea is that and we need as much data as possible. That is, we need a data bank, but these campaigns so that the Neural Network can learn to imitate this selfpropelled gun on video, for example, and when we say generative Neural Network. What does this mean, accuracy . But in principle, this is the same Neural Network as usual, just the only thing is that it is trained to show some new objects that are similar to what is in the training set, that is, we started with moreover, relatively speaking, correct me, we distinguished the image, learned to see the image, the differences were then decided. Why not just watch, but also create. Yes, we actually, this picture here is actually a good example of, uh, how we came to such models, which are very complex, which gave beautiful pictures. In fact, in this case, we set two Neural Networks to train with each other. One tried to distinguish good from bad, for example, generated from real data, and the second, accordingly, learned to generate something that would deceive the first. Please tell me when we ask her to generate something under such conditions and that the results are, for example, people, and not monsters. This is the result of simply training and our hopes that if it has already turned out to be people and not monsters 100,000 times, then it will still be people, will she be wrong . Or do we have some kind of guarantee that it will produce exactly this and there will not be any strange emissions of fluctuations of something. No, no impossible. This is because she doesnt know what is possible and what is not. This is a set of some crazy numbers, so how should we treat this . Yes, by the way, this is a very good question. In fact. And the Neural Network really doesnt know whats going on with the picture. And what are we going to do with this picture . And uh, to what extent does the fact that she generated reality exist . Well , that is, its like reality it looks like it exists, but for a Neural Network its very easy. Is it possible to create a person with a stirrup, or a cat . Oh with wings or or Something Else and how we ensure that the text is a Neural Network, except for what you use it, that they do not generate it, that is, uh, you dont want to say that we need control over what it will do. Do we like it or not . Because maybe we can control something superficial, but we cant control the depths. How to control that she does what we want, at the same time, seemingly, showing some kind of independence, and on the third hand , guided by the general ideas that would have been put into it initially, and we try control. E through learning, that is, in fact, our learning process, and it correlates with what was shown by the expert. Our Network Training process is correlated with that. And what, uh, seems real to us, that is, uh, we specifically created a learning process that, uh, selects for us, uh, the correct pictures are far away, these pictures will not always be correct, but in principle, if look at the first articles on Generative Networks, and and there was one of the criteria. Uh, something that a person could not recognize, that is, accordingly they showed it to an expert or just , uh, ordinary people. And they said this picture of a real person or not, and uh, then they looked at how the experts responded, how people responded. A and. Eh, then we looked at how this is a criterion for the quality of Network Training. Ahh, it correlates with what responds to people. And that they learned to deceive all people, yes, well, they didnt learn to deceive all people, but they learned to somehow deceive, like they all do. Well, at the very beginning it was now at this moment. Now look there. Eh, here, again, in the first place, going back to the first articles, and there was a very cool criterion. Uh, this criterion was the amount of time you look at this picture and there it was shown to the same person. Eh, several pictures on one could be looked at for half a second , then another for one second, for a third and a half seconds. And of course, the person understood in a second and a half. Well, this picture is wrong. Yes, now there are no salt pictures that are at least an hour. Look, you wont understand. And now there are such people, but its impossible to say that a Neural Network can generate any picture, and it will be super believable. That is, we always have some errors in the Neural Network, by which they can be caught; another thing is that these errors are good in the Neural Network. Eh, they dont do it all the time. And then lets say once in a thousand or once in a million, not that there is a small error everywhere, but errors still suddenly appear, these outliers. No matter how much you train, sometimes some combination of factors happens, since we have no guarantee, from this there is only hope that since she did it just once, and she she did it well, she does it a thousand times, 999 times well, and for the thousandth time, conventionally, she gives out some kind of nonsense. Well, about then. That is, as if well trained. It is not russia that will give you food. Sometimes we install a second Neural Network that recognizes this or not, or is it impossible, and then these two weeks the networks act like reinforced concrete, huh . Eh, in