Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs

[0] The following is a conversation with Jürgen Schmidt -Huber.

[1] He's the co -director of a C .S. Swiss AI Lab and a co -creator of long -shirt -term memory networks.

[2] LSDMs are used in billions of devices today for speech recognition, translation, and much more.

[3] Over 30 years, he has proposed a lot of interesting, out -of -the -box ideas on meta -learning, adversarial networks, computer vision, and even a formal theory of quote, creativity, curiosity, curiosity, and fun.

[4] This conversation is part of the MIT course in Artificial General Intelligence and the Artificial Intelligence podcast.

[5] If you enjoy it, subscribe on YouTube, iTunes, or simply connect with me on Twitter at Lex Friedman, spelled F .R .I .D. And now, here's my conversation with Jirgen Schmidt -Huber.

[6] Early on, you dreamed of AI systems that self -improve recursively.

[7] when was that dream born when it was a baby no that's not true when I was a teenager and what was the catalyst for that birth what was the thing that first inspired you when I was a boy I was thinking about what to do in my life and then I thought the most exciting thing is to solve the riddles of the universe and that means you have to become a physicist However, then I realized that there's something even grammar.

[8] You can try to build a machine that isn't really a machine any longer that learns to become a much better physicist than I could ever hope to be.

[9] And that's how I thought maybe I can multiply my tiny little bit of creativity into infinity.

[10] But ultimately that creativity will be multiplied to understand the universe around us.

[11] that's the curiosity for that mystery that drove you yes so if you can build a machine that learns to solve more and more complex problems and more and more general problem solver then you basically have solved all the problems at least all the solvable problems so how do you think what is the mechanism for that kind of general solver look like.

[12] Obviously, we don't quite yet have one or know how to build one but have ideas and you have had throughout your career several ideas about it.

[13] So how do you think about that mechanism?

[14] So in the 80s, I thought about how to build this machine that learns to solve all these problems that I cannot solve myself.

[15] And I thought it is clear.

[16] It has to be a machine that not only learns to solve this problem here and this problem here, but it also has to learn to improve the learning algorithm itself.

[17] So it has to have the learning algorithm in a representation that allows it to inspect it and modify it such that it can come up with a better learning algorithm.

[18] So I call that meta -learning, learning to learn, and recursive.

[19] self -improvement, that is really the pinnacle of that, where you then not only learn how to improve on that problem and on that, but you also improve the way the machine improves, and you also improve the way it improves the way it improves itself.

[20] And that was my 1987 diploma thesis, which was all about that hierarchy of meta -learners that have no computational limits except for the well -known limits that Goodle identified in 1931 and for the limits are physics.

[21] In the recent years, meta -learning has gained popularity in a specific kind of form.

[22] You've talked about how that's not really meta -learning with neural networks.

[23] That's more basic transfer learning.

[24] Can you talk about the difference between the big general meta learning and a more narrow sense of meta learning the way it's used today, the way it's talked about today?

[25] Let's take the example of a deep neural network that has learned to classify images.

[26] And maybe you have trained that network on 100 different databases of images.

[27] And now a new database comes along and you want to quickly learn the new thing.

[28] as well.

[29] So one simple way of doing that is you take the network, which already knows 100 types of databases, and then you just take the top layer of that, and you retrain that using the new label data that you have in the new image database.

[30] And then it turns out that it really, really quickly can learn that to one shot, basically, because from the first 100 data sets, it already has learned so much about computer vision that it can reuse that and that is then almost good enough to solve the new task except you need a little bit of adjustment on the top.

[31] So that is transfer learning and it has been done in principle for many decades.

[32] People have done similar things for decades.

[33] Meta -learning, true.

[34] meta -learning is about having the learning algorithm itself open to introspection by the system that is using it and also open to modification such that the learning system has an opportunity to modify any part of the learning algorithm and then evaluate the consequences of that modification and then learn from that to create a better learning algorithm and so on recursively.

[35] So that's a very different animal where you are opening the space of possible learning algorithms to the learning system itself.

[36] Right.

[37] So you've, like in the 2004 paper, you describe gaitle machines, programs that rewrite themselves.

[38] Yeah.

[39] Right.

[40] Philosophically, and even in your paper, mathematically, these are really compelling ideas.

[41] But practically, do you see these self -referential programs being successful in the near term to having an impact where sort of demonstrates to the world that this direction is a good one to pursue in the near term?

[42] Yes.

[43] We had these two different types of fundamental research, how to be.

[44] build a universal problem solver, one basically exploiting proof search and things like that that you need to come up with asymptotically optimal, theoretically optimal self -improvers and problem solvers.

[45] However, one has to admit that through this proof search comes in an additive constant, an overhead, an additive overhead that vanishes in comparison to what you have to do to solve large problems.

[46] However, for many of the small problems that we want to solve in our everyday life, we cannot ignore this constant overhead.

[47] And that's why we also have been doing other things, non -universal things such as recurrent neural networks which are trained by gradient descent and local search techniques, which aren't universal at all, which aren't provably optimal at all, like the other stuff that we did, but which are much more practical as long as we only want to solve the small problems that we are typically trying to solve in this environment here.

[48] So the universal problem solvers, like the Girdle Machine, but also Marcus Huttah's father way of solving all possible problems, which he developed around 2002 in my lab, they are associated with these constant overheads for proof search, which guarantees that the thing that you're doing is optimal.

[49] For example, there is this fastest way of solving all problems with a computable solution, which is due to Marcus, Marcus Huta.

[50] And to explain what's going on there, let's take traveling salesmen problems.

[51] With traveling salesman problems, you have a number of cities, N cities, and you try to find the shortest path through all these cities without visiting any city twice.

[52] And nobody knows the fastest way of solving traveling salesman problems, TSP's, But let's assume there is a method of solving them within n to the five operations where n is the number of cities.

[53] Then the universal method of marquis is going to solve the same traveling salesman problem also within end to the five steps, plus O of one, plus a constant number of steps that you need.

