Vladimir Vapnik: Statistical Learning
Guests: Vladimir Vapnik
Full transcript · Read the full conversation in order, with its original context.
§01Transcript · 00:00–11:08
The following is a conversation with Vladimir Vapnik.
He's the co-inventor of the support vector machine,
support vector clustering, VC theory,
and many foundational ideas in statistical learning.
He was born in the Soviet Union
and worked at the Institute of Controlled Sciences in Moscow.
Then in the United States,
he worked at AT&T, NEC Labs, Facebook Research,
and now is a professor at Columbia University.
His work has been cited over 170,000 times.
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.
This conversation is part of MIT course
on artificial general intelligence
and the Artificial Intelligence Podcast.
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.
And now, here's my conversation with Vladimir Vapnik.
Einstein famously said that God doesn't play dice.
- Yeah.
- 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, does God play dice?
- You don't know some factors.
And because you don't know some factors,
which could be important, it looks like God plays dice.
But we only should describe.
In philosophy, they distinguish between two positions,
positions of instrumentalism,
where you're creating theory for prediction,
and position of realism,
where you're trying to understand what God did.
- Can you describe instrumentalism and realism a little bit?
- For example, if you have some mechanical laws,
what is that?
Is it law which true always and everywhere,
or it is law which allow you to predict
position of moving element?
What you believe.
You believe that it is God's law,
that God created the world,
which obey to this physical law,
or it is just law for predictions.
- And which one is instrumentalism?
- For predictions.
If you believe that this is law of God,
and it's always true everywhere,
that means that you're realist.
So you're trying to really understand God's thought.
- So the way you see the world is as an instrumentalist?
- You know, I'm working for some models,
model of machine learning.
So in this model, we can see setting,
and we try to solve,
resolve the setting to solve the problem.
And you can do it in two different way,
from the point of view of instrumentalist,
and that's what everybody does now,
because they say the goal of machine learning
is to find the rule for classification.
That is true, but it is instrument for prediction.
But I can say the goal of machine learning
is to learn about conditional probability.
So how God played use,
and is he play, what is probability for one,
what is probability for another given situation?
But for prediction, I don't need this.
I need the rule.
But for understanding, I need conditional probability.
- So let me just step back a little bit first
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."
Such a beautiful paper, by the way.
Made me feel, to be honest,
to confess my own work in the past few years
on deep learning, heavily applied,
made me feel that I was missing out
on some of the beauty of nature
in the way that math can uncover.
So let me just step away from the poetry
of that for a second.
How do you see the role of math in your life?
Is it a tool, is it poetry?
Where does it sit?
And does math, for you, have limits
of what it can describe?
- Some people say that math is language which use God.
So I believe--
- Speak to God or use God?
- Use God. - Use God.
- Yeah.
So I believe that this article
about effectiveness, unreasonable effectiveness of math,
is that if you're looking at mathematical structures,
they know something about reality.
And the most scientists from natural science,
they're looking on equation and trying to understand reality.
So the same in machine learning.
If you try, very carefully look on all equations
which define conditional probability,
you can understand something about reality
more than from your fantasy.
- So math can reveal the simple underlying principles
of reality, perhaps.
- You know what means simple?
It is very hard to discover them.
But then when you discover them and look at them,
you see how beautiful they are.
And it is surprising why people did not see that before.
You're looking on equation and derive it from equations.
For example, I talked yesterday about least square method.
And people have a lot of fantasy
how to improve least square method.
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 describes position
of observation point.
In least square method, we throw out a lot of information.
We don't look in composition of point of observations,
we're looking only on residuals.
But when you understood that, that's very simple idea,
but it's not too simple to understand.
And you can derive this just from equations.
- So some simple algebra, a few steps will take you
to something surprising that when you think about--
- Absolutely, yes.
- You understand.
- And that is proof that human intuition
not too rich and very primitive.
And it does not see very simple situations.
- So let me take a step back.
In general, yes, right?
But what about human, as opposed to intuition,
ingenuity, moments of brilliance?
Do you have to be so hard on human intuition?
Are there moments of brilliance in human intuition
that can leap ahead of math,
and then the math will catch up?
- I don't think so.
I think that the best human intuition,
it is putting in axioms.
And then it is technical way--
- See where the axioms take you.
- Yeah.
But if they correctly take axioms,
but axiom polished during generations of scientists.
And this is integral wisdom.
- So that's beautifully put.
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?
So there's obviously a mix of math
and out of the box imagination there.
