Vladimir Vapnik: Statistical Learning

[0] The following is a conversation with Latimer Vapnick.

[1] He's the co -inventure of support vector machine, support vector clustering, VC theory, and many foundational ideas and statistical learning.

[2] He was born in the Soviet Union and worked at the Institute of Control Sciences in Moscow.

[3] Then, in the United States, he worked at AT &T, NEC Labs, Facebook Research, and now is a professor at Columbia University.

[4] His work has been cited over 170 ,000 times.

[5] He has some very interesting ideas about artificial intelligence and the nature of learning, especially on the limits of our current approaches and the open problems in the field.

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

[7] If you enjoy it, please subscribe on YouTube or rate it on iTunes or your podcast provider of choice, or simply connect with me on Twitter or other social networks at Lex Friedman, spelled F -R -I -D.

[8] And now here's my conversation with Vladimir Vapnik.

[9] Einstein famously said that God doesn't play dice.

[10] Yeah.

[11] You have studied the world through the eyes of statistics, so let me ask you in terms of the nature of reality, fundamental nature of reality.

[12] Does God play dice?

[13] We don't know some factors.

[14] Because we don't know some factors, which could be important, it looks like good play, dice.

[15] But we should describe it.

[16] In philosophy, they distinguish between two positions, positions of instrumentalism, where you're creating theory for prediction and position of realism.

[17] where you're trying to understand what God beat.

[18] Can you describe instrumentalism and realism a little bit?

[19] For example, if you have some mechanical laws, what is that?

[20] Is it law which true always and everywhere, or it is law which allow you to predict position of moving element?

[21] what you believe.

[22] You believe that it is God's law, that God created the world which obeyed to this physical law.

[23] Or it is just law for predictions.

[24] And which one is instrumentalism?

[25] For predictions.

[26] If you believe that this is law of God and it's always true everywhere, that means that you're a realist.

[27] You're trying to really understand the God thought.

[28] So the way you see the world is as an instrumentalist?

[29] You know, I'm working for some models, models of machine learning.

[30] So in this model, we can see setting, and we try to solve, resolve the setting, to solve the problem.

[31] And you can do in two different way, from the point of view of instrumentalists, and that's what everybody does now, because they say the goal of machine learning is to find the rule for classification.

[32] That is true, but it is instrument for prediction.

[33] But I can say the God of machine learning is to learn about conditional probability.

[34] So how God played use and is he play?

[35] What is probability for one?

[36] What is probability for another given situation?

[37] But for prediction, I don't need this.

[38] I need the role.

[39] But for understanding, I need conditional probability.

[40] So let me just step back a little bit.

[41] to talk about, you mentioned, which I read last night, the parts of the 1960 paper by Eugene Wigner, unreasonable effectiveness of mathematics and natural sciences.

[42] Such a beautiful paper, by the way.

[43] It made me feel, to be honest, to confess my own work in the past few years on deep learning, heavily applied.

[44] It made me feel that I was missing out.

[45] on some of the beauty of nature in the way that math can uncover.

[46] So let me just step away from the poetry of that for a second.

[47] How do you see the role of math in your life?

[48] Is it a tool?

[49] Is it poetry?

[50] Where does it sit?

[51] And does math for you have limits of what it can describe?

[52] Some people saying that mass is language which use God.

[53] So I believe Speak to God or use God or use God.

[54] Yeah.

[55] So I believe that this article about unreasonable effectiveness of mass is that if you're looking at mathematical structures, they know something about reality.

[56] And the most scientists from natural science, they're looking on equation and trying to understand reality.

[57] So the same in machine learning.

[58] If you try very carefully look on all equations which define conditional probability, you can understand something about reality more than.

[59] than from your fantasy.

[60] So math can reveal the simple underlying principles of reality, perhaps?

[61] You know what means simple?

[62] It is very hard to discover them.

[63] But then when you discover them and look at them, you see how beautiful they are.

[64] And it is surprising why people did not see that before.

[65] or you're looking on equation and derive it from equations.

[66] For example, I talked yesterday about least square -needed.