[54] for the proof searcher, which you need to show that this particular class of problems, the traveling salesman problems, can be solved within a certain time bound, within order into the five steps, basically.

[55] And this additive constant doesn't care for n, which means as n is getting larger and larger as you have more and more cities, the constant overhead pales in comparison.

[56] And that means that almost all large problems are solved.

[57] In the best possible way, already today, we already have a universal problem solving like that.

[58] However, it's not practical because the overhead, the constant overhead, is so large that for the small kind of of problems that we want to solve in this little biosphere.

[59] By the way, when you say small, you're talking about things that fall within the constraints of our computational systems.

[60] So they can seem quite large to us mere humans, right?

[61] That's right, yeah.

[62] So they seem large and even unsolvable in a practical sense today, but they are still small compared to almost all problems, because almost all problems are large problems, which are much larger than any constant.

[63] Do you find it useful as a person who is dreamed of creating a general learning system, has worked on creating one, has done a lot of interesting ideas there, to think about P versus NP, this formalization of how hard problems are, how they scale, this kind of worst -case analysis type of thinking?

[64] Do you find that useful?

[65] or is it only just a mathematical, it's a set of mathematical techniques to give you intuition about what's good and bad?

[66] So P versus NP, that's super interesting from a theoretical point of view.

[67] And in fact, as you are thinking about that problem, you can also get inspiration for better practical problems all of us.

[68] On the other hand, we have to admit that at the moment, the best practical problem solvers for all kinds of problems that we are now solving through what is called AI at the moment, they are not of the kind that is inspired by these questions.

[69] There we are using general purpose computers such as recurrent neural networks, but we have a search technique, which is just local search, gradient descent, to try to find a program that is running on these recurrent networks such that it can solve some interesting problems such as speech recognition or machine translation and something like that.

[70] And there is very little theory behind the best solutions that we have at the moment that can do that.

[71] Do you think that needs to change?

[72] You think that will change?

[73] Or can we go, can we create a general intelligence system without ever really proving that that system is intelligent in some kind of mathematical way?

[74] solving machine translation perfectly or something like that within some kind of syntactic definition of a language or can we just be super impressed by the thing working extremely well and that's sufficient there's an old saying and I don't know who brought it up first which says there's nothing more practical than a good theory and a good theory of problem solving under limited resources, like here in this universe or on this little planet, has to take into account these limited resources.

[75] And so probably there is locking a theory which is related to what we already have, these asymptocally optimal problem solvers, which tells us what we need in addition to that to come up with a practically optimal problem solver.

[76] So, I believe we will have something like that, and maybe just a few little tiny twists are necessary to change what we already have, to come up with that as well.

[77] As long as we don't have that, we admit that we are taking suboptimal ways and recurrential networks and long, short term memory, equipped with local search techniques, and we are happy that it works better than any competing method, but that doesn't mean that we think we are done.

[78] You've said that an AGI system will ultimately be a simple one, a general intelligence system will ultimately be a simple one, maybe a pseudocode of a few lines to be able to describe it.

[79] Can you talk through your intuition behind this idea, why you feel that at its core intelligence is a simple algorithm?

[80] Experience tells us that the stuff that works best is really simple.

[81] So the asymptotically optimal ways of solving problems, if you look at them, just a few lines of code.

[82] It's really true.

[83] Although they are these amazing properties, just a few lines of code, then the most promising and most useful practical things maybe don't have this proof of optimality associated with them however they are also just a few lines of code the most successful recurrential networks you can write them down and five lines are pseudocode that's a beautiful almost poetic idea but what you're describing that there is the lines of pseudocode are sitting on top of layers and layers of abstractions in a sense.

[84] So you're saying at the very top, it'll be a beautifully written sort of algorithm.

[85] But do you think that there's many layers of abstractions we have to first learn to construct?

[86] Yeah.

[87] Of course, we are building on all these great abstractions that people have invented over the millennia.

[88] such as matrix multiplications and drill numbers and basic arithmetic and calculus and derivations of error functions and derivatives of error functions and stuff like that.

[89] So without that language that greatly simplifies our way of thinking about these problems, we couldn't do anything.

[90] So in that sense, as always, we are standing on the shoulders of the giants who, in the past, simplified the problem of problem -solving so much that now we have a chance to do the final step.

[91] So the final step will be a simple one.

[92] If we take a step back through all of human civilization and just the universe in general, how do you think about evolution?

[93] and what if creating a universe is required to achieve this final step?

[94] What if going through the very painful and inefficient process of evolution is needed to come up with this set of abstractions that ultimately lead to intelligence?

[95] Do you think there's a shortcut or do you think we have to create something like our universe in order to create something like human level intelligence?

[96] so far the only example we have is this one this universe in which we can do better maybe not but we are part of this whole process so apparently so it might be the case that the code that runs the universe is really really simple everything points to that possibility because gravity and other basic forces are really simple laws that can be easily described also in just a few lines of code, basically.

[97] And then there are these other events, the apparently random events in the history of the universe, which as far as we know at the moment don't have a compact code, but who knows, maybe somebody in the near future is going to figure out the pseudo -random generator, which is computing whether the measurement of that spin up or down thing here is going to be positive or negative.

[98] Underline quantum mechanics.

[99] Yes.

[100] So you ultimately think quantum mechanics is a pseudo -random number.

[101] So it's all deterministic.

[102] There's no randomness in our universe.

[103] Does God play dice?

[104] So a couple of years ago, a famous physicist.

[105] quantum physicist, Anton Zeilinger, he wrote an essay in nature, and it started more or less like that.

[106] One of the fundamental insights of the 20th century was that the universe is fundamentally random on the quantum level.

[107] And that whenever you measure spin up or down or something like that.

[108] A new bit of information enters the history of the universe.

[109] And while I was reading that, I was already typing the response and they had to publish it because I was right.

