Latest / Core Memory / A Brain Researcher's Bid to Make Digital Twins of Humans

Transcript

A Brain Researcher's Bid to Make Digital Twins of Humans · Core Memory

0:00
but now imagine that if i had a way to create a model a
0:04
functional model a digital
0:05
SPEAKER_01: twin of that system
0:08
then if I have enough resources and I buy enough GPUs or enough computers,
0:14
SPEAKER_01: I could run what would take 10,000 years.
0:16
SPEAKER_01: I can run it in an hour.
0:18
SPEAKER_01: Simulations.
0:19
SPEAKER_01: Because, yeah, exactly.
0:20
I can just download all the video on the internet and then just
0:24
show it to this digital tweet.
0:26
It's like a digital representation of the mouse's visual cortex
0:31
that's watching the movie.
0:32
SPEAKER_00: Exactly.
0:33
SPEAKER_00: Exactly.
0:47
SPEAKER_00: Hello, everyone, and welcome to another episode of Core Memory.
0:51
SPEAKER_00: Our guest today is Andreas Delias, brain researcher extraordinaire.
0:57
SPEAKER_00: I'll give you a proper intro here in a second.
0:59
I just wanted to welcome people to our new podcasting studio,
1:03
which is in flux.
1:05
SPEAKER_00: It's going to be even more glorious and beautiful in the weeks to come.
1:10
SPEAKER_00: But thank you so much for joining us in person and our new digs.
1:14
SPEAKER_00: Yeah, thank you.
1:14
SPEAKER_00: Thank you for having me.
1:15
SPEAKER_00: And I'm very excited to be your first podcast.
1:18
You also have the honor of being the first guest on a,
1:23
when I read an ad for the podcast.
1:28
SPEAKER_00: We're some very kind people at E1 Ventures,
1:33
SPEAKER_00: a venture capital firm,
1:34
SPEAKER_00: have been,
1:35
SPEAKER_00: are now sponsoring the show.
1:38
SPEAKER_00: And they're a local VC firm.
1:40
SPEAKER_00: So we're not doing, we're not doing, uh,
1:43
SPEAKER_00: supplements or mattresses like every other podcast.
1:46
SPEAKER_00: We're going pure capitalism.
1:48
SPEAKER_00: And there you go.
1:51
SPEAKER_00: Everyone can check out E1 Ventures.
1:52
SPEAKER_00: And now, end of the show we go.
1:54
SPEAKER_00: Andres, I've known you for, I guess, a year, year and a half now.
2:00
SPEAKER_00: You're down the road at Stanford doing brain research.
2:04
SPEAKER_00: Before that,
2:04
I think you did your equivalent of your undergrad studies at
2:08
Cambridge and then at
2:10
SPEAKER_00: least PhD stuff at MIT.
2:13
SPEAKER_00: And you are in this area that has always fascinated me, which is...
2:20
SPEAKER_00: Well,
2:20
SPEAKER_00: you do a lot,
2:21
but one part of it is this idea of actually understanding how
2:27
the brain works,
2:28
SPEAKER_00: trying to map it.
2:31
SPEAKER_00: And then now,
2:32
SPEAKER_00: these days,
2:33
doing a bit of pairing between neuroscience and AI research and
2:40
trying to figure
2:40
SPEAKER_00: out where these fields meet together.
2:42
SPEAKER_01: Yeah, yeah.
2:43
SPEAKER_01: So basically, you know, the brain is...
2:48
you know,
2:48
the example of an intelligent system that we have,
2:52
you know,
2:52
animals,
2:53
humans.
2:54
And,
2:54
you know,
2:55
we're primarily interested to understand how the brain works.
2:57
SPEAKER_01: What are the algorithms of intelligence?
3:00
SPEAKER_01: And it so happens that in the last 10 years or so,
3:04
we have another form of intelligence that is also performing
3:09
and doing very
3:10
SPEAKER_01: sophisticated tasks,
3:11
SPEAKER_01: artificial intelligence.
3:13
SPEAKER_01: that got early inspiration from the brain.
3:17
SPEAKER_01: You know,
3:17
SPEAKER_01: like there was many different attempts over the years,
3:19
SPEAKER_01: over the decades to build AI systems,
3:22
SPEAKER_01: you know,
3:22
SPEAKER_01: symbolic approaches,
3:25
SPEAKER_01: other approaches in computer science departments.
3:28
SPEAKER_01: But what has ultimately succeeded in what we have right now as AI,
3:33
SPEAKER_01: it's artificial neural networks.
3:35
SPEAKER_01: And what is very fascinating to us in neuroscience is that
3:41
SPEAKER_01: Not only we can use the brain for inspiration,
3:45
SPEAKER_01: keep on hopefully developing better AI,
3:48
SPEAKER_01: safer AI,
3:48
SPEAKER_01: smarter AI,
3:50
SPEAKER_01: but also now we can use AI to help us understand the brain.