[67] And people had a lot of fantasy, have to improve least -square -netted.

[68] But if you're going step -by -step by solving some equations, you suddenly will get some term which, after thinking, you understand that it described position of observation.

[69] point.

[70] In this square method, we throw out a lot of information.

[71] We don't look in composition of points of observations.

[72] We're looking only on residuals.

[73] But when you understood that, that's a very simple idea.

[74] But it's not too simple to understand.

[75] And you can derive this just from equations.

[76] So some simple algebra, a few steps will take you to something surprising that when you think about, you understand.

[77] And that is proof that human intuition not to reach and very primitive and it does not see very simple situations.

[78] So let me take a step back.

[79] In general, yes, right?

[80] But what about human, as opposed to intuition, and ingenuity, the moments of brilliance.

[81] So, are you so, do you have to be so hard on human intuition?

[82] Are there moments of brilliance in human intuition?

[83] They can leap ahead of math, and then the math will catch up?

[84] I don't think so.

[85] I think that the best human intuition, it is putting in axioms.

[86] and then it is technical way.

[87] See where the axioms take you.

[88] Yeah.

[89] But if they correctly take axioms.

[90] But axiom polished during generations of scientists.

[91] And this is integral wisdom.

[92] So that's beautifully put.

[93] But if you maybe look at, when you think of Einstein and special relativity, What is the role of imagination coming first there in the moment of discovery of an idea?

[94] So there's obviously a mix of math and out -of -the -box imagination there.

[95] That's I don't know.

[96] Whatever I did, I exclude any imagination.

[97] Because whatever I saw in machine learning that come from imagination, like features, like deep learning, they are not relevant to the problem.

[98] When you're looking very carefully from mathematical equations, you're deriving very simple theory, which goes far by on.

[99] Theoretically.

[100] Then whatever people can imagine.

[101] Because it is not good fantasy.

[102] It is just interpretation, it is just fantasy, but it is not what you need.

[103] You don't need any imagination.

[104] to derive, say, main principle of machine learning.

[105] When you think about learning and intelligence, maybe thinking about the human brain and trying to describe mathematically the process of learning that is something like what happens in the human brain.

[106] Do you think we have the tools currently?

[107] Do you think we will ever have the tools to try to describe that process of learning?

[108] you it is not description what's going on it is interpretation it is your interpretation your vision can be wrong you know when guy invent microscope Levingook for a first time only he got this instrument and nobody he kept secret about microscope but he wrote reports in London Academy of Science and in his His report, when he's looking at the blood, he looked everywhere on the water, on the blood, on the spirit.

[109] But he described blood like fight between queen and king.

[110] So he saw blood cells, red cells, and he imagines that it is army fighting each other.

[111] And it was his interpretation of situation.

[112] And he sent this report in Academy of Science.

[113] they very carefully looked because they believed that he is right, he saw something.

[114] But he gave wrong interpretation.

[115] And I believe the same can happen with brain.

[116] Because the most important part, you know, I believe in human language, in some proverb is so much wisdom.

[117] For example, people say that It is better than a thousand days of diligent studies one day with a great teacher.

[118] But if I will ask you what teacher does, nobody knows.

[119] And that is intelligence.

[120] But we know from history and now from mass and machine learning.

[121] That teacher can do a lot.

[122] So what, from a mathematical point of view, is the great teacher?

[123] I don't know.

[124] No, no, but we can say what teacher can do.

[125] He can introduce some invariance, some predicate for creating invariants.

[126] Have he doing it?

[127] I don't know because teacher knows reality and can describe from this reality a predicate invariance.

[128] But we know that when you're using invariant, he can decrease number of observations hundred times that's so but maybe try to pull that apart a little bit I think you mentioned like a piano teacher saying to the student play like a butterfly yeah I played piano I played in guitar for a long time and it yeah that's there's maybe it's romantic poetic but it feels like there's a lot of truth in that statement like there's there is a lot of instruction in that statement And so can you pull that apart?

[129] What is that?

[130] The language itself may not contain this information.