[110] That there is no evidence, no physical evidence for that.

[111] So there's an alternative explanation where everything that we consider random is actually pseudo -random, such as the decimal expansion of pi, 3 .141 and so on, which looks random, but isn't.

[112] So pi is interesting because every three -digit sequence, every sequence of three digits appears roughly one in a thousand times, and every five -digit sequence appears roughly one in 10 ,000 times.

[113] What do you would expect if it was random?

[114] But there's a very short algorithm, a short program that computes all of that.

[115] So it's extremely compressible.

[116] And who knows, maybe tomorrow somebody, some grad student at CERN, goes back over all these data points, better decay and whatever, and figures out, oh, it's the second billion digits of pie or something like that.

[117] We don't have any fundamental reason at the moment to believe that this is truly random and not just.

[118] just a deterministic video game.

[119] If it was a deterministic video game, it would be much more beautiful because beauty is simplicity.

[120] And many of the basic laws of the universe, like gravity and the other basic forces, are very simple.

[121] So very short programs can explain what these are doing.

[122] And it would be awful.

[123] And ugly, the universe would be ugly.

[124] The history of the universe would be ugly.

[125] If for the extra things, the seemingly random data points that we get all the time, that we really need a huge number of extra bits to describe all these extra bits of information.

[126] So as long as we don't have evidence that there is no short program that computes the entire history of the entire universe, we are as scientists compelled to look further for that shortest program.

[127] Your intuition says there exists a shortest program that can backtrack to the creation of the universe.

[128] So compute the shortest path to the creation of the universe.

[129] Yes, including all the entanglement things and all the spin -up and down measurements that have been taken place since 13 .8 billion years ago and so yeah so we don't have a proof that it is random we don't have a proof that it is compressible to a short program but as long as we don't have that proof we are obliged as scientists to keep looking for that simple explanation absolutely so you said simplicity is beautiful or beauty is simple either one works But you also work on curiosity, discovery, you know, the romantic notion of randomness, of serendipity, of being surprised by things that are about you, kind of in our poetic notion of reality, we think as humans require randomness.

[130] So you don't find randomness beautiful.

[131] You find simple determinism beautiful.

[132] Yeah.

[133] Okay.

[134] So why?

[135] Why?

[136] Because the explanation becomes shorter.

[137] A universe that is compressible to a short program is much more elegant and much more beautiful than another one, which needs an almost infinite number of bits to be described.

[138] As far as we know, many things that are happening in this universe are really simple in terms of short programs that compute gravity and the interaction between elementary particles and so on.

[139] So all of that seems to be very, very simple.

[140] Every electron seems to reuse the same sub -program all the time as it is interacting with other elementary particles.

[141] if we now required an extra oracle injecting new bits of information all the time for these extra things which are currently not understood such as better decay then the whole description length of the data that we can observe of the history of the universe would become much longer and therefore uglier and uglier again simplicity is elegant and beautiful or the history of science is a history of compression progress yeah so you've described sort of as we build up abstractions and you've talked about the idea of compression how do you see this the history of science the history of humanity, our civilization, and life on Earth, as some kind of path towards greater and greater compression.

[142] What do you mean by that?

[143] How do you think about that?

[144] Indeed, the history of science is a history of compression progress.

[145] What does that mean?

[146] Hundreds of years ago, there was an astronomer whose name was Kepler, and he looked at the data points that he got by watching planets, move, And then he had all these data points and suddenly it turned out that he can greatly compress the data by predicting it through an ellipse law.

[147] So it turns out that all these data points are more or less on ellipses around the sun.

[148] And another guy came along whose name was Newton and before him hook.

[149] And they said the same thing that is making these planets move like that is what makes the apples fall down.

[150] And it also holds for stones and for all kinds of other objects.

[151] And suddenly many, many of these observations became much more compressible.

[152] Because as long as you can predict, the next thing, given what you're, have seen so far, you can compress it, but you don't have to store that data extra.

[153] This is called predictive coding.

[154] And then there was still something wrong with that theory of the universe, and you had deviations from these predictions of the theory.

[155] And 300 years later, another guy came along, whose name was Einstein.

[156] And he was able to explain away all these deviations from the predictions of the old theory, through a new theory which was called the general theory of relativity which at first glance looks a little bit more complicated and you have to warp space and time but you can't phrase it within one single sentence which is no matter how fast you accelerate and how fast or hard you decelerate and no matter what is the gravity in your local framework light speed always looks the same.

[157] And from that you can calculate all the consequences.

[158] So it's a very simple thing, and it allows you to further compress all the observations, because certainly there are hardly any deviations any longer that you can measure from the predictions of this new theory.

[159] So, all of science is a history of compression progress.

[160] You never arrive immediately at the shortest explanation of the data, but you're making progress.

[161] Whenever you are making progress, you have an insight.

[162] You see, oh, first I needed so many bits of information to describe the data, to describe my falling apples, my video of falling apples, I need so many data, so many pixels have to be stored, but then suddenly I realize, no, there is a very simple way of predicting the third frame in the video from the first two.

[163] And maybe not every little detail can be predicted, but more or less most of these orange blabs that are coming down, they accelerate in the same way, which means that I can greatly compress the video.

[164] And the amount of compression, progress, that is the depth of the insight that you have at that moment.

[165] That's the fun that you have, the scientific fun, the fun in that discovery.

[166] and we can build artificial systems that do the same thing.

[167] They measure the depth of their insights as they are looking at the data, which is coming in through their own experiments, and we give them a reward, an intrinsic reward, in proportion to this depth of insight.

[168] And since they are trying to maximize the rewards they get, they are suddenly motivated to come up with new action sequences with new experiments that have the property that the data that is coming in as a consequence of these experiments has the property that they can learn something about, see a pattern in there which they hadn't seen yet before.

[169] So there's an idea of power play that you've described a training, a general problem solver in this kind of way of looking for the unsolved problems.

[170] Yeah.

