episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
has been that the main breakthroughs,
https://karpathy.ai/lexicap/0013-large.html#00:11:18.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
the main recent breakthroughs, really start from neuroscience.
https://karpathy.ai/lexicap/0013-large.html#00:11:23.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I can mention reinforcement learning as one.
https://karpathy.ai/lexicap/0013-large.html#00:11:32.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It's one of the algorithms at the core of AlphaGo,
https://karpathy.ai/lexicap/0013-large.html#00:11:35.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
which is the system that beat the kind of an official world
https://karpathy.ai/lexicap/0013-large.html#00:11:41.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
champion of Go, Lee Sedol, two, three years ago in Seoul.
https://karpathy.ai/lexicap/0013-large.html#00:11:45.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
That's one.
https://karpathy.ai/lexicap/0013-large.html#00:11:52.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And that started really with the work of Pavlov in 1900,
https://karpathy.ai/lexicap/0013-large.html#00:11:53.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Marvin Minsky in the 60s, and many other neuroscientists
https://karpathy.ai/lexicap/0013-large.html#00:12:00.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
later on.
https://karpathy.ai/lexicap/0013-large.html#00:12:05.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And deep learning started, which is at the core, again,
https://karpathy.ai/lexicap/0013-large.html#00:12:07.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of AlphaGo and systems like autonomous driving
https://karpathy.ai/lexicap/0013-large.html#00:12:12.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
systems for cars, like the systems that Mobileye,
https://karpathy.ai/lexicap/0013-large.html#00:12:17.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
which is a company started by one of my ex postdocs,
https://karpathy.ai/lexicap/0013-large.html#00:12:22.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Amnon Shashua, did.
https://karpathy.ai/lexicap/0013-large.html#00:12:25.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So that is at the core of those things.
https://karpathy.ai/lexicap/0013-large.html#00:12:28.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And deep learning, really, the initial ideas
https://karpathy.ai/lexicap/0013-large.html#00:12:30.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in terms of the architecture of these layered
https://karpathy.ai/lexicap/0013-large.html#00:12:34.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
hierarchical networks started with work of Torsten Wiesel
https://karpathy.ai/lexicap/0013-large.html#00:12:37.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and David Hubel at Harvard up the river in the 60s.
https://karpathy.ai/lexicap/0013-large.html#00:12:43.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So recent history suggests that neuroscience played a big role
https://karpathy.ai/lexicap/0013-large.html#00:12:47.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in these breakthroughs.
https://karpathy.ai/lexicap/0013-large.html#00:12:53.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
My personal bet is that there is a good chance they continue
https://karpathy.ai/lexicap/0013-large.html#00:12:54.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to play a big role.
https://karpathy.ai/lexicap/0013-large.html#00:12:58.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Maybe not in all the future breakthroughs,
https://karpathy.ai/lexicap/0013-large.html#00:12:59.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
but in some of them.
https://karpathy.ai/lexicap/0013-large.html#00:13:01.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
At least in inspiration.
https://karpathy.ai/lexicap/0013-large.html#00:13:03.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
At least in inspiration, absolutely, yes.
https://karpathy.ai/lexicap/0013-large.html#00:13:05.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So you studied both artificial and biological neural networks.
https://karpathy.ai/lexicap/0013-large.html#00:13:07.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
You said these mechanisms that underlie deep learning
https://karpathy.ai/lexicap/0013-large.html#00:13:12.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and reinforcement learning.
https://karpathy.ai/lexicap/0013-large.html#00:13:17.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But there is nevertheless significant differences
https://karpathy.ai/lexicap/0013-large.html#00:13:19.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
between biological and artificial neural networks
https://karpathy.ai/lexicap/0013-large.html#00:13:23.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
as they stand now.
https://karpathy.ai/lexicap/0013-large.html#00:13:26.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So between the two, what do you find
https://karpathy.ai/lexicap/0013-large.html#00:13:27.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is the most interesting, mysterious, maybe even
https://karpathy.ai/lexicap/0013-large.html#00:13:30.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
beautiful difference as it currently
https://karpathy.ai/lexicap/0013-large.html#00:13:33.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
stands in our understanding?
https://karpathy.ai/lexicap/0013-large.html#00:13:35.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I must confess that until recently, I
https://karpathy.ai/lexicap/0013-large.html#00:13:37.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
found that the artificial networks, too simplistic
https://karpathy.ai/lexicap/0013-large.html#00:13:41.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
relative to real neural networks.
https://karpathy.ai/lexicap/0013-large.html#00:13:46.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But recently, I've been starting to think that, yes,
https://karpathy.ai/lexicap/0013-large.html#00:13:49.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
there is a very big simplification of what
https://karpathy.ai/lexicap/0013-large.html#00:13:53.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
you find in the brain.
https://karpathy.ai/lexicap/0013-large.html#00:13:57.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But on the other hand, they are much closer
https://karpathy.ai/lexicap/0013-large.html#00:13:59.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in terms of the architecture to the brain
https://karpathy.ai/lexicap/0013-large.html#00:14:03.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
than other models that we had, that computer science used
https://karpathy.ai/lexicap/0013-large.html#00:14:07.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
as model of thinking, which were mathematical logics, LISP,
https://karpathy.ai/lexicap/0013-large.html#00:14:11.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Prologue, and those kind of things.