- That I don't know.
Whatever I did, I exclude any imagination.
Because whatever I saw in machine learning
that come from imagination,
like features, like deep learning,
they are not relevant to the problem.
When you're looking very carefully
from mathematical equations,
you're deriving very simple theory,
which goes far beyond theoretically
than whatever people can imagine.
Because it is not good fantasy.
It is just interpretation, it is just fantasy,
but it is not what you need.
You don't need any imagination
to derive, say, main principle of machine learning.
- 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.
Do you think we have the tools currently?
Do you think we will ever have the tools
to try to describe that process of learning?
- 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,
Levenguk, for the first time,
only he got this instrument and nobody,
only he kept secret about microscope.
But he wrote report in London Academy of Science.
In his report, when he looking at the blood,
he look everywhere, on the water,
on the blood, on the skin.
But he described blood like fight
between queen and king.
So he saw blood cells, red cells,
and he imagined that it is army fighting each other.
§02Transcript · 11:11–21:43
And it was his interpretation of situation.
And he sent this report in Academy of Science.
They very carefully look because they believe
that he is right, he saw something.
But he gave wrong interpretation.
And I believe the same can happen with brain.
- With brain, yeah.
- Because the most important part,
you know, I believe in human language.
In some proverb, is so much wisdom.
For example, people say that it is better
than thousand days of diligent studies one day
with great teacher.
But if I will ask you what teacher does, nobody knows.
And that is intelligence.
And but we know from history,
and now from math and machine learning,
that teacher can do a lot.
- So what, from a mathematical point of view,
is the great teacher?
- I don't know.
- That's an open question.
- No, no, no, but we can say what teacher can do.
He can introduce some invariants,
some predicate for creating invariants.
How he doing it, I don't know,
because teacher knows reality,
and can describe from this reality,
a predicate invariants.
But he knows that when you're using invariant,
he can decrease number of observations 100 times.
- 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 guitar for a long time.
And 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 is a lot of instruction in that statement.
And so can you pull that apart?
What is that?
The language itself may not contain this information.
- It's not blah, blah, blah.
- It is not blah, blah, blah, yeah.
- It affects you.
- It's what?
- Affect you.
- Yeah.
- Affect your playing.
- Yes, it does, but it's not the lang,
it feels like a, what is the information
being exchanged there?
What is the nature of information?
What is the representation of that information?
- I believe that it is sort of predicate, but I don't know.
That is exactly what intelligence
in machine learning should be.
- Yes.
- Because the rest is just mathematical technique.
I think that what was discovered recently
is that there is two type, two mechanism of learning.
One called strong convergence mechanism
and weak convergence mechanism.
Before people use only one convergent.
In weak convergence mechanism, you can use predicate.
That's what play like butterfly
and it will immediately affect your playing.
You know, there is English proverb, great.
If it looks like a duck, swims like a duck,
and quack like a duck, then it is probably duck.
- Yes.
- But this is exact about predicate.
Looks like a duck, what it means.
So you saw many ducks that you're training data.
So you have description of how looks,
integral looks ducks.
- Yeah, the visual characteristics of a duck, yeah.
- Yeah, but you want, and you have model
for the cognition doubts.
So you would like to the theoretical description
from model coincide with empirical description
which you saw on teletext data.
So about looks like a duck, it is general.
But what about swims like a duck?
You should know the duck swims.
You can say it play chess like a duck.
Okay, duck doesn't play chess.
And it is completely legal predicate, but it is useless.
So half teacher can recognize not useless predicate.
So up to now, we don't use this predicate
in existing machine learning.
- And you think that's absolutely--
- So why we need zillions of data?
But in this English proverb,
they use only three predicate.
Looks like a duck, swims like a duck, and quack like a duck.
- So you can't deny the fact that swims like a duck
and quacks like a duck has humor in it, has ambiguity.
- Let's talk about swim like a duck.
It doesn't say jumps, jump like a duck.
Why?
Because-- - It's not relevant.
- But that means that you know ducks,
you know different birds, you know animals.
And you derive from this that it is relevant
to say swim like a duck.
- 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
which we pick up along the way.
You don't think so.
There doesn't need to be this knowledge base.
In those statements carries some rich information
that helps us understand the essence of duck.
How far are we from integrating predicates?
- You know that when you consider complete
theory of machine learning.
So what it does, you have a lot of functions.
And then you're talking, it looks like a duck.
You see your training data.