[131] It is not blah, blah, blah, yeah.

[132] It's what?

[133] Effect you.

[134] Effect your playing.

[135] Yes, it does.

[136] But it's not the language.

[137] It feels like what is the information being exchanged there?

[138] What is the nature of information?

[139] What is the representation of that information?

[140] I believe that it is sort of predicate, but I don't know.

[141] That is exactly what intelligence in machine learning should be.

[142] Yes.

[143] Because the rest is just mathematical technique.

[144] I think that what was discovered recently is that there is two mechanisms of learning.

[145] One called strong convergence mechanism and weak convergence mechanism.

[146] Before people use only one conversion.

[147] In weak convergence mechanism, you can use predicate.

[148] That's what play like butterfly, and it will immediately affect your playing.

[149] You know, there is English proverb, great.

[150] If it looks like a dog, swims like a dog, and quack like a dog, then it is probably dog.

[151] Yes.

[152] But this is exact about predicate.

[153] Looks like a duck, what it means.

[154] So you saw many ducks that you're training data.

[155] So you have description of how looks, integral looks, ducks.

[156] Yeah, the visual characteristics of a duck, yeah.

[157] Yeah, but you want, and you have model for the cognition doubts.

[158] So you would like so that theoretical description from model coincide with empirical description.

[159] which you saw on teletext day.

[160] So about looks like a duck, it is general.

[161] But what about swims like a dog?

[162] You should know that dog swims.

[163] You can say it plays chess like a dog.

[164] Okay, dog doesn't play chess.

[165] And it is completely legal predicate, but it is useless.

[166] So half teacher can recognize not useless predicate.

[167] So up to now, we don't use this predicate in existing machine learning.

[168] So why we need zillions of data.

[169] But in this English proverb, they use only three predicates.

[170] It looks like a duck, swims like a duck, and quack like a duck.

[171] So you can't deny the fact that swims like a duck and quacks like a duck has humor in it, has ambiguity.

[172] Let's talk about swim like a duck.

[173] It does not say jumps, jump like a duck.

[174] Why?

[175] Because...

[176] It's not relevant.

[177] But that means that you know ducks, you know, different birds, you know, animals.

[178] And you derive from this that it is relevant to say swim like a jump.

[179] So underneath, in order for us to understand swims like a duck, it feels like we need to know millions of other little pieces of information.

[180] Which we pick up along the way.

[181] You don't think so.

[182] There doesn't need to be this knowledge base.

[183] In those statements, carries some rich information that helps us understand the essence of duck.

[184] Yeah.

[185] How far are we from integrating predicates?

[186] You know that when you consider complete theory, machine learning, so what it does, you have a lot of functions.

[187] And then you're talking, it looks like a dog.

[188] You see your training data.

[189] From training data, you recognize like expected dog should look.

[190] Then you remove all functions, which does not look like you think it should look from training data.

[191] So you decrease amount of function from which we, you pick up one, then you give a second predicate, and again, decrease the set of function.

[192] And after that, you pick up the best function you can find.

[193] It is standard machine learning.

[194] So why you need not too many examples?

[195] Because your predicates aren't very good?

[196] Yeah, that means that predicate very good.

[197] Because every predicate is invented to decrease.

[198] admissible set of functions.

[199] So you talk about admissible set of functions and you talk about good functions.

[200] So what makes a good function?

[201] So admissible set of function is set of function which has small capacity or small diversity, small VC dimension example, which contain good function and such.

[202] By the way, for people who don't know, VC, you're the V in the VC.

[203] So how would you describe to a layperson, what VC theory is, how would you describe VC?

[204] So when you have a machine, so machine capable to pick up one function from the admissible set of function.

[205] But a set of admissible function can be big.

[206] So they contain all continuous functions and it useless.

[207] You don't have so many examples to pick up function.

[208] But it can be small.

[209] Small, we call it capacity, but maybe it better call diversity.

[210] So not very different function in the set.

[211] It's infinite set of function, but not very diverse.

[212] So it is small VC dimension.

[213] When the C dimension is small, you need small amount of training data.