3:54
SPEAKER_01: So we can use AI for science.
3:56
And what is particularly exciting is that we can build models
4:01
of the brain
4:03
SPEAKER_01: that are in the same substrate as AI.
4:05
SPEAKER_01: So we can compare artificial intelligence with natural intelligence.
4:09
SPEAKER_01: So this is like tight loop from neuroscience to AI,
4:12
SPEAKER_01: AI to neuroscience,
4:13
and we hope this loop will accelerate and enhance both fields
4:17
in the next decade.
4:19
SPEAKER_00: Yeah, I mean, and this is why I get so excited.
4:22
SPEAKER_00: We have,
4:24
SPEAKER_00: like you said,
4:25
SPEAKER_00: you have this collision almost of these two things happening,
4:28
which is new tools and abilities to understand the physical
4:33
brain as we're racing
4:35
SPEAKER_00: to make an artificial intelligence.
4:37
And to your point,
4:38
I mean,
4:38
they called these things neural nets from the outset.
4:41
SPEAKER_00: I used to feel like it was...
4:44
SPEAKER_00: loose and vague at best in terms of representing the brain or being akin,
4:50
but as time has gone on and we see the complexity of these
4:55
neural networks,
4:55
SPEAKER_00: I mean,
4:56
SPEAKER_00: it feels more apt and realistic to me these days.
5:00
SPEAKER_01: Yeah, definitely.
5:03
They are not implementing intelligence at what we call
5:08
implementation level.
5:09
SPEAKER_01: They don't have like,
5:10
SPEAKER_01: you know,
5:11
SPEAKER_01: they're not built from,
5:12
SPEAKER_01: you know,
5:13
the same complexity of neurons that are built in our brain or
5:15
in animals' brains.
5:16
SPEAKER_01: They don't have ion channels.
5:17
SPEAKER_01: They don't have like the biological complexity.
5:20
SPEAKER_01: But they have,
5:22
SPEAKER_01: in many ways,
5:23
they are built by units that essentially integrate information
5:27
from other units.
5:29
SPEAKER_01: They have synapses that are plastic, like our brains do.
5:33
And what is particularly remarkable that has come out in the
5:37
last few years is this
5:39
SPEAKER_01: concept of universality.
5:41
That if you look at an artificial neural network and you take
5:45
different artificial
5:45
neural networks with different architectures and you train them
5:49
on similar tasks,
5:51
They have a lot of similarities in the way they are
5:53
representing information at
5:55
SPEAKER_01: this sort of information processing level.
5:57
SPEAKER_01: So you can think of it like if you can write down,
6:01
SPEAKER_01: let's say an algorithm of how to compute something,
6:04
SPEAKER_01: you can implement it like with biological substrate or with silicon or,
6:08
SPEAKER_01: you know,
6:08
SPEAKER_01: maybe even other kinds of stuff.
6:11
The algorithm,
6:12
the way it's representing information is present information.
6:15
SPEAKER_01: You can ask the question how similar it is.
6:17
SPEAKER_01: So what people have found in the last few years is that there's a lot of
6:20
SPEAKER_01: similarities within networks that have different architectures.
6:25
And also we're finding now and others that there's similarities
6:28
between biological networks,
6:31
SPEAKER_01: the way they present information that's about vision in particular,
6:35
SPEAKER_01: and these artificial neural networks.
6:37
And there's some very nice work that one of the things is
6:39
called the platonic idea.
6:41
SPEAKER_01: So it's basically like,
6:43
SPEAKER_01: it seems that as systems are becoming intelligent,
6:48
SPEAKER_01: they share a lot of similarities,
6:50
SPEAKER_01: even if it's almost like conversion evolution of some sort.
6:53
SPEAKER_01: Yeah.
6:54
But of course,
6:54
there are so many differences,
6:56
and that's where the excitement comes.
6:57
Like,
6:57
what are the similarities and what are the differences between
7:01
these two systems?
7:02
So,
7:03
given all this,
7:04
when you started out your career,
7:06
you know,
7:06
I mean,
7:06
AI...
7:09
SPEAKER_00: As you know,
7:09
SPEAKER_00: I mean,
7:09
SPEAKER_00: it was really,
7:10
we had gone through these fallow periods where people did not
7:15
think this neural net
7:16
SPEAKER_00: approach would work.
7:17
SPEAKER_00: I remember when I would,
7:18
SPEAKER_00: even in,
7:19
SPEAKER_00: I'd say like 2012,
7:20
SPEAKER_00: when it starts lighting up a bit,
7:24
SPEAKER_00: and then even 2015,
7:25
SPEAKER_00: 2016,
7:25
SPEAKER_00: when I would interview neuroscientists,
7:28
SPEAKER_00: you know,
7:28
SPEAKER_00: they were a bit skeptical about AI.