principle, yes, but not necessarily. That is , uh, in principle, it is possible to distinguish using Neural Networks. Uh, what one Neural Network generated, but now Neural Network training procedure. These differences are also included. That is, its like, that is, they are trying to compete with each other. Uh, trying somehow. Find the very minimum and, accordingly, get the correct picture. What do you think . In what sense . Is she creating something new . Well, this is a non existent person. In what sense . This is new in what sense . She can write a new literary work. To start. Let it be small. In what sense is it new and how does this relate to when a person doing something new and creative. Ah, look. Uh, so i said that it does interpolation, yes, there we have, uh, a Neural Network that takes some numbers and somehow presents a picture and represents the text in the form of numbers and inside this space. Shes doing interpolation there, shes doing it. Eh, there is some kind of specific generation of holes that they see. Well, for example, yes, its just another thing that for us these holes can sometimes be something new. Yeah, but sometimes. Uh, sometimes sometimes its just a mixture of several things. Uh, dont know about a few photos. That is, it s just that, depending on how she learned, in principle. Uh, if we look at these are our real objects, then we can hope that this is super. Well, its not the Russian Network that generates. So we have hope that it will generate more, uh, classical objects. But the super new one does not generate, but uh, here we need to understand what is for us. What is in our object, for example, a photograph, is super new. Yes there is Something Like that. Heres a super new one that never happened at all. But in principle, in the understanding of a Neural Network, this could be something old, that is, in the sense that it is a drill obtained by interpolation. This is a cat with very st with square ears. Eh, well , for example, yes, that is, like a cat with square ears, which nevertheless was bred or can be bred. But look, lets take some literary work, lets assume that they are writing russian ones. They seem to be able to write them, and great Literary Works were a cat with square ears. They were up in arms against them; contemporaries said that this was not literature at all. Well, its the same with music, whats that there . It is not clear that different words were pronounced there, and then it turned out that this was a step forward in literature. There in human development, in culture in general. Here its like, if we do something and suddenly, uh, everyone will say, this is brilliant, this is super. Nobody has ever done this. Well, you can and in this sense she is a better person. Can you imagine a moment when Neural Networks in this sense will create a new genre will create some new images that will appeal to people and will be perceived as something new and, in general, close to human nature and close to us. Yes , look, lets assume that we have a dictionary, without Neural Networks at all, without anything. Just a dictionary, which means we begin to randomly compose words. Yes. Now, can we compose shakespeares . Well, you need to try a lot. Yes, in the life of the universe, the Neural Network reduces it a little. That is, she, for example, knows after which word it should not be placed. Eh, another word here. Well, or they put the usual one there. Eh, accordingly she shortens this set of texts. And some of them will be brilliant, some will mean nothing and accordingly, it will work out. It accelerates the opportunity to create a new genre. Thats another thing, at the moment, at least. Ah, a persons task is not to generate new text, yes, that is, to generate, of course, there we have graphomaniacs who generate new text every day. They get up there in the morning and do something. Heres the task good writing is about catching, uh, something. Uh, for example, here in history in a historical event that is happening now or catching something in the future. Soon the future or there in 100 years , depending on the Science Fiction writer, for example, some kind and write a text that answers, uh, some questions. Well, Neural Networks. This is even easier to do. After all, i have. Teach her that she just needs to be able to highlight these columns, which may not be columns at all, but Something Else, and she will be great at plugging holes and making new ones. Catch trends. Yes, but she cant choose. The only thing is new. These are the great ones. There are artists, writers, rivers , musicians, they knew how, from straw. And, perhaps their generated ones. And here are these sets of words. They chose the same set of words that are right here in the dust. Ingeniously , a sculpture is created according to the principle of removing it from stone. Everything is unnecessary. Exactly yes, thats roughly what we do there. This is when we write, for example, a great work. I dont know how much more you need me. Lets put the second third university, which will watch the first one so that she creates a great work. You see, what i am leading to. I am leading to where the actual boundaries of the human are and where they can be invaded. So that people will be forced to admit that