[214] So the goal is to create administrative.

[215] set of functions which have small VC dimension and contain good function.

[216] Then you should, then you will be able to pick up the function using small amount of observations.

[217] So that is the task of learning.

[218] Yeah.

[219] Is creating a set of admissible functions that has a small VC dimension.

[220] And then you've figured out a clever way of picking up.

[221] No, that is the goal of learning, which I formulated yesterday.

[222] Statistical learning theory does not involve in creating admissible set of function.

[223] In classical learning theory, everywhere, 100 % in textbook, the set of function, admissible set of function is given.

[224] But this is science about nothing, because the most difficult problem to create admissible set of functions.

[225] given, say, a lot of functions, continuous set of function, create admissible set of functions, that means that it has finite VC dimension, small VC dimension, and contain good function.

[226] So this was out of consideration.

[227] So what's the process of doing that?

[228] I mean, it's fascinating.

[229] What is the process of creating this admissible set of functions?

[230] That is invariance.

[231] That's invariance.

[232] Can you describe the variance?

[233] Yeah, you're looking of properties of training data, and properties means that you have some function, and you just count what is value, average value of function on training data.

[234] You have model, and what is the expectation of this function on the model, and they should coincide.

[235] So the problem is about how to pick up functions.

[236] It can be any function.

[237] In fact, it is true for all functions.

[238] But because when I'm talking to, say, duck does not jumping, so you don't ask questions, jump like a dog.

[239] because it is trivial, it does not jumping and doesn't help you to recognize you.

[240] But you know something, which question to ask.

[241] And you're asking, it swims like a dog.

[242] But looks like a duck at his general situation.

[243] It looks like, say, a guy who has this illness, is this disease.

[244] It is legal.

[245] Yeah.

[246] So there is a general type of predicate looks like, and special type of predicate which related to this specific problem.

[247] And that is intelligence part of all this business.

[248] And that where teachers involved.

[249] Incorporating the specialized predicates.

[250] Yeah.

[251] Okay.

[252] What do you think about deep learning as neural networks, these arbitrary architectures, as helping accomplish some of the tasks you're thinking about.

[253] Their effectiveness or lack thereof.

[254] What are the weaknesses and what are the possible strengths?

[255] You know, I think that this is fantasy.

[256] Everything, like deep learning, like features.

[257] Let me give you this example.

[258] One of the greatest book, this Churchill book, about history of Second World War and he's starting this book describing that in all time when war is over so the great kings they gathered together and most all of them were relatives and they discussed what should be done have to create peace and they came to agreement and that when happened First World War, the general public came in power.

[259] And they were so greedy that robbed Germany.

[260] And it was clear for everybody that it is not peace.

[261] That peace will last only 20 years because they were not professionals.

[262] And the same I see in machine learning.

[263] There are mathematicians who are looking for the problem from very deep point of view, of mathematical point.

[264] And there are computer scientists with mostly does not know mathematics.

[265] They just have interpretation of that.

[266] And they invented a lot of blah, blah, blah interpretations like deep learning.

[267] Why you did deep learning, mathematics does not know deep learning.

[268] Mathematic does not know neurons.

[269] It is just function.

[270] If you like to say, piecewise linear function, say that.

[271] and do in class of piecewise linear function but they invent something and then they try to to prove advantage of that through interpretations which mostly wrong and when it not enough they they appeal to brain which they know nothing about that nobody knows what going on in the brain so I think that more reliable look on mass this is mathematical problem, do your best to solve this problem.

[272] Try to understand that there is no only one way of convergence, which is a strong way of convergence, there is a weak way of convergence, which requires predicate.

[273] And if you will go through all this stuff, you will see that you don't need deep learning.

[274] Even more, I would say one of the theorem, which called a representative theory.

[275] It says, That optimal solution of mathematical problem, which described learning, is on shadow network, not on deep learning.

[276] And a shallow network, yeah.

[277] The ultimate problem is there.

[278] Absolutely.

[279] So in the end, what you're saying is exactly right.

[280] The question is, you have no value for throwing something on the table.