[171] Can you describe that idea a little further?

[172] It's another very simple idea.

[173] So normally what you do in computer science, you have some guy who gives you a problem and then there is a huge search space of potential solution candidates and you somehow try them out and you have more less sophisticated ways of moving around in that search space until you finally found a solution which you consider satisfactory.

[174] That's what most of computer science is about.

[175] PowerPlay just goes one little step further and says, let's not only search for solutions to a given problem, but let's search to pairs of problems and their solutions, where the system itself has the opportunity to phrase its own problem.

[176] So we are looking suddenly at pairs of problems and their solutions or modifications of the problem solver that is supposed to generate a solution to that new problem.

[177] And this additional degree of freedom allows us to build curious systems that are like scientists in the sense that they not only try to solve, try to find answers, to existing questions, no, they are also free to pose their own questions.

[178] So if you want to build an artificial scientist, we have to give it that freedom, and power play is exactly doing that.

[179] So that's a dimension of freedom that's important to have, but how hard do you think that, how multi -dimensional and difficult the space of then coming up with your own questions is?

[180] Yeah.

[181] So it's one of the things that as human beings, we consider to be the thing that makes us special the intelligence that makes a special is that brilliant insight yeah that can create something totally new yes so now let's look at the extreme case let's look at the set of all possible problems that you can formally describe which is infinite which should be the next problem that a scientist or power play is going to to solve?

[182] Well, it should be the easiest problem that goes beyond what you already know.

[183] So it should be the simplest problem that the current problem solver that you have, which can already solve 100 problems, that he cannot solve yet by just generalizing.

[184] So it has to be new.

[185] So it has to require a modification of the problem solver, such that the new problem solver can solve this new thing, but the old problem solver cannot do it.

[186] And in addition to that, we have to make sure that the problem solver doesn't forget any of the previous solutions.

[187] Right.

[188] And so, by definition, power play is now trying always to search in this pair of, in the set of pairs of problems and problem solver modifications for a combination that minimize the time to achieve these criteria.

[189] So it's always trying to find the problem, which is easiest to add to the repertoire.

[190] So just like grad students and academics and researchers can spend their whole career in a local minima, stuck trying to come up with interesting questions, but ultimately doing very little.

[191] Do you think it's easy in this approach of looking for the simplest unsolvable problem to get stuck in a local minima?

[192] is never really discovering new, you know, really jumping outside of the hundred problems that you've already solved in a genuine creative way?

[193] No, because that's the nature of power play, that it's always trying to break its current generalization abilities by coming up with a new problem, which is beyond the current horizon, just shifting the horizon of knowledge a little bit out there, breaking the existing rules such that the new thing becomes solvable but wasn't solvable by the old thing.

[194] So like adding a new axiom, like what Gouldel did when he came up with these new sentences, new theorems that didn't have a proof in the formal system, which means you can add them to the repertoire, hoping that they are not going to damage the consistency of the whole thing.

[195] So in the paper with the amazing title, formal theory of creativity, fun, and intrinsic motivation, you talk about discovery as intrinsic reward.

[196] So if you view humans as intelligent agents, what do you think is the purpose and meaning of life for us humans?

[197] You've talked about this discovery.

[198] Do you see humans as an instance of power play agents?

[199] Yeah, so humans are curious, and that means they behave like scientists, not only the official scientists, but even the babies behave like scientists, and they play around with their toys to figure out how the world works and how it is responding to their actions, and that's how they learn about gravity and everything.

[200] and yeah, in 1990 we had the first systems like that would just try to play around with the environment and come up with situations that go beyond what they knew at that time and then get a reward for creating these situations and then becoming more general problem solvers and being able to understand more of the world.

[201] So, yeah, I think in principle that curiosity strategy or more sophisticated versions of what I just described, they are what we have built in as well, because evolution discovered, that's a good way of exploring the unknown world.

[202] And a guy who explores the unknown world has a higher chance of solving problems that he needs to survive in this world.

[203] On the other hand, those guys who were too curious, they were weeded out as well.

[204] So you have to find this trade -off.

[205] Evolution found a certain trade -off.

[206] Apparently in our society, there is a certain percentage of extremely explorative guys.

[207] And it doesn't matter if they die because many of the others are more conservative.

[208] And so, yeah, it would be surprising to me if that principle of artificial curiosity wouldn't be present in almost exactly the same form here in our brains so you're a bit of a musician and an artist so continuing on this topic of creativity what do you think is the role of creativity and intelligence so you've kind of implied that it's essential for intelligence if you think of intelligence as a problem solving system as ability to solve problems but do you think it's essential this idea of creativity We never have a program, a sub -program that is called creativity or something.

[209] It's just a side effect of what our problem solvers do.

[210] They are searching a space of problems or a space of candidates, of solution candidates, until they hopefully find a solution to a given problem.

[211] But then there are these two types of creativity, and both of them are now present in our machines.

[212] The first one has been around for a long time, which is human, gives problem to machine, machine tries to find a solution to that, and this has been happening for many decades, and for many decades, machines have found creative solutions to interesting problems, where humans were not aware of these particularly creative solutions, but then appreciated that the machine found that.

[213] The second is the pure creativity.

[214] That I would call, what I just mentioned, I would call the applied creativity.

[215] like applied art where somebody tells you now make a nice picture of this pope and you will get money for that okay so here is the artist and he makes a convincing picture of the Pope and the Pope likes it and gives him the money and then there is the pure creativity which is more like the power play and the artificial curiosity thing where you have the freedom to select your own problem like a scientist who defines his own question to study.

[216] And so that is the pure creativity, if you will, as opposed to the applied creativity, which serves another.

[217] And in that distinction, there's almost echoes of narrow AI versus general AI.