this is a thing that has no brains and is engaged in plugging a column. The way you teach is no worse. We are indistinguishable from us, but look at the basic texts. Uh, in principle, uh can be generated using a Neural Network and hmm, people who, uh, for example, dont look at the text for long. You just have to get closer with your eyes. They are no longer different. Yes, by and large, but again you need the only thing that you need to install a proofreader or some kind of editorinchief who will check that the Neural Network has not done something completely stupid , remember . Yes, i initially said that its even a good Neural Network. It may be wrong once in a thousand or once in a million. So, accordingly, you need a person to read and say, well, here, here , there is some kind of stupidity, there is something wrong. Well, wait this way, then james joyce sends his hand. That is, the editor tells you, come on, cross out this, this, this , this, chapters are not written like that. You cant use that kind of language here, and so on. He knows within himself roughly that he is a genius. And that he wants to write like that. Eh, right here. Eh, how would i put it, but in this way. We dont let her create Neural Networks. We simply artificially keep it within some framework, putting human control over it. Or maybe she will actually come up with everything better than us. How do we know this . She doesnt come up with any better ideas. She hes probably making it up. Eh, more quickly. Here, and here is the quality, that is, we really need to say what to say. Yes we need to figure it out. Distractions, uh, a good Neural Network of a brilliant text, uh, from a text that is almost brilliant, but its like something that you can buy consumer goods somewhere in a simple book stall near the metro, and here it is accordingly. And which method is not very good, its clear, for example, what i did. What made, for example, parisian artists great, but brussels artists not so much. Well, yeah, uh, joan rowling wrote under a pseudonym and no one began to publish it until they found out that it was her, but here it is very difficult for us. Its very difficult for us. In fact, this is the last rule. Uh, uh, well, that is, in fact, its very difficult for us to come up with these last rules. That is, this is some kind of human feeling. Here he looks, perhaps he was inspired by the car that splashed him in the puddle. And so he made a novel about how a little person. It started with this is exactly where it started. E splashing. It turned out better there, yes, ah happened. Yes, but Something Like this happened. What are you saying, you are saying that the network cannot do this. And i thought that you came here to say that on this network we dont train, it doesnt cost us anything to shut up the same idiots. And we can, we can train the network so that, for example, it generates a million texts for you. Or , for example, no, no, 20. Well, generally super brilliant. But whoever will say in general brilliant, at least three of them will say. Well, at least one super genius, well, we need to make sure that genius was in place. And i say that genius is often out of place. Well, why did you bring it here now . Example of james joyce, yes. Well, he was. He was a revolutionary, right . In what sense, that is , he had some kind of fox there. Yes, there are ways of owning the text that were inaccessible to the people around him, yes, but nevertheless, in a few years , in 100 years. For example, we think that he is a genius. Yes there or there, right . Well, even in two, well, or in 20 years. Yes he was he was already recognized as a genius there and, in general, in principle, everything was pretty good. But for us, well, in principle, this does not mean that besides james joyce there were no other writers who could also write brilliantly and, accordingly, it just happened that way. Here she also coincided with him well, but looked at ireland, looked at what was happening in ireland and , based on this, she wrote, that is, say no, and you say so in the next thing. You say that they are growing up, they just dont have enough Life Experience. Well, somehow, yes, then there is, in principle, in principle, you can imagine that you enter, for example, eh. I saw this today. Uh, such a long long ride there, like a subway ride, then i had some kind of bus ride there, and then again i read some book, some scientific article, and uh, at the end of the day. I say the Neural Network will generate for me, uh, a Short Description of my day. Here is a Short Description of my day, then someone else will read it and accordingly will be influenced by the Short Description. They are my day because i, well, that is, wait, we we are working on this so that the university has Life Experience, and they are already just like people. And n why, that is, how could we already in the sense that uh yes and no. That is, as if we can already have some thoughts in the form of outside the network. This is called industrial engineering. Yes, that is, but here are modern text generators. We enter some line and based on the line the network begins to generate. Text, that is, we can ask a question. Or, for example, lets say we continue the work we started, and it begins