[281] playing with it, not math.

[282] It's like a neural network or you said throwing something in the bucket or the biological example and looking at kings and queens or the cells or the microscope.

[283] You don't see value in imagining the cells or kings and queens and using that as inspiration and imagination for where the math will eventually lead you.

[284] You think that interpretation basically deceives you in a way that's not productive.

[285] I think that if you're trying to analyze this business of learning, and especially discussion about deep learning, it is discussion about interpretation, not about things, about what you can say about things.

[286] That's right, but aren't you surprised by the beauty of it?

[287] So not mathematical beauty, but the fact that it works at all.

[288] Or are you criticizing that very beauty, our human desire to interpret, to find our silly interpretations in these constructs?

[289] Let me ask you this.

[290] Are you surprised and does it inspire you?

[291] How do you feel about the success of a system like AlphaGo beating the game of Go?

[292] using neural networks to estimate the quality of a board and the quality of the position.

[293] That is your interpretation, quality of the board.

[294] Yeah, yes.

[295] Yeah.

[296] But it works.

[297] So it's not our interpretation.

[298] The fact is a neural network system doesn't matter, a learning system that we don't, I think, mathematically understand that well, beats the best human player.

[299] Does something that was thought impossible.

[300] That means that it's not very difficult problem.

[301] So we empirically have discovered that this is not a very difficult problem.

[302] It's true.

[303] So maybe I can't argue.

[304] Even more, I would say, that if they use deep learning, it is not the most effective way of learning theory.

[305] And usually when people use deep learning, they're using zillions of training data.

[306] Yeah, but you don't need this.

[307] So I describe challenge.

[308] Can we do some problems which do well deep learning method with deep net using 100 times less training data?

[309] Even more, some problems deep learning cannot solve because it's not necessarily they create admissible set of function when they get.

[310] To create deep architecture means to create admissible set of functions.

[311] You cannot say that you're creating good admissible set of functions.

[312] You're just, it's your fantasy.

[313] It does not come from mass. but it is possible to create admissible set of functions because you have your training data.

[314] Actually, for mathematicians, when you consider invariant, you need to use law of large numbers.

[315] When you're making training in existing algorithm, you need uniform law of large numbers, which is much more difficult.

[316] requires.

[317] We see dimension and all this stuff.

[318] But nevertheless, if you use balls, weak and stroke way of convergence, you can decrease a lot of training data.

[319] Yeah, you could do the three, the swims like a duck and quacks like a duck.

[320] But our, so let's let's step back and think about human intelligence in general.

[321] And clearly that has evolved in a non -mathematical away.

[322] It wasn't, as far as we know, God or whoever didn't come up with a model and place in our brain of admissible functions.

[323] It kind of evolved.

[324] I don't know.

[325] Maybe you have a view on this, but so Alan Turing in the 50s in his paper asked and rejected the question, can machines think as not a very useful question.

[326] But can you briefly entertain this useful question?

[327] useless question.

[328] Can machines think?

[329] So talk about intelligence and your view of it?

[330] I don't know that.

[331] I know that Turing describe imitation.

[332] If computer can imitate human being, let's call it intelligent.

[333] And he understands that it is not thinking computer.

[334] He completely understand what he's doing but he's set up problem of imitation so now we understand that the problem not in imitation I'm not sure that intelligence just inside of us it may be also outside of us I have several observations so when I prove some theory it's very difficult theory in couple of years in several places.

[335] People proved the same theory.

[336] I'm saying, Soyri Lema, after us was done.

[337] Then another guys proved the same theory.

[338] In history of science, it's happened all the time.

[339] For example, geometry.

[340] It's happened simultaneously.

[341] First, it did Lobachevsky, and then Gauss and Bojai and another guy's.

[342] And it approximately in 10 times period, 10 years period.

[343] of time.

[344] And I saw a lot of examples like that.

[345] And many mathematicians think that when they develop something, they develop something in general which affect everybody.

[346] So maybe our model that intelligence only inside of us is incorrect.

[347] It's our interpretation.

[348] Yeah.

