episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
In learning.
https://karpathy.ai/lexicap/0005-large.html#00:42:59.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Yeah.
https://karpathy.ai/lexicap/0005-large.html#00:43:00.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
You will know it when we see it?
https://karpathy.ai/lexicap/0005-large.html#00:43:01.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
So for example, in my talk, the last slide was a challenge.
https://karpathy.ai/lexicap/0005-large.html#00:43:02.920
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
So you have say NIST digit recognition problem and deep learning claims that they did it
https://karpathy.ai/lexicap/0005-large.html#00:43:07.100
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
very well, say 99.5% of correct answers.
https://karpathy.ai/lexicap/0005-large.html#00:43:14.680
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But they use 60,000 observations.
https://karpathy.ai/lexicap/0005-large.html#00:43:22.100
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Can you do the same using hundred times less?
https://karpathy.ai/lexicap/0005-large.html#00:43:25.280
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But incorporating invariants, what it means, you know, digit one, two, three.
https://karpathy.ai/lexicap/0005-large.html#00:43:29.560
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But looking on that, explain to me which invariant I should keep to use hundred examples or say
https://karpathy.ai/lexicap/0005-large.html#00:43:35.280
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
hundred times less examples to do the same job.
https://karpathy.ai/lexicap/0005-large.html#00:43:44.040
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Yeah, that last slide, unfortunately your talk ended quickly, but that last slide was
https://karpathy.ai/lexicap/0005-large.html#00:43:47.800
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
a powerful open challenge and a formulation of the essence here.
https://karpathy.ai/lexicap/0005-large.html#00:43:56.520
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
What is the exact problem of intelligence?
https://karpathy.ai/lexicap/0005-large.html#00:44:01.960
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Because everybody, when machine learning started and it was developed by mathematicians, they
https://karpathy.ai/lexicap/0005-large.html#00:44:06.300
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
immediately recognized that we use much more training data than humans needed.
https://karpathy.ai/lexicap/0005-large.html#00:44:15.040
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But now again, we came to the same story, have to decrease.
https://karpathy.ai/lexicap/0005-large.html#00:44:22.540
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
That is the problem of learning.
https://karpathy.ai/lexicap/0005-large.html#00:44:27.640
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
It is not like in deep learning, they use zillions of training data because maybe zillions
https://karpathy.ai/lexicap/0005-large.html#00:44:30.660
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
are not enough if you have a good invariants.
https://karpathy.ai/lexicap/0005-large.html#00:44:37.320
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Maybe you will never collect some number of observations.
https://karpathy.ai/lexicap/0005-large.html#00:44:44.720
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But now it is a question to intelligence, how to do that?
https://karpathy.ai/lexicap/0005-large.html#00:44:49.520
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Because statistical part is ready, as soon as you supply us with predicate, we can do
https://karpathy.ai/lexicap/0005-large.html#00:44:56.080
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
good job with small amount of observations.
https://karpathy.ai/lexicap/0005-large.html#00:45:03.200
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And the very first challenge is well known digit recognition.
https://karpathy.ai/lexicap/0005-large.html#00:45:06.880
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And you know digits, and please tell me invariants.
https://karpathy.ai/lexicap/0005-large.html#00:45:11.040
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
I think about that, I can say for digit three, I would introduce concept of horizontal symmetry.
https://karpathy.ai/lexicap/0005-large.html#00:45:15.560
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
So the digit three has horizontal symmetry, say more than, say, digit two or something
https://karpathy.ai/lexicap/0005-large.html#00:45:25.760
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
like that.
https://karpathy.ai/lexicap/0005-large.html#00:45:32.440
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But as soon as I get the idea of horizontal symmetry, I can mathematically invent a lot
https://karpathy.ai/lexicap/0005-large.html#00:45:33.440
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
of measure of horizontal symmetry, or then vertical symmetry, or diagonal symmetry, whatever,
https://karpathy.ai/lexicap/0005-large.html#00:45:40.480
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
if I have idea of symmetry.
https://karpathy.ai/lexicap/0005-large.html#00:45:47.360
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But what else?