to generate. This is what we started work, but i want to be attached to talking about the example of writers. Lets make Neural Networks a real life. We will provide her with experience every day. Well , of course, we will do neither daily. We will not stretch it out for 20 years until she grows up and matures. And what is there, it is clear that we will compress this in time on our chest into it immediately, but she will receive Life Experience and will analyze it. Approximately the same as we do, but it will begin to treat and begin to produce the same thing as we can do on whether to train this or not is unclear. How to teach, and we know, how to teach her the correct reactions to imitation of Life Experience. We know how to train a Neural Network in this way. Yes, that is , it is not so scary, maybe we are not training them now as well as we will be in 10 years. But in principle, we understand how all this about chinese in front of you is artificial. Ah, well, if you understand, she behaves like a person, has Life Experience , thinks like a person, talks like a person, but on the phone talking like a person writes may not be leo tolstoy writes letters as a person with some kind of character of his own. Um, when did she get sprayed . Is there a car . Uh, from a puddle. She reacts to it there in one way or another, like this or that person, what about this . A . Well , basically, yes, she is three, she seems to walk like a duck quacks. Here. Well, it seems like she really is a duck. And here. Well, then this is a very interesting question. That is, this is, in principle, uh, here i lack philosopher, philosophical knowledge. No, there is no such thing yet. Or it will be around tomorrow. But look, as i said, yes, with us, for example, about the text. Yes, e we can generate. Uh, some text. Yes, a and this text, in principle, fits quite well, for example, under a column in evening newspapers, in which there is a stove. And we generate quite normal horoscopes. And in principle, uh, we, uh, accordingly, can now kind of give uh on the latest uh Neural Networks, uh, and videos and pictures, that is, it already turns out that hmm how how in some sense, we can uh imitate e person. Heres the basic one level yes, but another thing is that at some point we can, we can get lost, that is, we can say. But here you can see that this one is not a human after all, he wrote human things and did human things. So , accordingly, it turns out that, well, we are close to this, well, i cant say that we are that close. That is , maybe five years, maybe here. Please tell me a conclusion with your experience from your point of view. Well, you are dealing with these artificial formations, which are the very things in a person that make him a person, that the grid will be more difficult. They may never achieve this in total. Or it can be achieved, but if so, then last of all, what is most characteristic of a person. And in a person, uh, the most characteristic thing is the novelty, that is, uh, the grid interpolates, and a person, in principle, can achieve. Heres some here with the correct extrapolation, that is, he seems to be saying that here with uh, here here you need to find, and and accordingly, accordingly, in a sense, we are so far this is some kind of. Then a fantasy, but with this is this fantasy. Uh, as i said, yes, in some cases, fantasy is simply an interpolation between experiences. And its not interesting. That is, in the sense that a Neural Network can do this, but in some cases, this is not entirely true. Uh, uh, it doesnt quite come from experience. That is some kind of generalization. E of your knowledge and this knowledge in some way. Uh, they go beyond. Heres what uh in order to rather not that rather that the Neural Network in front of this video. Here , accordingly, it is easy for a person to say here is a blue cat. Well, thats how it happened. Code blue and the man says, huh . Well, of course, yes, this is a cat with such blue and all that. Here. And a Neural Network, for example, if it has learned a representation in which there are no blue cats, and it will be quite difficult, that is , a person, as if a little more quickly than us , will adjust how good it is. E hmm and how quickly this will be produced by the Neural Network. I dont know, well, most likely, if we Pay Attention to this and hide in life, we will probably be able to, they will catch up with us. We live we live in an amazing time when we have a competitor who is not human, but wants to catch up with us on the scale of uh, human human measurements of human values of our cognitive capabilities. We are closely monitoring what is happening. I thank you for this interesting story. Thank you very much. Best wishes. Goodbye. Thanks a lot. We watch Educational Programs and documentaries to explore the world. The west manipulates University Ratings and lures talented students from asia and africa, head of the Foreign Intelligence ServiceSergei Naryshkin stated this at the Eastern Economic forum. According to him, the west is thereby preventing the development of the educational systems of the countries of the east and the south. What was also discussed at the conference was colonialism in the east and its influence on the modern world. Watch it now