From training data you recognize like
expected duck should look.
Then you remove all functions which does not look
like you think it should look from training data.
So you decrease amount of function
from which you pick up one.
Then you give a second predicate.
And again decrease the set of function.
And after that you pick up the best function you can.
Fine, it is standard machine learning.
So why you need not too many examples.
- 'Cause your predicates aren't very good.
Or you're not--
- Yeah, that means the predicate variable.
Because every predicate is invented
to decrease admissible set of function.
- So you talk about admissible set of functions
and you talk about good functions.
So what makes a good function?
- So admissible set of function is set of function
which has small capacity or small diversity,
small VC dimension example.
Which contain good function inside.
- So by the way for people who don't know,
VC, you're the V in the VC.
So how would you describe to a layperson what VC theory is?
How would you describe VC?
- When you have a machine,
so machine capable to pick up one function
from the admissible set of function.
But set of admissible function can be big.
So it contain all continuous functions and it's useless.
You don't have so many examples to pick up function.
But it can be small.
Small, we call it capacity,
but maybe better call diversity.
So not very different function in the set.
It's infinite set of function, but not very diverse.
So it is small VC dimension.
When VC dimension is small,
you need small amount of training data.
So the goal is to create admissible set of functions
which is, have small VC dimension
and contain good function.
Then you should, you will be able to pick up the function
using small amount of observations.
- So that is the task of learning.
- Yeah.
- Is creating a set of admissible functions
that has a small VC dimension.
And then you've figured out a clever way of picking up.
- No, that is goal of learning,
which I formulated yesterday.
Statistical learning theory does not involve
in creating admissible set of function.
In classical learning theory,
everywhere, 100% in textbook,
the set of function, admissible set of function is given.
But this is science about nothing
because the most difficult problem
to create admissible set of functions
given, say, a lot of functions,
continual set of function.
Create admissible set of functions,
that means that it has finite VC dimension,
small VC dimension, and contain good function.
So this was out of consideration.
- So what's the process of doing that?
I mean, it's fascinating.
What is the process of creating this
admissible set of functions?
- That is invariance.
- That's invariance.
Can you describe invariance?
- 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.
You have model and what is expectation
of this function on the model.
§03Transcript · 21:44–32:25
And they should coincide.
So the problem is about how to pick up functions.
It can be any function.
In fact, it is true for all functions.
But because when we're talking, say,
duck does not jumping,
so you don't ask question, jump like a duck.
Because it is trivial, it does not jumping
and doesn't help you to recognize jump.
But you know something, which question to ask.
When you're asking, it swims like a duck.
But looks like a duck at this general situation.
Looks like, say, guy who have this illness,
this disease, it is legal.
So there is a general type of predicate,
looks like, and special type of predicate,
which related to this specific problem.
And that is intelligence part of all this business.
And that where teacher is involved.
- Incorporating the specialized predicates.
Okay, what do you think about deep learning
as neural networks, these arbitrary architectures
as helping accomplish some of the tasks
you're thinking about?
Their effectiveness or lack thereof?
What are the weaknesses and what are the possible strengths?
- You know, I think that this is fantasy.
Everything, which like deep learning, like features.
Let me give you this example.
One of the greatest book, this Churchill book
about history of Second World War.
And he's starting this book,
describing that in old time, when war is over,
so the great kings, they gather together,
and most all of them were relatives.
And they discussed what should be done,
how to create peace.
And they came to agreement.
And when happens First World War,
the general public came in power.
And they were so greedy that robbed Germany.
And it was clear for everybody that it is not peace.
That peace will last only 20 years.
Because they was not professionals.
And the same I see in machine learning.
There are mathematicians who are looking for the problem
from very deep point of view, mathematical point of view.
And there are computer scientists,
who mostly does not know mathematics.
They just have interpretation of that.
And they invented a lot of blah, blah, blah interpretations
like deep learning.
Why you did deep learning?
Mathematic does not know deep learning.
Mathematic does not know neurons.
It is just function.
If you like to say, piecewise linear function,
say that.
And do it in class of piecewise linear function.
But they invent something.
And then they try to prove advantage of that
through interpretations, which mostly wrong.
And when it not enough, 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 math.
This is mathematical problem.
Do your best to solve this problem.
Try to understand that there is not only one way
of convergence, which is strong way of convergence.
There is a weak way of convergence,
which requires predicate.
And if you will go through all this stuff,
you will see that you don't need deep learning.
Even more, I would say one of the theorem,
which called representative theory.
