episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
How far can we get in this kind of way?
https://karpathy.ai/lexicap/0013-large.html#00:35:44.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So here is almost philosophy.
https://karpathy.ai/lexicap/0013-large.html#00:35:47.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Well, let's go there.
https://karpathy.ai/lexicap/0013-large.html#00:35:51.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Yeah, let's go there.
https://karpathy.ai/lexicap/0013-large.html#00:35:53.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So a friend of mine, Max Tegmark, who is a physicist at MIT.
https://karpathy.ai/lexicap/0013-large.html#00:35:54.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I've talked to him on this thing.
https://karpathy.ai/lexicap/0013-large.html#00:36:00.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Yeah, and he disagrees with you, right?
https://karpathy.ai/lexicap/0013-large.html#00:36:01.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
A little bit.
https://karpathy.ai/lexicap/0013-large.html#00:36:03.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Yeah, we agree on most.
https://karpathy.ai/lexicap/0013-large.html#00:36:04.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But the conclusion is a bit different.
https://karpathy.ai/lexicap/0013-large.html#00:36:07.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
His conclusion is that for images, for instance,
https://karpathy.ai/lexicap/0013-large.html#00:36:10.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
the compositional structure of this function
https://karpathy.ai/lexicap/0013-large.html#00:36:14.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that we have to learn or to solve these problems
https://karpathy.ai/lexicap/0013-large.html#00:36:19.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
comes from physics, comes from the fact
https://karpathy.ai/lexicap/0013-large.html#00:36:23.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that you have local interactions in physics
https://karpathy.ai/lexicap/0013-large.html#00:36:27.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
between atoms and other atoms, between particle
https://karpathy.ai/lexicap/0013-large.html#00:36:31.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of matter and other particles, between planets
https://karpathy.ai/lexicap/0013-large.html#00:36:37.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and other planets, between stars and other.
https://karpathy.ai/lexicap/0013-large.html#00:36:41.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It's all local.
https://karpathy.ai/lexicap/0013-large.html#00:36:44.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And that's true.
https://karpathy.ai/lexicap/0013-large.html#00:36:48.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But you could push this argument a bit further.
https://karpathy.ai/lexicap/0013-large.html#00:36:51.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Not this argument, actually.
https://karpathy.ai/lexicap/0013-large.html#00:36:56.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
You could argue that maybe that's part of the truth.
https://karpathy.ai/lexicap/0013-large.html#00:36:57.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But maybe what happens is kind of the opposite,
https://karpathy.ai/lexicap/0013-large.html#00:37:02.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is that our brain is wired up as a deep network.
https://karpathy.ai/lexicap/0013-large.html#00:37:06.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So it can learn, understand, solve
https://karpathy.ai/lexicap/0013-large.html#00:37:11.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
problems that have this compositional structure
https://karpathy.ai/lexicap/0013-large.html#00:37:18.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and it cannot solve problems that don't have
https://karpathy.ai/lexicap/0013-large.html#00:37:22.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
this compositional structure.
https://karpathy.ai/lexicap/0013-large.html#00:37:27.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So the problems we are accustomed to, we think about,
https://karpathy.ai/lexicap/0013-large.html#00:37:29.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
we test our algorithms on, are this compositional structure
https://karpathy.ai/lexicap/0013-large.html#00:37:34.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
because our brain is made up.
https://karpathy.ai/lexicap/0013-large.html#00:37:40.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And that's, in a sense, an evolutionary perspective
https://karpathy.ai/lexicap/0013-large.html#00:37:42.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that we've.
https://karpathy.ai/lexicap/0013-large.html#00:37:45.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So the ones that didn't have, that weren't
https://karpathy.ai/lexicap/0013-large.html#00:37:46.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
dealing with the compositional nature of reality died off?
https://karpathy.ai/lexicap/0013-large.html#00:37:50.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Yes, but also could be maybe the reason
https://karpathy.ai/lexicap/0013-large.html#00:37:55.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
why we have this local connectivity in the brain,
https://karpathy.ai/lexicap/0013-large.html#00:38:00.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
like simple cells in cortex looking
https://karpathy.ai/lexicap/0013-large.html#00:38:05.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
only at the small part of the image, each one of them,
https://karpathy.ai/lexicap/0013-large.html#00:38:08.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and then other cells looking at the small number
https://karpathy.ai/lexicap/0013-large.html#00:38:11.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of these simple cells and so on.
https://karpathy.ai/lexicap/0013-large.html#00:38:14.680
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
The reason for this may be purely
https://karpathy.ai/lexicap/0013-large.html#00:38:16.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that it was difficult to grow long range connectivity.
https://karpathy.ai/lexicap/0013-large.html#00:38:19.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So suppose it's for biology.
https://karpathy.ai/lexicap/0013-large.html#00:38:25.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It's possible to grow short range connectivity but not
https://karpathy.ai/lexicap/0013-large.html#00:38:28.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
long range also because there is a limited number of long range
https://karpathy.ai/lexicap/0013-large.html#00:38:34.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that you can.
https://karpathy.ai/lexicap/0013-large.html#00:38:38.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And so you have this limitation from the biology.
https://karpathy.ai/lexicap/0013-large.html#00:38:39.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And this means you build a deep convolutional network.
https://karpathy.ai/lexicap/0013-large.html#00:38:45.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
This would be something like a deep convolutional network.