[349] It may be there exists some connection with world intelligence.

[350] I don't know that.

[351] You're almost like plugging in into yeah exactly and contributing to this into a big network into a big maybe in your own network no no no no on the flip side of that maybe you can comment on a big O complexity and how you see classifying algorithms by worst case running time in relation to their input so that way of thinking about functions do you think P equals NP Do you think that's an interesting question?

[352] Yeah, it is an interesting question.

[353] But let me talk about complexity and about worst -case scenario.

[354] There is a mathematical setting.

[355] When I came to United States in 1990, so people did not know.

[356] This is the theory they did not know, Statistical Learning.

[357] So in Russia, it was published to monographs, our monographs, but in America, didn't know.

[358] Then they learned and somebody told me that it is worst case theory and they will create real case theory but till now I did not because it is mathematical tool.

[359] You can do only what you can do using mathematics and which has a clear understanding and clear description and for this reason we introduce complexity and we need this because using actually it is diversity like this more DC dimension you can prove some theorems but we also create theory for case when you know probability measure and that is the best case which can happen in this entropy theory so from a mathematical point of view you know the best possible case, the worst possible case.

[360] You can derive different model in medium.

[361] But it's not so interesting.

[362] You think the edges are interesting?

[363] The edges is interesting.

[364] Because it is not so easy to get good bound, exact bound.

[365] It's not many cases where you have the bond is not exact.

[366] But interesting principles.

[367] which discover the mass. Do you think it's interesting because it's challenging and reveals interesting principles that allow you to get those bounds?

[368] Or do you think it's interesting because it's actually very useful for understanding the essence of a function, of an algorithm?

[369] So it's like me judging your life as a human being by the worst thing you did and the best thing you did versus all the stuff in the middle.

[370] it seems not productive.

[371] I don't think so because you cannot describe situation in the middle or it will be not general.

[372] So you can describe ages cases and it is clear it has some model but you cannot describe model for every new case.

[373] So you will be nervous.

[374] were accurate when you're using more.

[375] But from a statistical point of view, the way you've studied functions and the nature of learning and the world, don't you think that the real world has a very long tail that the edge cases are very far away from the mean the stuff in the middle or no?

[376] I don't know that.

[377] Because...

[378] I seem I think that, but from my point of view, if you will use formal statistic, you need uniform law of large numbers.

[379] If you will use this invariance business, you will need just law of large numbers.

[380] You don't, and there's this huge difference between uniform law of large numbers and launch numbers.

[381] Is it useful to describe that a little more?

[382] Or should we just take it to...

[383] No, for example, when I'm talking about doc, I gave three predicates and that was enough.

[384] But if you will try to do formal, distinguish, you will need a lot of observations.

[385] So that means that information about looks like a duck.

[386] contain a lot of bit of information, form of bits of information.

[387] So we don't know that, how much bit of information contain things from artificial, from intelligence, and that is the subject of analysis.

[388] Till now, all business, I don't like how people consider artificial intelligence.

[389] they consider as some codes which imitate activity of human being.

[390] It is not science, it is applications.

[391] You would like to imitate go ahead.

[392] It is very useful and good problem.

[393] But you need to learn something more.

[394] How people can to develop, say, predicate.

[395] swims like a dog or play like butterfly or something like that.

[396] Then not the teacher says you how it came in his mind, how he chose this image.

[397] So that process.

[398] That is the problem of intelligence.

[399] That is the problem of intelligence.

[400] And you see that connected to the problem of learning?

[401] Absolutely.

[402] Because you immediately give this predicate, like specific predicates, swims like a dog or quack like a dog.

[403] It was chosen somehow.

[404] So what is the line of work, would you say?

[405] If you were to formulate as a set of open problems, that will take us there.

[406] To play like a butterfly, we'll get a system to be able to...

[407] Let's separate two stories.

[408] One mathematical story, that if you have predicate, you can do something.

[409] And another story, you have to get predicate.

[410] It is an intelligence problem, and people even did not start to understand intelligence.

[411] Because to understand intelligence, first of all, try to understand what doing teachers.