It says that optimal solution of mathematical problem,
which is described learning, is on shadow network.
Not on deep learning.
- And a shallow network, yeah.
The ultimate problem is there.
Absolutely, so in the end,
what you're saying is exactly right.
The question is, you have no value
for throwing something on the table,
playing with it, not math.
It's like in neural network,
where you said throwing something in the bucket
or the biological example and looking at kings and queens
or the cells with a microscope.
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.
You think that interpretation basically deceives you
in a way that's not productive.
- 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.
- That's right, but aren't you surprised
by the beauty of it?
So, not mathematical beauty,
but the fact that it works at all.
Or you're criticizing that very beauty,
our human desire to interpret,
to find our silly interpretations in these constructs.
Like, let me ask you this.
Are you surprised, does it inspire you?
How do you feel about the success of a system
like AlphaGo at beating the game of Go?
Using neural networks to estimate the quality of a board
and the quality of the--
- That is your interpretation, quality of the board.
- Yeah, yes.
- Yeah.
(laughing)
- But it's not our interpretation.
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,
does something that was thought impossible.
- That means that it's not a very difficult problem.
- That's it.
So, you empirically, we empirically have discovered
that this is not a very difficult problem.
- Yeah.
(laughing)
- It's true.
So, maybe, can't argue.
So--
- Even more, I would say,
that if they use deep learning,
it is not the most effective way of learning theory.
And usually, when people use deep learning,
they're using zillions of training data.
- Yeah.
- But you don't need this.
So, I describe challenge,
can we do some problems which do well,
deep learning method,
with deep net, using 100 times less training data.
Even more, some problems deep learning cannot solve,
because it's not necessary,
they create admissible set of function.
To create deep architecture means
to create admissible set of functions.
You cannot say that you're creating
good admissible set of functions.
You're just, it's your fantasy.
It is not comes from us.
But it is possible to create admissible set of functions,
because you have your training data.
That actually, for mathematicians,
when you consider invariant,
you need to use law of large numbers.
When you're making training in existing algorithm,
you need uniform law of large numbers,
which is much more difficult,
it requires dimension and all this stuff.
But, nevertheless, if you use both,
weak and strong way of convergence,
you can decrease a lot of training data.
- Yeah, you could do the three,
the swims like a duck and quacks like a duck.
- Yeah, yeah.
- So, let's step back and think about
human intelligence in general.
Clearly, that has evolved in a non-mathematical way.
It wasn't, as far as we know,
God or whoever didn't come up with a model
and place it in our brain of admissible functions.
It kind of evolved.
I don't know, maybe you have a view on this.
So, Alan Turing in the '50s, in his paper,
asked and rejected the question,
can machines think?
It's not a very useful question,
but can you briefly entertain this useless question?
Can machines think?
So, talk about intelligence and your view of it.
- I don't know that.
I know that Turing described imitation.
If computer can imitate human being,
let's call it intelligent.
And he understands that it is not thinking computer.
- Yes.
- He completely understands what he's doing.
But he 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 theorem,
it's very difficult theorem,
but in couple of years, in several places,
people prove the same theorem.
Say, Sauer Lemma after us was done.
Then another guys prove the same theorem.
§04Transcript · 32:28–43:31
In the history of science, it's happened all the time.
For example, geometry.
It's happened simultaneously.
First it did Lobachevsky, and then Gauss,
and Boyai, and another guys.
And it approximately in 10 times period,
10 years period of time.
And I saw a lot of examples like that.
And many mathematicians thinks that
when they develop something,
they develop something in general which affect everybody.
So, maybe our models that intelligence
only inside of us is incorrect.
- It's our interpretation, yeah.
It may be there exist some connection
with world intelligence.
I don't know.
- You're almost like plugging in into--
- Yeah, exactly.
- And contributing to this--
- Into 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 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?
- Yeah, it is interesting question.
But let me talk about complexity
in about worst case scenario.
There is a mathematical setting.
When I came to United States in 1990,
people did not know.
They see theories, they did not know statistical learning.
So in Russia it was published to monographs,
our monographs, but in America they did not know.
Then they learned it.
And somebody told me that it is worst case theory,
and they will create real case theory.
But till now it did not.
Because it is mathematical tool.
You can do only what you can do using mathematics.
And which has clear understanding, and clear description.
And for this reason, we introduce complexity.
And we need this, because using,
actually it is diversity, I like this one more.
You see 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,
it is entropy theory.