https://karpathy.ai/lexicap/0013-large.html#00:38:50.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And this is great for solving certain class of problems.
https://karpathy.ai/lexicap/0013-large.html#00:38:53.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
These are the ones we find easy and important for our life.
https://karpathy.ai/lexicap/0013-large.html#00:38:57.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And yes, they were enough for us to survive.
https://karpathy.ai/lexicap/0013-large.html#00:39:02.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And you can start a successful business
https://karpathy.ai/lexicap/0013-large.html#00:39:07.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
on solving those problems with Mobileye.
https://karpathy.ai/lexicap/0013-large.html#00:39:10.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Driving is a compositional problem.
https://karpathy.ai/lexicap/0013-large.html#00:39:14.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So on the learning task, we don't
https://karpathy.ai/lexicap/0013-large.html#00:39:17.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
know much about how the brain learns
https://karpathy.ai/lexicap/0013-large.html#00:39:21.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in terms of optimization.
https://karpathy.ai/lexicap/0013-large.html#00:39:24.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So the thing that's stochastic gradient descent
https://karpathy.ai/lexicap/0013-large.html#00:39:26.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is what artificial neural networks use for the most part
https://karpathy.ai/lexicap/0013-large.html#00:39:29.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to adjust the parameters in such a way that it's
https://karpathy.ai/lexicap/0013-large.html#00:39:33.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
able to deal based on the label data,
https://karpathy.ai/lexicap/0013-large.html#00:39:37.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
it's able to solve the problem.
https://karpathy.ai/lexicap/0013-large.html#00:39:40.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So what's your intuition about why it works at all?
https://karpathy.ai/lexicap/0013-large.html#00:39:42.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
How hard of a problem it is to optimize
https://karpathy.ai/lexicap/0013-large.html#00:39:50.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
a neural network, artificial neural network?
https://karpathy.ai/lexicap/0013-large.html#00:39:53.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Is there other alternatives?
https://karpathy.ai/lexicap/0013-large.html#00:39:56.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Just in general, your intuition is
https://karpathy.ai/lexicap/0013-large.html#00:39:58.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
behind this very simplistic algorithm
https://karpathy.ai/lexicap/0013-large.html#00:40:01.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that seems to do pretty good, surprisingly so.
https://karpathy.ai/lexicap/0013-large.html#00:40:03.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Yes.
https://karpathy.ai/lexicap/0013-large.html#00:40:06.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So I find neuroscience, the architecture of cortex,
https://karpathy.ai/lexicap/0013-large.html#00:40:07.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is really similar to the architecture of deep networks.
https://karpathy.ai/lexicap/0013-large.html#00:40:13.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So there is a nice correspondence there
https://karpathy.ai/lexicap/0013-large.html#00:40:17.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
between the biology and this kind
https://karpathy.ai/lexicap/0013-large.html#00:40:20.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of local connectivity, hierarchical architecture.
https://karpathy.ai/lexicap/0013-large.html#00:40:23.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
The stochastic gradient descent, as you said,
https://karpathy.ai/lexicap/0013-large.html#00:40:28.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is a very simple technique.
https://karpathy.ai/lexicap/0013-large.html#00:40:30.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It seems pretty unlikely that biology could do that
https://karpathy.ai/lexicap/0013-large.html#00:40:35.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
from what we know right now about cortex and neurons
https://karpathy.ai/lexicap/0013-large.html#00:40:41.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and synapses.
https://karpathy.ai/lexicap/0013-large.html#00:40:47.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So it's a big question open whether there
https://karpathy.ai/lexicap/0013-large.html#00:40:50.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
are other optimization learning algorithms that
https://karpathy.ai/lexicap/0013-large.html#00:40:53.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
can replace stochastic gradient descent.
https://karpathy.ai/lexicap/0013-large.html#00:40:59.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And my guess is yes, but nobody has found yet a real answer.
https://karpathy.ai/lexicap/0013-large.html#00:41:02.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I mean, people are trying, still trying,
https://karpathy.ai/lexicap/0013-large.html#00:41:11.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and there are some interesting ideas.
https://karpathy.ai/lexicap/0013-large.html#00:41:13.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
The fact that stochastic gradient descent
https://karpathy.ai/lexicap/0013-large.html#00:41:18.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is so successful, this has become clearly not so
https://karpathy.ai/lexicap/0013-large.html#00:41:22.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
mysterious.
https://karpathy.ai/lexicap/0013-large.html#00:41:26.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And the reason is that it's an interesting fact.
https://karpathy.ai/lexicap/0013-large.html#00:41:27.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It's a change, in a sense, in how
https://karpathy.ai/lexicap/0013-large.html#00:41:33.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
people think about statistics.
https://karpathy.ai/lexicap/0013-large.html#00:41:36.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And this is the following, is that typically when
https://karpathy.ai/lexicap/0013-large.html#00:41:39.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
you had data and you had, say, a model with parameters,
https://karpathy.ai/lexicap/0013-large.html#00:41:45.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
you are trying to fit the model to the data,
https://karpathy.ai/lexicap/0013-large.html#00:41:51.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to fit the parameter.
https://karpathy.ai/lexicap/0013-large.html#00:41:54.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Typically, the kind of crowd wisdom type idea
https://karpathy.ai/lexicap/0013-large.html#00:41:55.960