https://karpathy.ai/lexicap/0005-large.html#00:45:49.980
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
I think on digit I see that it is meta predicate, which is not shape, it is something like symmetry,
https://karpathy.ai/lexicap/0005-large.html#00:45:52.800
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
like how dark is whole picture, something like that, which can self rise a predicate.
https://karpathy.ai/lexicap/0005-large.html#00:46:07.600
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
You think such a predicate could rise out of something that is not general, meaning
https://karpathy.ai/lexicap/0005-large.html#00:46:16.240
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
it feels like for me to be able to understand the difference between two and three, I would
https://karpathy.ai/lexicap/0005-large.html#00:46:29.800
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
need to have had a childhood of 10 to 15 years playing with kids, going to school, being
https://karpathy.ai/lexicap/0005-large.html#00:46:35.640
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
yelled by parents, all of that, walking, jumping, looking at ducks, and then I would be able
https://karpathy.ai/lexicap/0005-large.html#00:46:48.080
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
to generate the right predicate for telling the difference between two and a three.
https://karpathy.ai/lexicap/0005-large.html#00:46:57.880
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Or do you think there's a more efficient way?
https://karpathy.ai/lexicap/0005-large.html#00:47:03.120
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
I don't know.
https://karpathy.ai/lexicap/0005-large.html#00:47:05.720
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
I know for sure that you must know something more than digits.
https://karpathy.ai/lexicap/0005-large.html#00:47:06.720
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Yes.
https://karpathy.ai/lexicap/0005-large.html#00:47:12.200
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And that's a powerful statement.
https://karpathy.ai/lexicap/0005-large.html#00:47:13.200
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Yeah.
https://karpathy.ai/lexicap/0005-large.html#00:47:15.000
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But maybe there are several languages of description, these elements of digits.
https://karpathy.ai/lexicap/0005-large.html#00:47:16.000
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
So I'm talking about symmetry, about some properties of geometry, I'm talking about
https://karpathy.ai/lexicap/0005-large.html#00:47:24.600
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
something abstract.
https://karpathy.ai/lexicap/0005-large.html#00:47:32.000
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
I don't know that.
https://karpathy.ai/lexicap/0005-large.html#00:47:33.000
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But this is a problem of intelligence.
https://karpathy.ai/lexicap/0005-large.html#00:47:34.780
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
So in one of our articles, it is trivial to show that every example can carry not more
https://karpathy.ai/lexicap/0005-large.html#00:47:38.900
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
than one bit of information in real.
https://karpathy.ai/lexicap/0005-large.html#00:47:47.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Because when you show example and you say this is one, you can remove, say, a function
https://karpathy.ai/lexicap/0005-large.html#00:47:50.240
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
which does not tell you one, say, is the best strategy.
https://karpathy.ai/lexicap/0005-large.html#00:48:00.660
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
If you can do it perfectly, it's remove half of the functions.
https://karpathy.ai/lexicap/0005-large.html#00:48:05.080
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But when you use one predicate, which looks like a duck, you can remove much more functions
https://karpathy.ai/lexicap/0005-large.html#00:48:10.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
than half.
https://karpathy.ai/lexicap/0005-large.html#00:48:17.080
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And that means that it contains a lot of bit of information from formal point of view.
https://karpathy.ai/lexicap/0005-large.html#00:48:18.920
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But when you have a general picture of what you want to recognize and general picture
https://karpathy.ai/lexicap/0005-large.html#00:48:26.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
of the world, can you invent this predicate?
https://karpathy.ai/lexicap/0005-large.html#00:48:34.640
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And that predicate carries a lot of information.
https://karpathy.ai/lexicap/0005-large.html#00:48:40.960
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Beautifully put.
https://karpathy.ai/lexicap/0005-large.html#00:48:47.560
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Maybe just me, but in all the math you show, in your work, which is some of the most profound
https://karpathy.ai/lexicap/0005-large.html#00:48:48.960
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
mathematical work in the field of learning AI and just math in general, I hear a lot
https://karpathy.ai/lexicap/0005-large.html#00:48:56.000
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
of poetry and philosophy.
https://karpathy.ai/lexicap/0005-large.html#00:49:02.320
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
You really kind of talk about philosophy of science.