[412] Half teacher teach.

[413] Why one teacher better than another one?

[414] Yeah.

[415] So you think we really even haven't started on the journey of generating the predicates?

[416] We don't understand.

[417] We even don't understand this.

[418] problem exists.

[419] Because did you care?

[420] You do.

[421] No, I just know a name.

[422] I want to understand why one teacher better than another.

[423] And how affect teacher, student.

[424] It is not because he's repeating the problem which is in textbook.

[425] He makes some remarks.

[426] He makes some philosophy of reasoning.

[427] you know, that's a beautiful.

[428] So it is a formulation of a question that is the open problem.

[429] Why is one teacher better than another?

[430] Right.

[431] But he does better.

[432] Yeah.

[433] What, what, what, what, what, what, what, what, what, why in, at every level?

[434] But people.

[435] How do they get better?

[436] What does it mean to be better?

[437] The, the whole.

[438] Yeah.

[439] Yeah, yeah.

[440] From, from, from whatever model I have.

[441] Yeah.

[442] One teacher can give a very good predicate.

[443] One teacher can say swims like a dog, and another can say jump like a duck.

[444] And jump like a dog carries zero information.

[445] Yeah.

[446] So what is the most exciting problem in statistical learning you've ever worked on or are working on now?

[447] I just finished this invariant story.

[448] And I am happy that I believe that it is ultimate learning story.

[449] At least I can show that there are no another mechanism, only two mechanisms.

[450] But they separate statistical part from intelligent part.

[451] And I know nothing about intelligent part.

[452] And if we will know this intelligent part, So it will help us a lot in teaching, in learning.

[453] In learning.

[454] Yeah.

[455] You don't know it when we see it?

[456] So for example, in my talk, the last slide was the challenge.

[457] So you have, say, NIST digital recognition problem.

[458] And deep learning claims that they did it very well, say, 99 .5 % of correct answers.

[459] But say you use 60 ,000 observations.

[460] Can you do the same using 100 times less?

[461] But incorporating invariants.

[462] What it means, you know, digit 1, 2, 3, but just looking at that, explain me which invariant I should keep to use 100 examples or say 100 times less examples to do the same job.

[463] Yeah, that last slide.

[464] Unfortunately, your talk ended quickly, but that last slide was a powerful open challenge and a formulation of the essence here.

[465] That is the exact problem of intelligence.

[466] Because everybody, when machine learning started and it was developed by mathematician, they immediately recognize that we use much more training data than human needed.

[467] But now again we came to the same story.

[468] Have to decrease.

[469] And that is the problem of learning.

[470] It is not like in deep learning they use zillions of training data.

[471] Because maybe zillions are not enough if you have a good invariance.

[472] Maybe you will never collect some number of observations.

[473] But now it is a question to intelligence.

[474] How to do that?

[475] Because statistical part is ready.

[476] As soon as you supply us with predicate, we can do a good job with small amount of observations.

[477] And the very first challenge is we'll know digit recognition.

[478] And you know digits?

[479] And please tell me invariants.

[480] I think about that I can say, for digit three, I would introduce concept of horizontal symmetry.

[481] So the digit three has horizontal symmetry, say more than say digit two or something like that.

[482] As soon as I get the idea of horizontal symmetry, I can mathematically invent a lot of measure of horizontal symmetry or then vertical symmetry or diagonal symmetry, whatever, if I have idea of symmetry.

[483] But what else?

[484] Looking on digit, I see that it is metapredicate, which is not shape.

[485] It is something like symmetry, like half dark is whole picture, something like that, which can self -rise a predicate.

[486] You think such a predicate could rise.

[487] out of something that's not general, meaning it feels like for me to be able to understand the difference between a two and three, I would need to have had a childhood of 10 to 15 years playing with kids, going to school, being yelled by parents, all of that, walking, jumping, looking at ducks, and now, then I would be able to generate the right predicate for telling the difference between two and a three.

[488] Or do you think there's a more efficient way?

[489] I know, for sure, then you must know something more than digits.

[490] Yes.