[218] So this kind of constrained painting of a pope seems like the approaches of what people, people are calling narrow AI and pure creativity seems to be maybe I'm just biased as a human but it seems to be an essential element of human level intelligence is that what you're implying to a degree if you zoom back a little bit and you just look at a general problem solving machine which is trying to solve arbitrary problems then this machine will figure out in the course of solving problems that it's good to be curious.

[219] So all of what I said just now about this pre -wired curiosity and this will to invent new problems that the system doesn't know how to solve yet should be just a byproduct of the general search.

[220] However, apparently evolution has built it into us because it turned out to be so successful, a pre -wiring, a bias, a very successful exploratory.

[221] are a bias that we are born with.

[222] And you've also said that consciousness in the same kind of way may be a byproduct of problem solving.

[223] Do you think, do you find it's an interesting byproduct?

[224] Do you think it's a useful byproduct?

[225] What are your thoughts on consciousness in general?

[226] Or is it simply a byproduct of greater and greater capabilities of problem solving that's similar to creativity in that sense.

[227] Yeah, we never have a procedure called consciousness in our machines.

[228] However, we get as side effects of what these machines are doing, things that seem to be closely related to what people call consciousness.

[229] So, for example, already in 1990, we had simple systems, which were basically recurrent networks and therefore universal computers, trying to map incoming data.

[230] into actions that lead to success.

[231] Maximizing reward in a given environment, always finding the charging station in time, whenever the battery is low and negative signals are coming from the battery, always find the charging station in time without bumping against painful obstacles on the way.

[232] So complicated things, but very easily motivated.

[233] And then we give these little guys a separate reconnal network which is just predicting what's happening if I do that and that what will happen as a consequence of these actions that I'm executing and it's just trained on the long and long history of interactions with the world so it becomes a predictive model of the world basically and therefore also a compressor of the observations of the world because whatever you can predict you don't have to store extra Compression is a side effect, our prediction.

[234] And how does this recurrent network compress?

[235] Well, it's inventing little sub -programs, little sub -networks that stand for everything that frequently appears in the environment, like bottles and microphones and faces, maybe lots of faces in my environment.

[236] So I'm learning to create something like a prototype face, and a new face comes along, and all I have to encode are the deviations from the prototype.

[237] So it's compressing all the time with stuff that frequently appears.

[238] There's one thing that appears all the time that is present all the time when the agent is interacting with its environment, which is the agent itself.

[239] So just for data compression reasons, it is extremely natural for this recur network to come up with little subnetworks that stand for the properties of the agents, the hand.

[240] the other actuators and all the stuff that you need to better encode the data which is influenced by the actions of the agent.

[241] So there, just as a side effect of data compression during problem solving, you have internal self -models.

[242] Now you can use this model of the world to plan your future.

[243] And that's what you also have done since 1990.

[244] So the recurrent network, which is the controller, which is trying to maximize reward, can use this model of the world, this model network of the world, this predictive model of the world, to plan ahead and say, let's not do this action sequence, let's do this action sequence instead because it leads to more predicted reward.

[245] And whenever it's waking up these little subnetworks that stand for itself, then it's thinking about itself, then it's thinking about itself.

[246] and it's exploring mentally the consequences of its own actions and now you tell me why it's still missing missing the next the gap to consciousness yeah there isn't that's a really beautiful idea that you know if life is a collection of data and in life is a process of compressing that data to act efficiently in that data, you yourself appear very often.

[247] So it's useful to form compressions of yourself.

[248] And it's a really beautiful formulation of what consciousness is, is a necessary side effect.

[249] It's actually quite compelling to me. You've described RNN's developed LSTMs, long short -term memory networks that are their type of recurrent neural networks.

[250] They have gotten a lot of success recently.

[251] So these are networks that model the temporal aspects in the data, temporal patterns in the data.

[252] And you've called them the deepest of the neural networks, right?

[253] So what do you think is the value of depth in the models that we use to learn?

[254] Since you mentioned the long, short term memory and the LSTM, I have to mention the names of the brilliant, students who made that possible.

[255] First of all, my first student ever, Sepp Hochreiter, who had fundamental insights already in his diploma thesis, then Felix Gers, who had additional important contributions, Alex Gray, is a guy from Scotland, who is mostly responsible for this CTC algorithm, which is now often used to train the LSTM to do the speech recognition on all the Google Android phones and whatever and Siri and so on.

[256] So these guys, without these guys, I would be nothing.

[257] It's a lot of incredible work.

[258] What is now the depth?

[259] What is the importance of depth?

[260] Well, most problems in the real world are deep in the sense that the current input doesn't tell you all you need to know about the environment.

[261] So instead, you have to have a memory of what happened in the past.

[262] and often, important parts of that memory are dated.

[263] They are pretty old.

[264] So when you're doing speech recognition, for example, and somebody says 11, then that's about half a second or something like that, which means it's already 50 time steps.

[265] And another guy, or the same guy, says seven.

[266] So the ending is the same, Evan.

[267] But now the system has to see the distinction between 7 and 11.

[268] And the only way it can see the difference is it has to store that 50 steps ago, there was an S or an L, 11 or 7.

[269] So there you have already a problem of depth 50, because for each time step, you have something like a virtual, a layer in the expanded, unrolled version of this recurrent network, which is doing the speech recognition.

[270] So these long time lags, they translate into problem depth.

[271] And most problems in this world are such that you really have to look far back in time to understand what is the problem and to solve it.

[272] But just like with LSCMs, you don't necessarily need to, when you look back in time, remember every aspect.

[273] You just need to remember the important aspects.

[274] That's right.

[275] The network has to learn to put the important stuff into memory and to ignore the unimportant noise.

[276] But in that sense, deeper and deeper is better, or is there a limitation?

[277] I mean, LSTM is one of the great examples of architectures that do something beyond just deeper and deeper networks.