So from mathematical point of view,
you know the best possible case,
and the worst possible case.
You can derive different model in medium.
But it's not so interesting.
- You think the edges are interesting?
- The edges are interesting.
Because it is not so easy to get good bound, exact bound.
It's not many cases where you have the bound is not exact.
But interesting principles 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?
Or do you think it's interesting
because it's actually very useful
for understanding the essence of a function,
of an algorithm?
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.
It seems not productive.
- I don't think so because you cannot describe
situation in the middle.
Or it will be not general.
So you can describe edges cases,
and it is clear it has some model.
But you cannot describe model for every new case.
So you will be never accurate when you're using model.
- But from a statistical point of view,
the way you've studied functions
and the nature of learning in 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?
- I don't know that.
- Because--
- I think that, from my point of view,
if you will use formal statistics,
you need uniform law of large numbers.
If you will use this invariance business,
you will need just law of large numbers.
You don't, and there's this huge difference
between uniform law of large numbers and large numbers.
- Is it useful to describe that a little more?
Or should we just take it to--
- No, for example, when I'm talking about duck,
I gave three predicates and that was enough.
But if you will try to do formal, distinguish,
you will need a lot of observation.
- I gotcha.
- So that means that information about looks like a duck
contain a lot of bit of information,
formal bits of information.
So we don't know that, how much bit of information
contain things from artificial intelligence.
And that is the subject of analysis.
Till now, all business, I don't look at it
I don't like how people consider artificial intelligence.
They consider us some codes which imitate
activity of human being.
It is not science, it is applications.
You would like to imitate, go ahead,
it is very useful and good problem.
But you need to learn something more.
How people try to, how people can to develop,
say, predicates, swims like a duck,
or play like butterfly or something like that.
They're not, not the teacher says you,
how it came in his mind.
How he chooses image.
- So that process--
- That is problem of intelligence.
- That is the problem of intelligence.
And you see that connected to the problem of learning?
- Absolutely.
Because you immediately give this predicate,
like specific predicate, swims like a duck,
or quack like a duck.
It was chosen somehow.
- So what is the line of work, would you say?
If you were to formulate as a set of open problems,
that will take us there, to play like a butterfly,
we'll get a system to be able to--
- Let's separate two stories.
One mathematical story, that if you have predicate,
you can do something.
And another story, how to get predicate.
It is intelligence problem,
and people even did not start understanding intelligence.
Because to understand intelligence, first of all,
try to understand what doing teachers.
How teacher teach.
Why one teacher better than another one?
- Yeah, so you think we really even haven't started
on the journey of generating the predicates?
- No, we don't understand.
We even don't understand that this problem exists.
Because did you hear-- - You do.
- No, I just know name.
I want to understand why one teacher better than another.
And how affect teacher, student.
It is not because he repeating the problem
which is in textbook.
He make some remarks.
He make some philosophy of reasoning.
- Yeah, that's a beautiful, so it is a formulation
of a question that is the open problem.
Why is one teacher better than another?
- Right.
What he does better.
- Yeah, what, why at every level?
How do they get better?
What does it mean to be better?
The whole-- - Yeah, yeah.
From whatever model I have,
one teacher can give a very good predicate.
One teacher can say, "Swims like a dog."
And another can say, "Jump like a dog."
And jump like a dog carries zero information.
- Yeah. (laughs)
So what is the most exciting problem
in statistical learning?
Ever worked on or are working on now?
- I just finished this invariant story.
And I'm happy that I believe
that it is ultimate learning story.
At least I can show that there are no another mechanism,
only two mechanisms.
But they separate statistical part
from intelligent part.
And I know nothing about intelligent part.
And if we will know this intelligent part,
so it will help us a lot in teaching,
in learning.
- In learning. - Yeah.
- Do you know we'll know it when we see it?
- So for example, in my talk,
the last slide was a challenge.
So you have, say, NIST, digital recognition problem.
And deep learning claims that they did it very well.
Say, 99.5% of correct answers.
But they use 60,000 observations.
- Yeah.
- Can you do the same using 100 times less?
But incorporating invariants.
What it means, you know, digit one, two, three.
§05Transcript · 43:34–54:00
- Yeah.
- 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.
- Yeah, that last slide,
and unfortunately, your talk ended quickly,
but that last slide was a powerful open challenge
in a formulation of the essence here.
- Yeah, that is exact problem of intelligence.
