episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
What we proved is that if you have deep layers,
https://karpathy.ai/lexicap/0013-large.html#00:47:28.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
hierarchical architecture with the local connectivity
https://karpathy.ai/lexicap/0013-large.html#00:47:32.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of the type of convolutional deep learning,
https://karpathy.ai/lexicap/0013-large.html#00:47:36.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and if you're dealing with a function that
https://karpathy.ai/lexicap/0013-large.html#00:47:39.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
has this kind of hierarchical architecture,
https://karpathy.ai/lexicap/0013-large.html#00:47:42.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
then you avoid completely the curse.
https://karpathy.ai/lexicap/0013-large.html#00:47:46.680
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
You've spoken a lot about supervised deep learning.
https://karpathy.ai/lexicap/0013-large.html#00:47:50.680
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
What are your thoughts, hopes, views
https://karpathy.ai/lexicap/0013-large.html#00:47:54.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
on the challenges of unsupervised learning
https://karpathy.ai/lexicap/0013-large.html#00:47:56.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
with GANs, with Generative Adversarial Networks?
https://karpathy.ai/lexicap/0013-large.html#00:47:59.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Do you see those as distinct?
https://karpathy.ai/lexicap/0013-large.html#00:48:05.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
The power of GANs, do you see those
https://karpathy.ai/lexicap/0013-large.html#00:48:08.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
as distinct from supervised methods in neural networks,
https://karpathy.ai/lexicap/0013-large.html#00:48:09.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
or are they really all in the same representation ballpark?
https://karpathy.ai/lexicap/0013-large.html#00:48:13.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
GANs is one way to get estimation of probability
https://karpathy.ai/lexicap/0013-large.html#00:48:16.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
densities, which is a somewhat new way that people have not
https://karpathy.ai/lexicap/0013-large.html#00:48:24.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
done before.
https://karpathy.ai/lexicap/0013-large.html#00:48:28.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I don't know whether this will really play an important role
https://karpathy.ai/lexicap/0013-large.html#00:48:30.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in intelligence.
https://karpathy.ai/lexicap/0013-large.html#00:48:36.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Or it's interesting.
https://karpathy.ai/lexicap/0013-large.html#00:48:39.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I'm less enthusiastic about it than many people in the field.
https://karpathy.ai/lexicap/0013-large.html#00:48:43.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I have the feeling that many people in the field
https://karpathy.ai/lexicap/0013-large.html#00:48:48.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
are really impressed by the ability
https://karpathy.ai/lexicap/0013-large.html#00:48:50.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of producing realistic looking images in this generative way.
https://karpathy.ai/lexicap/0013-large.html#00:48:54.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Which describes the popularity of the methods.
https://karpathy.ai/lexicap/0013-large.html#00:49:01.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But you're saying that while that's exciting and cool
https://karpathy.ai/lexicap/0013-large.html#00:49:03.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to look at, it may not be the tool that's useful for it.
https://karpathy.ai/lexicap/0013-large.html#00:49:06.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So you describe it kind of beautifully.
https://karpathy.ai/lexicap/0013-large.html#00:49:11.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Current supervised methods go n to infinity
https://karpathy.ai/lexicap/0013-large.html#00:49:13.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in terms of number of labeled points.
https://karpathy.ai/lexicap/0013-large.html#00:49:16.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And we really have to figure out how to go to n to 1.
https://karpathy.ai/lexicap/0013-large.html#00:49:18.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And you're thinking GANs might help,
https://karpathy.ai/lexicap/0013-large.html#00:49:21.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
but they might not be the right.
https://karpathy.ai/lexicap/0013-large.html#00:49:23.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I don't think for that problem, which I really think
https://karpathy.ai/lexicap/0013-large.html#00:49:25.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is important, I think they may help.
https://karpathy.ai/lexicap/0013-large.html#00:49:28.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
They certainly have applications,
https://karpathy.ai/lexicap/0013-large.html#00:49:32.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
for instance, in computer graphics.
https://karpathy.ai/lexicap/0013-large.html#00:49:33.680
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And I did work long ago, which was
https://karpathy.ai/lexicap/0013-large.html#00:49:35.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
a little bit similar in terms of saying, OK, I have a network.
https://karpathy.ai/lexicap/0013-large.html#00:49:41.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And I present images.
https://karpathy.ai/lexicap/0013-large.html#00:49:47.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And I can input its images.
https://karpathy.ai/lexicap/0013-large.html#00:49:49.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And output is, for instance, the pose of the image.
https://karpathy.ai/lexicap/0013-large.html#00:49:54.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
A face, how much is smiling, is rotated 45 degrees or not.
https://karpathy.ai/lexicap/0013-large.html#00:49:57.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
What about having a network that I train with the same data
https://karpathy.ai/lexicap/0013-large.html#00:50:02.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
set, but now I invert input and output.
https://karpathy.ai/lexicap/0013-large.html#00:50:07.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Now the input is the pose or the expression, a number,
https://karpathy.ai/lexicap/0013-large.html#00:50:10.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
set of numbers.
https://karpathy.ai/lexicap/0013-large.html#00:50:15.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And the output is the image.
https://karpathy.ai/lexicap/0013-large.html#00:50:16.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And I train it.
https://karpathy.ai/lexicap/0013-large.html#00:50:18.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And we did pretty good, interesting results
https://karpathy.ai/lexicap/0013-large.html#00:50:20.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in terms of producing very realistic looking images.