[491] And that's a powerful statement.

[492] Yeah, but maybe there are several languages of description of these elements of digits.

[493] So I'm talking about symmetry, about some properties of geometry.

[494] I'm talking about something abstract.

[495] I don't know that.

[496] But this is a problem of intelligence.

[497] So in one of our article, it is trivial to show that every example, can carry not more than one bit of information in real.

[498] Because when you show example, and you say this is one, you can remove a function which does not tell you one.

[499] It's the best strategy.

[500] If you can do it perfectly, it's remove half of the functions.

[501] But when you use one predicate, which looks like a duck, you can remove much more functions than half.

[502] And that means that it contains a lot of bit of information from formal point of view.

[503] But when you have a general picture on what you want to recognize, and general picture of the world, can you invent this predicate?

[504] and that predicates carry a lot of information.

[505] Beautifully put, maybe just me, but in all the math you show, in your work, which is some of the most profound mathematical work in the field of learning AI and just math in general, I hear a lot of poetry and philosophy.

[506] You really kind of to talk about philosophy of science.

[507] There's a, poetry and music to a lot of the work you're doing and the way you're thinking about it.

[508] So do you, where does that come from?

[509] Do you escape to poetry?

[510] Do you escape to music or not?

[511] I think that there exists ground truth.

[512] There exists ground truth.

[513] Yeah, and that can be seen everywhere.

[514] Yeah.

[515] The smart guy, philosopher, sometimes I surprise how they deep see.

[516] sometimes I see that some of them are completely out of subject but the ground truth I see in music music is the ground truth yeah and in poetry many poets they believe that they take dictation so what piece of music as a piece of a empirical evidence gave you a sense that they are touching something in the ground truth.

[517] It is structure.

[518] The structure with the math of music.

[519] Yeah, because when you're listening to Bach, you see this structure.

[520] Very clear, very classic, very simple.

[521] And the same in Maas, when you have axioms in geometry, you have the same feeling.

[522] Yeah.

[523] And in poetry sometimes you see the same.

[524] Yeah.

[525] and if you look back at your childhood, you grew up in Russia, you maybe were born as a researcher in Russia, you've developed as a researcher in Russia, you've came to the United States and in a few places.

[526] If you look back, what were, what was some of your happiest moments as a researcher?

[527] Some of the most profound moments, not in terms of their impact on society, but in terms of of their impact on how damn good you feel that day and you remember that moment.

[528] You know, every time when you found something, it is great.

[529] Yeah, one of the things in life.

[530] Every simple things.

[531] But my general feeling that most of my time was wrong.

[532] You should go again and again and again and try to be honest.

[533] in front of yourself, not to make interpretation, but try to understand that it related to grand truth.

[534] It is not my blah, blah, blah, interpretation and something like that.

[535] But you're allowed to get excited at the possibility of discovery.

[536] Oh, yeah.

[537] You'll have to double check it, but...

[538] No, but how it's related to the other ground truth?

[539] Is it just temporary or it is forever?

[540] You know, you always have a feeling when you found something, how big is that?

[541] So 20 years ago when we discovered statistical learning theory, nobody believed, except for one guy, Dudley from MIT.

[542] And then in 20 years it became.

[543] fashion and the same with support vector machines.

[544] That's kernel machines.

[545] So with support vector machines and learning theory when you were working on it you had a sense that you had a sense of the the profundity of it how this seems to be right this seems to be powerful.

[546] Right.

[547] Absolutely immediately.

[548] I recognize that it will last forever.

[549] And now when I found this invariance story.

[550] You feel the side?

[551] So I have a feeling that it is complete long.

[552] Because I have proved that there are no different mechanisms.

[553] You can have some, say, cosmetic improvement you can do.

[554] But in terms of invariance, you need both environments.

[555] variants and statistical learning and they should work together.

[556] But also I'm happy that we can formulate what is intelligence with that and to separate from technical part.

[557] That is completely better.

[558] Absolutely.

[559] Well, Vladimir, thank you so much for talking today.

[560] Thank you.

[561] That's an honor.