Because everybody,
when machine learning started,
and it was developed by mathematicians,
they immediately recognized
that we use much more training data than humans needed.
But now again, we came to the same story.
Have to decrease.
And that is the problem of learning.
It is not like in deep learning,
they use zillions of training data.
Because maybe zillions are not enough
if you have a good invariants.
Maybe you will never collect some number of observations.
But now it is a question to intelligence.
How to do that?
Because statistical part is ready.
As soon as you supply us with predicate,
we can do a good job with small amount of observations.
And the very first challenge is well-known
digit recognition.
And you know digits.
And please, tell me invariants.
I think about that, I can say,
for digit three, I would introduce concept
of horizontal symmetry.
So the digit three has horizontal symmetry,
say more than say digit two or something like that.
But as soon as I get the idea of horizontal symmetry,
I can mathematically invent a lot of measure
of horizontal symmetry.
Or even vertical symmetry, or diagonal symmetry, whatever.
If I have idea of symmetry.
But what else?
Looking on digit, I see that it is meta-predicate.
Which is not shape.
It is something like symmetry,
like how dark is whole picture, something like that.
Which can self-rise a predicate.
- You think such a predicate could rise
out of something that's not general?
Meaning, it feels like for me to be able
to understand the difference between a two and a 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.
Or do you think there's a more efficient way?
- I don't know.
I know for sure that you must know something
more than digits.
- Yes, and that's a powerful statement.
- Yeah, but maybe there are several languages
of description, these elements of digits.
So I'm talking about symmetry,
about some properties of geometry,
I'm talking about something abstract.
I don't know that.
But this is a problem of intelligence.
So in one of our articles, it is trivial to show
that every example can carry not more than one bit
of information in real.
Because when you show example,
and you say this is one,
you can remove, say, a function which does not tell you one.
Say, it's the best strategy.
If you can do it perfectly,
it's remove half of the functions.
But when you use one predicate,
which looks like a duck,
you can remove much more functions than half.
And that means that it contains a lot of bit of information
from formal point of view.
But when you have a general picture
of what you want to recognize,
a general picture of the world,
can you invent this predicate?
And that predicate carry a lot of information.
- 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.
You really kind of talk about philosophy of science.
There's a poetry and music to a lot of the work you're doing
and the way you're thinking about it.
So where does that come from?
Do you escape to poetry?
Do you escape to music or not?
- I think that there exists ground truth.
- There exists ground truth?
- Yeah, and that can be seen everywhere.
The smart guy philosopher,
sometimes I surprise how they deep see.
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.
(laughing)
- So what piece of music,
as a piece of empirical evidence,
gave you a sense that they are touching something
in the ground truth?
- It is structure.
- The structure of the math of music.
- Yeah, because when you're listening to Bach,
you see the structure.
Very clear, very classic, very simple.
And the same in maths,
when you have axioms in geometry,
you have the same feeling.
And in poetry, sometimes you see the same.
- 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 United States,
you've been in a few places.
If you look back,
what was some of your happiest moments as a researcher?
Some of the most profound moments,
not in terms of their impact on society,
but in terms of their impact
on how damn good you feel that day,
and you remember that moment.
- You know, every time when you found something,
it is great.
- Yeah, one of the things in life.
- Every simple things.
- Just even-- - But my general feeling
that most of my time was wrong.
You should go again and again and again,
and try to be honest in front of yourself.
Not to make interpretation,
but try to understand that it related to ground truth.
It is not my blah, blah, blah interpretation
and something like that.
- But you're allowed to get excited
at the possibility of discovery.
- Oh yeah.
- You have to double check it, but--
- No, but how it related to the ground truth,
is it just temporary or it is forever?
- You know, you always have a feeling
when you found something, how big is that?
So 20 years ago when we discovered statistical learning,
so nobody believed, except for one guy, Dudley, from MIT.
And then in 20 years, it became fashion.
And the same with support vector machines.
That's kernel machines.
- 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 profundity of it?
How that this seems to be right, this seems to be powerful.
- Right, absolutely, immediately.
I recognize that it will last forever.
And now, when I found this invariance story,
I have a feeling that it is completely wrong.
Because I have proof that there are no different mechanism.
You can have some cosmetic improvement, you can do,
but in terms of invariance, you need both invariance
and statistical learning, and they should work together.
But also, I'm happy that we can formulate
what is intelligence, from that.
And to separate from technical part.
That is completely different.
- Absolutely.
Well, Vladimir, thank you so much for talking today.
- Thank you.