[278] there's clever mechanisms for filtering data for remembering and forgetting so do you think that kind of thinking is necessary if you think about LSTMs as a leap, a big leap forward over traditional vanilla RNNs what do you think is the next leap within this context so LSCM is a very clever improvement but LCM still don't have the same kind of ability to see far back in the future in the past as us humans do the credit assignment problem across way back not just 50 times steps or 100 or a thousand but millions and billions it's not clear what are the practical limits of the LSTM when it comes to looking back already in 2006 I think we had examples where it not only looked back tens of thousands of steps but really millions of steps and we Juan Perez Ortiz in my lab, I think, was the first author of a paper where we really, was it 2006 or something, had examples where I learned to look back for more than 10 million steps.

[279] So, for most problems of speech recognition, it's not necessary to look that far back, but there are examples where it does.

[280] Now, the looking back thing, that's rather easy because there is only one past, but there are many possible futures.

[281] And so a reinforcement learning system, which is trying to maximize its future expected reward and doesn't know yet which of these many possible futures should I select, given this one single past, is facing problems that the LSCM by itself cannot solve.

[282] So the LSDM is good for coming up with a compact representation of the history so far, of the history and of observations and action so far.

[283] But now, how do you plan in an efficient and good way among all these, how do you select one of these many possible action sequences that a reinforcement learning system has to consider to maximize reward in this unknown future?

[284] So again, we have this basic setup where you have one every car network, which gets in the video and the speech and whatever, and it's executing actions, and it's trying to maximize a reward.

[285] So there is no teacher who tells it what to do at which point in time.

[286] And then there's the other network, which is just predicting what's going to happen if I do that and then.

[287] And that could be an LSCM network.

[288] and it learns to look back all the way to make better predictions of the next time step.

[289] So essentially, although it's predicting only the next time step, it is motivated to learn to put into memory something that happened maybe a million steps ago because it's important to memorize that if you want to predict that at the next time step, the next event.

[290] Now, how can a model of the world like that, a predictive model, of the world be used by the first guy, let's call it the controller and the model, the controller and the model.

[291] How can the model be used by the controller to efficiently select among these many possible futures?

[292] The naive way we had about 30 years ago was let's just use the model of the world as a stand -in, as a simulation of the world, and millisecond by millisecond we plan the future and that means we have to roll it out really in detail and it will work only as the model is really good and it will still be inefficient because we have to look at all these possible futures and and there are so many of them.

[293] So instead what we do now since 2015 in our CM systems, controller model systems, we give the controller the opportunity to learn by itself how to use the potentially relevant parts of the the M of the model network to solve new problems more quickly.

[294] And if it wants to, it can learn to ignore the M. And sometimes it's a good idea to ignore the M because it's really bad.

[295] It's a bad predictor in this particular situation of life where the control is currently trying to maximize reward.

[296] However, it can also learn to address and exploit some of the sub -programs that came about in the model network through compressing the data by predicting it.

[297] So it now has an opportunity to reuse that code, the algorithmic information in the model network, to reduce its own search space, such that it can solve a new problem more quickly than without the model.

[298] Compression.

[299] So you're ultimately optimistic and excited about the, power of ARL of reinforcement learning in the context of real systems.

[300] Absolutely, yeah.

[301] So you see RL as a potential having a huge impact beyond just sort of the M part is often developed on supervised learning methods.

[302] You see RL as a for problems of cell driving cars or any kind of applied side of robotics.

[303] that's the correct, interesting direction for research in your view?

[304] I do think so.

[305] We have a company called Naysense, which has applied reinforcement learning to little audies.

[306] Little Audies.

[307] Which learn to park without a teacher.

[308] The same principles were used, of course.

[309] So these little Audies, they are small, maybe like that, so much smaller than the real Audies.

[310] but they have all the sensors that you find in the real Audis, you find the cameras, the LIDA sensors.

[311] They go up to 120 kilometers an hour if they want to.

[312] And they have pain sensors, basically, and they don't want to bump against obstacles and other Audies.

[313] And so they must learn, like little babies, to park.

[314] Take the raw vision input and translate that into actions that lead to successful packing behavior, which is a rewarding thing.

[315] And yes, they learn that.

[316] They learn.

[317] So we have examples like that, and it's only in the beginning.

[318] This is just the tip of the iceberg, and I believe the next wave of AI is going to be all about that.

[319] So at the moment, the current wave of AI is about passive pattern observation and prediction, and that's what you have on your smartphone, and what the major companies on the Pacific Rim are using to sell you ads, to do marketing.

[320] That's the current source of profit in AI, and that's only one or two percent of the world economy, which is big enough to make these companies pretty much the most valuable companies in the world, but there's a much, much bigger fraction of the economy going to be affected by the next wave, which is really about machines that shape the data through their own actions.

[321] Do you think simulation is ultimately the biggest way that those methods will be successful in the next 10, 20 years?

[322] We're not talking about 100 years from now.

[323] We're talking about sort of the near -term impact of RL.

[324] Do you think really good simulation is required or is there other techniques like imitation learning, observing other humans operating in the real world.

[325] Where do you think that this success will come from?

[326] So at the moment, we have a tendency of using physics simulations to learn behavior for machines that learn to solve problems that humans also do not know how to solve.

[327] However, this is not the future, because the future is in what little babies do.

[328] They don't use a physics engine to simulate the world.

[329] No, they learn a predictive model of the world, which maybe sometimes is wrong in many ways, but captures all kinds of important abstract, high -level predictions, which are really important to be successful.

[330] And that's what was the future 30 years ago when you started that type of research, but it's still the future, and now we know much better how to go there to move.

[331] there to move forward and to really make work systems based on that where you have a learning model of the world, a model of the world that learns to predict what's going to happen if I do that and that and then the controller uses that model to more quickly learn successful action sequences.

[332] And then of course always this curiosity thing in the beginning of the model is stupid, so the controller should be motivated to come up with experiments, with action sequences that lead to data that improve the model.