https://karpathy.ai/lexicap/0005-large.html#00:49:04.400
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
There's a poetry and music to a lot of the work you're doing and the way you're thinking
https://karpathy.ai/lexicap/0005-large.html#00:49:09.920
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
about it.
https://karpathy.ai/lexicap/0005-large.html#00:49:13.320
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
So do you, where does that come from?
https://karpathy.ai/lexicap/0005-large.html#00:49:14.320
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Do you escape to poetry?
https://karpathy.ai/lexicap/0005-large.html#00:49:16.680
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Do you escape to music or not?
https://karpathy.ai/lexicap/0005-large.html#00:49:18.880
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
I think that there exists ground truth.
https://karpathy.ai/lexicap/0005-large.html#00:49:21.360
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
There exists ground truth?
https://karpathy.ai/lexicap/0005-large.html#00:49:23.840
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Yeah.
https://karpathy.ai/lexicap/0005-large.html#00:49:25.760
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And that can be seen everywhere.
https://karpathy.ai/lexicap/0005-large.html#00:49:26.760
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
The smart guy, philosopher, sometimes I'm surprised how they deep see.
https://karpathy.ai/lexicap/0005-large.html#00:49:30.720
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Sometimes I see that some of them are completely out of subject.
https://karpathy.ai/lexicap/0005-large.html#00:49:39.000
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But the ground truth I see in music.
https://karpathy.ai/lexicap/0005-large.html#00:49:45.560
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Music is the ground truth?
https://karpathy.ai/lexicap/0005-large.html#00:49:50.960
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Yeah.
https://karpathy.ai/lexicap/0005-large.html#00:49:51.960
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And in poetry, many poets, they believe, they take dictation.
https://karpathy.ai/lexicap/0005-large.html#00:49:52.960
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
So what piece of music as a piece of empirical evidence gave you a sense that they are touching
https://karpathy.ai/lexicap/0005-large.html#00:50:01.880
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
something in the ground truth?
https://karpathy.ai/lexicap/0005-large.html#00:50:12.360
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
It is structure.
https://karpathy.ai/lexicap/0005-large.html#00:50:14.560
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
The structure of the math of music.
https://karpathy.ai/lexicap/0005-large.html#00:50:16.720
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Yeah, because when you're listening to Bach, you see the structure.
https://karpathy.ai/lexicap/0005-large.html#00:50:17.720
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
Very clear, very classic, very simple, and the same in math when you have axioms in geometry,
https://karpathy.ai/lexicap/0005-large.html#00:50:22.360
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
you have the same feeling.
https://karpathy.ai/lexicap/0005-large.html#00:50:31.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And in poetry, sometimes you see the same.
https://karpathy.ai/lexicap/0005-large.html#00:50:32.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
And if you look back at your childhood, you grew up in Russia, you maybe were born as
https://karpathy.ai/lexicap/0005-large.html#00:50:38.360
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
a researcher in Russia, you've developed as a researcher in Russia, you've came to United
https://karpathy.ai/lexicap/0005-large.html#00:50:44.580
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
States and a few places.
https://karpathy.ai/lexicap/0005-large.html#00:50:48.680
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
If you look back, what was some of your happiest moments as a researcher, some of the most
https://karpathy.ai/lexicap/0005-large.html#00:50:51.800
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
profound moments, not in terms of their impact on society, but in terms of their impact on
https://karpathy.ai/lexicap/0005-large.html#00:51:00.000
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
how damn good you feel that day and you remember that moment?
https://karpathy.ai/lexicap/0005-large.html#00:51:09.960
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
You know, every time when you found something, it is great in the life, every simple things.
https://karpathy.ai/lexicap/0005-large.html#00:51:15.400
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
But my general feeling is that most of my time was wrong.
https://karpathy.ai/lexicap/0005-large.html#00:51:26.600
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
You should go again and again and again and try to be honest in front of yourself, not
https://karpathy.ai/lexicap/0005-large.html#00:51:32.160
Vladimir Vapnik: Statistical Learning | Lex Fridman Podcast #5
to make interpretation, but try to understand that it's related to ground truth, it is not
https://karpathy.ai/lexicap/0005-large.html#00:51:39.520