https://karpathy.ai/lexicap/0013-large.html#00:14:16.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So in comparison to those, they're
https://karpathy.ai/lexicap/0013-large.html#00:14:19.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
much closer to the brain.
https://karpathy.ai/lexicap/0013-large.html#00:14:21.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
You have networks of neurons, which
https://karpathy.ai/lexicap/0013-large.html#00:14:23.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is what the brain is about.
https://karpathy.ai/lexicap/0013-large.html#00:14:26.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And the artificial neurons in the models, as I said,
https://karpathy.ai/lexicap/0013-large.html#00:14:27.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
caricature of the biological neurons.
https://karpathy.ai/lexicap/0013-large.html#00:14:32.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But they're still neurons, single units communicating
https://karpathy.ai/lexicap/0013-large.html#00:14:35.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
with other units, something that is absent
https://karpathy.ai/lexicap/0013-large.html#00:14:38.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in the traditional computer type models of mathematics,
https://karpathy.ai/lexicap/0013-large.html#00:14:41.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
reasoning, and so on.
https://karpathy.ai/lexicap/0013-large.html#00:14:48.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So what aspect would you like to see
https://karpathy.ai/lexicap/0013-large.html#00:14:50.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in artificial neural networks added over time
https://karpathy.ai/lexicap/0013-large.html#00:14:53.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
as we try to figure out ways to improve them?
https://karpathy.ai/lexicap/0013-large.html#00:14:57.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So one of the main differences and problems
https://karpathy.ai/lexicap/0013-large.html#00:14:59.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in terms of deep learning today, and it's not only
https://karpathy.ai/lexicap/0013-large.html#00:15:07.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
deep learning, and the brain, is the need for deep learning
https://karpathy.ai/lexicap/0013-large.html#00:15:11.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
techniques to have a lot of labeled examples.
https://karpathy.ai/lexicap/0013-large.html#00:15:16.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
For instance, for ImageNet, you have
https://karpathy.ai/lexicap/0013-large.html#00:15:23.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
like a training set, which is 1 million images, each one
https://karpathy.ai/lexicap/0013-large.html#00:15:24.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
labeled by some human in terms of which object is there.
https://karpathy.ai/lexicap/0013-large.html#00:15:29.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And it's clear that in biology, a baby
https://karpathy.ai/lexicap/0013-large.html#00:15:34.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
may be able to see millions of images
https://karpathy.ai/lexicap/0013-large.html#00:15:42.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in the first years of life, but will not
https://karpathy.ai/lexicap/0013-large.html#00:15:44.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
have millions of labels given to him or her by parents
https://karpathy.ai/lexicap/0013-large.html#00:15:47.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
or caretakers.
https://karpathy.ai/lexicap/0013-large.html#00:15:52.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So how do you solve that?
https://karpathy.ai/lexicap/0013-large.html#00:15:56.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I think there is this interesting challenge
https://karpathy.ai/lexicap/0013-large.html#00:15:59.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that today, deep learning and related techniques
https://karpathy.ai/lexicap/0013-large.html#00:16:03.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
are all about big data, big data meaning
https://karpathy.ai/lexicap/0013-large.html#00:16:08.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
a lot of examples labeled by humans,
https://karpathy.ai/lexicap/0013-large.html#00:16:11.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
whereas in nature, you have this big data
https://karpathy.ai/lexicap/0013-large.html#00:16:18.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is n going to infinity.
https://karpathy.ai/lexicap/0013-large.html#00:16:24.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
That's the best, n meaning labeled data.
https://karpathy.ai/lexicap/0013-large.html#00:16:26.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But I think the biological world is more n going to 1.
https://karpathy.ai/lexicap/0013-large.html#00:16:30.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
A child can learn from a very small number
https://karpathy.ai/lexicap/0013-large.html#00:16:34.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of labeled examples.
https://karpathy.ai/lexicap/0013-large.html#00:16:38.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Like you tell a child, this is a car.
https://karpathy.ai/lexicap/0013-large.html#00:16:42.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
You don't need to say, like in ImageNet, this is a car,
https://karpathy.ai/lexicap/0013-large.html#00:16:44.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
this is a car, this is not a car, this is not a car,
https://karpathy.ai/lexicap/0013-large.html#00:16:48.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
1 million times.
https://karpathy.ai/lexicap/0013-large.html#00:16:51.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And of course, with AlphaGo, or at least the AlphaZero
https://karpathy.ai/lexicap/0013-large.html#00:16:54.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
variants, because the world of Go
https://karpathy.ai/lexicap/0013-large.html#00:16:57.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
is so simplistic that you can actually
https://karpathy.ai/lexicap/0013-large.html#00:17:01.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
learn by yourself through self play,
https://karpathy.ai/lexicap/0013-large.html#00:17:05.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
you can play against each other.
https://karpathy.ai/lexicap/0013-large.html#00:17:06.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
In the real world, the visual system
https://karpathy.ai/lexicap/0013-large.html#00:17:08.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that you've studied extensively is a lot more complicated
https://karpathy.ai/lexicap/0013-large.html#00:17:10.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
than the game of Go.
https://karpathy.ai/lexicap/0013-large.html#00:17:14.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
On the comment about children, which
https://karpathy.ai/lexicap/0013-large.html#00:17:16.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
are fascinatingly good at learning new stuff,
https://karpathy.ai/lexicap/0013-large.html#00:17:18.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
how much of it do you think is hardware,
https://karpathy.ai/lexicap/0013-large.html#00:17:23.000