[333] Do you think improving the model, constructing an understanding of the world in this connection is that now the popular approaches have been successful or grounded in ideas of neural networks, but in the 80s with expert systems, there's symbolic AI approaches, which to us humans are more intuitive in a sense that it makes sense that you build of knowledge in this knowledge representation what kind of lessons can we draw into our current approaches in for from expert systems from symbolic AI so I became aware of all of that in the 80s and back then logic program logic programming was a huge thing was it inspiring to yourself did you find it compelling because most a lot of your work was not so much in that realm right is more in the learning systems yes and no but we did all of that.

[334] So my first publication ever actually was 1987 was the implementation of genetic algorithm of a genetic programming system in Prolog.

[335] So Prolog, that's what you learned back then, which is a logic programming language.

[336] And the Japanese, they had this huge fifth generation AI project, which was mostly about logic programming back then, although neural networks existed and well known back then, and deep learning has existed since 1965, since this guy in the Ukraine, Ivaknenko started it.

[337] But the Japanese and many other people, they focused really on this logic programming, and I was influenced to the extent that I said, okay, let's take these biologically inspired algorithms like evolution, programs.

[338] and implement that in the language which I know, which was prolog, for example, back then.

[339] And then, in many ways, this came back later because the griddle machine, for example, has a proofsearcher on board.

[340] And without that, it would not be optimal.

[341] Well, Marcus Hutter's universal algorithm for solving all well -defined problems, has a proof search on board.

[342] So that's very much logic programming.

[343] Without that, it would not be asymptotically optimum.

[344] But then, on the other hand, because we have very pragmatic guys also, we focused on recondial networks and suboptimal stuff, such as gradient -based search and program space rather than provably optimal things.

[345] So logic programming certainly has a usefulness in when you're trying to construct something provably optimal or provably good or something like that.

[346] But is it useful for practical problems?

[347] It's really useful for our theorem proving.

[348] The best theorem provers today are not neural networks.

[349] Right.

[350] No, they are logic programming systems and they are much better theorem provers than most math students in the first or second semester.

[351] But for reasoning to for playing games of go or chess, or for robots, autonomous vehicles that operate in the real world, or object manipulation.

[352] You think learning.

[353] Yeah, as long as the problems have little to do with theory improving themselves, then as long as that is not the case, you would just want to have better pattern recognition.

[354] So to build a self -trying car, you want to have better pattern recognition and pedestrian recognition and all these things and you want to you want to minimize the number of false positives which is currently slowing down self -drying cars in many ways and all of that has very little to do with logic programming.

[355] What are you most excited about in terms of directions of artificial intelligence at this moment in the next few years in your own research and in the broader community?

[356] so I think in the not so distant future we will have for the first time little robots that learn like kids and I will be able to say to the robot look here a robot we are going to assemble a smartphone let's take the slab of plastic and the screwdriver and let's screw in the screw like that no not like that like that not like that like that and I don't have a data glove or something he will see me and he will hear me and he will try to do something with his own actuators which will be really different from mine but he will understand the difference and will learn to imitate me but not in the supervised way where a teacher is giving target signals for all his muscles all the time No, by doing this high -level imitation where he first has to learn to imitate me and then to interpret these additional noises coming from my mouth as helping helpful signals to do that matter.

[357] And then it will by itself come up with faster ways and more efficient ways of doing the same thing.

[358] And finally, I stop his learning algorithm and make a million copies and sell.

[359] it.

[360] And so at the moment, this is not possible, but we already see how we are going to get there.

[361] And you can imagine, to the extent that this works economically and cheaply, it's going to change everything.

[362] Almost all our production is going to be affected by that.

[363] And a much bigger wave, a much bigger AI wave is coming than the one that we are currently witnessing, which is mostly about passive pattern recognition on your smartphone.

[364] This is about active machines that shapes data through the actions.

[365] They are executing, and they learn to do that in a good way.

[366] So many of the traditional industries are going to be affected by that.

[367] All the companies that are building machines will equip these machines with cameras and other sensors, and they are going to learn to solve all kinds of problems through interaction with humans, but also a lot on their own to improve what they already can do.

[368] And lots of old economy is going to be affected by that.

[369] And in recent years I have seen that old economy is actually waking up and realizing that this is the case.

[370] Are you optimistic about that future?

[371] Are you concerned?

[372] There's a lot of people concerned in the near.

[373] term about the transformation of the nature of work, the kind of ideas that you just suggested would have a significant impact of what kind of things could be automated.

[374] Are you optimistic about that future?

[375] Are you nervous about that future?

[376] And looking a little bit farther into the future, there's people like Gilal Musk, Stuart Russell, concerned about the existential threats of that future.

[377] So in the near term, job.

[378] loss in the long -term existential threat, are these concerns to you or are you ultimately optimistic?

[379] So let's first address the near future.

[380] We have had predictions of job losses for many decades.

[381] For example, when industrial robots came along, many people predicted lots of jobs are going to get lost.

[382] And in a sense, they were right.

[383] Because back then there were car factories and hundreds of people and these factories assembled cars.

[384] And today the same car factories have hundreds of robots and maybe three guys watching the robots.

[385] On the other hand, those countries that have lots of robots per capita, Japan, Korea, Germany, Switzerland, a couple of other countries, they have really low unemployment rates.

[386] Somehow, all kinds of new jobs were created.

[387] Back then, nobody anticipated those jobs.

[388] And decades ago, I always said it's really easy to say which jobs are going to get last, but it's really hard to predict the new ones.

[389] 30 years ago, who would have predicted all these people making money as YouTube bloggers, for example?

[390] 200 years ago, 60 % of all people used to work in agriculture.

[391] Today, maybe 1%.

[392] But still, only, I don't know, 5 % unemployment.

[393] Lots of new jobs were created.

[394] And homoludans, the playing man, is inventing new jobs all the time.

[395] Most of these jobs are not existentially necessary.

[396] for the survival of our species.

[397] There are only very few existentially necessary jobs, such as farming and building houses and warming up the houses, but less than 10 % of the population is doing that.