https://karpathy.ai/lexicap/0013-large.html#00:50:22.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It was a less sophisticated mechanism.
https://karpathy.ai/lexicap/0013-large.html#00:50:27.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But the output was pretty less than GANs.
https://karpathy.ai/lexicap/0013-large.html#00:50:31.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But the output was pretty much of the same quality.
https://karpathy.ai/lexicap/0013-large.html#00:50:35.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So I think for a computer graphics type application,
https://karpathy.ai/lexicap/0013-large.html#00:50:38.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
yeah, definitely GANs can be quite useful.
https://karpathy.ai/lexicap/0013-large.html#00:50:43.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And not only for that, but for helping,
https://karpathy.ai/lexicap/0013-large.html#00:50:46.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
for instance, on this problem of unsupervised example
https://karpathy.ai/lexicap/0013-large.html#00:50:52.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of reducing the number of labeled examples.
https://karpathy.ai/lexicap/0013-large.html#00:50:58.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I think people, it's like they think they can get out
https://karpathy.ai/lexicap/0013-large.html#00:51:02.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
more than they put in.
https://karpathy.ai/lexicap/0013-large.html#00:51:07.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
There's no free lunch, as you said.
https://karpathy.ai/lexicap/0013-large.html#00:51:11.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
What do you think, what's your intuition?
https://karpathy.ai/lexicap/0013-large.html#00:51:14.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
How can we slow the growth of N to infinity in supervised,
https://karpathy.ai/lexicap/0013-large.html#00:51:17.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
N to infinity in supervised learning?
https://karpathy.ai/lexicap/0013-large.html#00:51:22.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So for example, Mobileye has very successfully,
https://karpathy.ai/lexicap/0013-large.html#00:51:25.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I mean, essentially annotated large amounts of data
https://karpathy.ai/lexicap/0013-large.html#00:51:29.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to be able to drive a car.
https://karpathy.ai/lexicap/0013-large.html#00:51:33.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Now one thought is, so we're trying
https://karpathy.ai/lexicap/0013-large.html#00:51:34.680
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to teach machines, school of AI.
https://karpathy.ai/lexicap/0013-large.html#00:51:37.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And we're trying to, so how can we become better teachers,
https://karpathy.ai/lexicap/0013-large.html#00:51:41.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
maybe?
https://karpathy.ai/lexicap/0013-large.html#00:51:45.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
That's one way.
https://karpathy.ai/lexicap/0013-large.html#00:51:46.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
No, I like that.
https://karpathy.ai/lexicap/0013-large.html#00:51:47.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Because again, one caricature of the history of computer
https://karpathy.ai/lexicap/0013-large.html#00:51:51.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
science, you could say, begins with programmers, expensive.
https://karpathy.ai/lexicap/0013-large.html#00:51:57.680
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Continuous labelers, cheap.
https://karpathy.ai/lexicap/0013-large.html#00:52:05.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And the future will be schools, like we have for kids.
https://karpathy.ai/lexicap/0013-large.html#00:52:09.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Yeah.
https://karpathy.ai/lexicap/0013-large.html#00:52:14.680
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Currently, the labeling methods were not
https://karpathy.ai/lexicap/0013-large.html#00:52:16.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
selective about which examples we teach networks with.
https://karpathy.ai/lexicap/0013-large.html#00:52:20.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So I think the focus of making networks that learn much faster
https://karpathy.ai/lexicap/0013-large.html#00:52:25.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is often on the architecture side.
https://karpathy.ai/lexicap/0013-large.html#00:52:31.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But how can we pick better examples with which to learn?
https://karpathy.ai/lexicap/0013-large.html#00:52:33.680
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Do you have intuitions about that?
https://karpathy.ai/lexicap/0013-large.html#00:52:37.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Well, that's part of the problem.
https://karpathy.ai/lexicap/0013-large.html#00:52:39.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But the other one is, if we look at biology,
https://karpathy.ai/lexicap/0013-large.html#00:52:42.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
a reasonable assumption, I think,
https://karpathy.ai/lexicap/0013-large.html#00:52:50.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is in the same spirit that I said,
https://karpathy.ai/lexicap/0013-large.html#00:52:52.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
evolution is opportunistic and has weak priors.
https://karpathy.ai/lexicap/0013-large.html#00:52:58.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
The way I think the intelligence of a child,
https://karpathy.ai/lexicap/0013-large.html#00:53:03.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
the baby may develop is by bootstrapping weak priors
https://karpathy.ai/lexicap/0013-large.html#00:53:08.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
from evolution.
https://karpathy.ai/lexicap/0013-large.html#00:53:16.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
For instance, you can assume that you
https://karpathy.ai/lexicap/0013-large.html#00:53:17.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
have in most organisms, including human babies,
https://karpathy.ai/lexicap/0013-large.html#00:53:24.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
built in some basic machinery to detect motion
https://karpathy.ai/lexicap/0013-large.html#00:53:28.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and relative motion.
https://karpathy.ai/lexicap/0013-large.html#00:53:35.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And in fact, we know all insects from fruit flies
https://karpathy.ai/lexicap/0013-large.html#00:53:38.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to other animals, they have this,
https://karpathy.ai/lexicap/0013-large.html#00:53:42.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
even in the retinas, in the very peripheral part.
https://karpathy.ai/lexicap/0013-large.html#00:53:49.760