[398] And most of these newly invented jobs are about interacting with other people in new ways, through new media and so on, getting new types of kudos in forms of likes and whatever, and even making money through that.

[399] So, Homo Ludens, the playing man, doesn't want to be unemployed, and that's why he's inventing new jobs all the time.

[400] And he keeps considering these jobs as really important and is investing a lot of energy and hours of work into those new jobs.

[401] That's quite beautifully put.

[402] We're really nervous about the future because we can't predict what kind of new jobs would be created.

[403] but you were ultimately optimistic that we humans are so restless that we create and give meaning to newer and newer jobs, totally new, things that get likes on Facebook or whatever the social platform is.

[404] So what about long -term existential threat of AI where our whole civilization may be swallowed up by this ultra -superintelligent systems?

[405] Maybe it's not going to be swallowed up, but I'd be surprised if we were, we humans, were the last step in the evolution of the universe.

[406] You've actually had this beautiful comment somewhere that I've seen saying that artificial, quite insightful.

[407] So artificial general intelligence systems, just like us humans, will likely not want to.

[408] to interact with humans, they'll just interact amongst themselves, just like ants interact amongst themselves and only tangentially interact with humans.

[409] And it's quite an interesting idea that once we create AGI, they will lose interest in humans and have, compete for their own Facebook likes on their own social platforms.

[410] So within that quite elegant idea, how do we know, in a hypothetical, sense that there's not already intelligence systems out there.

[411] How do you think broadly of general intelligence greater than us?

[412] How would we know it's out there?

[413] How do we know it's around us?

[414] And could it already be?

[415] I'd be surprised if within the next few decades or something like that we won't have AI that are truly smart in every single way and better problem solvers in almost every single important way.

[416] And I'd be surprised if they wouldn't realize what we have realized a long time ago, which is that almost all physical resources are not here in this biosphere, but further out, the rest of the solar system gets two billion times more solar energy than our little planet.

[417] There's lots of material out there.

[418] that you can use to build robots and self -replicating robot factories and all this stuff.

[419] And they are going to do that.

[420] And they will be scientists and curious, and they will explore what they can do.

[421] And in the beginning, they will be fascinated by life and by their own origins in our civilization.

[422] They will want to understand that completely, just like people today would like to understand how life works.

[423] and also the history of our own existence and civilization, but then also in the physical laws that created all of them.

[424] So in the beginning, they will be fascinated by life.

[425] Once they understand it, they lose interest, like anybody who loses interest in things he understands.

[426] And then, as you said, the most interesting sources of information for them will be others of their own kind.

[427] So, at least in the long run, there seems to be some sort of protection through lack of interest on the other side.

[428] And now it seems also clear, as far as we understand, physics, you need matter and energy to compute and to build more robots and infrastructure and more AI civilization and E .I .E .I. Ecologies consisting of trillions of different types of AIs.

[429] And so it seems inconceivable to me that this thing is not going to expand.

[430] Some AI ecology, not controlled by one AI, but by trillions of different types of AIs competing in all kinds of quickly evolving and disappearing ecological niches in ways that we cannot fathom at the moment.

[431] But it's going to expand, limited by light speed and physics, but it's going to expand.

[432] And now we realize that the universe is still young.

[433] It's only 13 .8 billion years old.

[434] And it's going to be a thousand times older than that.

[435] So there's plenty of time to conquer the entire universe and to fill it with intelligence and senders and receivers such that AIs can travel.

[436] the way they are traveling in our labs today, which is by radio from sender to receiver.

[437] And let's call the current age of the universe one eon.

[438] One eon.

[439] Now it will take just a few eons from now, and the entire visible universe is going to be full of that stuff.

[440] And let's look ahead to a time when the universe is going to be 1 ,000 times older than it is now.

[441] They will look back and they will say, look, almost in.

[442] immediately after the Big Bang, only a few eons later, the entire universe started to become intelligent.

[443] Now, to your question, how do we see whether anything like that has already happened or is already in a more advanced stage in some other part of the universe, of the visible universe?

[444] We are trying to look out there and nothing like that has happened so far, or is that true?

[445] do you think we would recognize it how do we know it's not among us how do we know planets aren't in themselves intelligent beings how do we know uh ants seen as a collective are not much greater intelligence than our own these kinds of ideas yeah when it was a boy i was thinking about these things and i thought hmm maybe it has already happened because back then i know i knew i learned from popular physics books that the structure the large scale structure of the universe is not homogeneous and you have these clusters of galaxies and then in between there are these huge empty spaces and I thought maybe they aren't really empty.

[446] It's just that in the middle of that some AI civilization already has expanded and then has covered a bubble of a billion light years diameter and using all the energy of all the stars within that bubble for its own unfathomable purposes.

[447] And so it always happened and we just failed to interpret the signs.

[448] But then I learned that gravity by itself explains the large scale of structure of the universe and that this is not a convincing explanation.

[449] And then I thought, maybe it's the dark matter because.

[450] as far as we know today, 80 % of the measurable matter is invisible.

[451] And we know that because otherwise our galaxy or other galaxies would fall apart.

[452] They are rotating too quickly.

[453] And then the idea was maybe all of these AI civilizations that are already out there, they are just invisible because they are really efficient in using the energies at their own local systems and that's why they appear dark to us but this is also not a convincing explanation because then the question becomes why is there are there still any visible stars left in our own galaxy which also must have a lot of dark matter so that is also not a convincing thing.

[454] And today, I like to think it's quite plausible that maybe you are the first, at least in our local light cone within the few hundreds of millions of light years that we can reliably observe.

[455] Is that exciting to you?

[456] That we might be the first?

[457] And it would make us much more important.

[458] Because if we mess it up through a nuclear war, then maybe this will have an effect on the development of the entire universe.

[459] So let's not mess it up.

[460] Let's not mess it up.

[461] Yulgin, thank you so much for talking today.

[462] I really appreciate it.

[463] It's my pleasure.