7:31
SPEAKER_00: There wasn't this huge overlap in the fields.
7:34
SPEAKER_00: I know that you've like embraced it and are doing...
7:38
now pioneering work uniting the physical brain in AI,
7:43
but what was your path?
7:45
Did you see any of this coming or you were just purely in the
7:49
neuroscience world?
7:49
SPEAKER_01: No, that's actually a very good question.
7:51
SPEAKER_01: When I started my PhD at MIT,
7:56
SPEAKER_01: You know,
7:57
SPEAKER_01: AI for a lot of years,
8:01
SPEAKER_01: I would say until 12,
8:04
SPEAKER_01: 2015,
8:04
SPEAKER_01: maybe even,
8:05
SPEAKER_01: or even later was like a bad term.
8:07
SPEAKER_01: Like people use the word machine learning.
8:09
It wasn't a very popular thing because of its over promise over
8:13
the years and
8:14
SPEAKER_01: failures and the winters,
8:16
SPEAKER_01: the so-called AI winters.
8:19
And in particular,
8:20
artificial neural networks were not popular at all for many,
8:24
many years.
8:25
SPEAKER_01: It was like there was a few people like Jeff Hinton,
8:28
SPEAKER_01: Jan LeCun,
8:29
SPEAKER_01: Joshua Benjo,
8:30
SPEAKER_01: a few academic labs that were,
8:34
SPEAKER_01: and even those were having a hard time getting funding sometimes,
8:37
SPEAKER_01: that were pioneering this field.
8:39
SPEAKER_01: There was other approaches,
8:40
let's say the mainstream approaches in computer science or in
8:44
computer vision was
8:45
SPEAKER_01: like other approaches.
8:46
SPEAKER_01: So...
8:48
SPEAKER_01: You know,
8:48
SPEAKER_01: of course,
8:49
SPEAKER_01: as being neuroscientists,
8:51
SPEAKER_01: you know,
8:51
SPEAKER_01: like my main interest always has been trying to understand the brain.
8:54
SPEAKER_01: You know, I'm a scientist.
8:56
SPEAKER_01: I want to understand the brain.
8:58
SPEAKER_01: But AI was,
8:59
SPEAKER_01: you know,
9:00
SPEAKER_01: as a way that,
9:02
SPEAKER_01: you know,
9:02
SPEAKER_01: if we understand the brain,
9:04
one way to test our understanding is to build systems that
9:08
mimic the biological behavior,
9:11
SPEAKER_01: right?
9:11
SPEAKER_01: Whether this is human or animal or whatever, right?
9:15
SPEAKER_01: So there was always been,
9:17
SPEAKER_01: as in neuroscience,
9:17
SPEAKER_01: there will always be an interest in artificial neural networks,
9:21
but it wasn't like a mainstream thing from an AI or computer
9:26
vision point of view.
9:28
SPEAKER_01: And then these two things came together,
9:30
SPEAKER_01: as you said,
9:30
SPEAKER_01: after 2010,
9:32
SPEAKER_01: 2012,
9:32
SPEAKER_01: and everybody then switched over very fast.
9:35
SPEAKER_00: And embraced it.
9:36
SPEAKER_00: Do you know this guy, I think his name is Gary Lynch at UC Irvine?
9:41
SPEAKER_00: Yeah, I know the name, but, you know, yeah.
9:43
SPEAKER_00: Okay, he's a neuroscientist I've interviewed over the years.
9:46
SPEAKER_00: There was this book about him called 101 Theory Drive.
9:49
SPEAKER_00: Yeah, I don't know the book, though.
9:51
SPEAKER_00: It's like one of my favorite books of all time.
9:52
SPEAKER_00: I mean,
9:52
SPEAKER_00: this guy is,
9:54
SPEAKER_00: he let this reporter in to his lab,
9:56
SPEAKER_00: just got to hang out for,
9:58
SPEAKER_00: I forget how long,
9:58
SPEAKER_00: a year or two.
9:59
SPEAKER_00: And Gary's one of these sort of eccentric researcher types.
10:05
SPEAKER_00: But, you know, yeah, I just remember calling.
10:07
SPEAKER_00: I got to know him a bit, and then I would call him about...
10:10
SPEAKER_00: And he was just so skeptical.
10:12
SPEAKER_00: It's funny to me now to see how intertwined these two fields are.
10:17
SPEAKER_00: Okay.
10:17
When you started out and you were doing your PhD work and then
10:22
thinking what you
10:23
SPEAKER_00: wanted to do after,
10:24
SPEAKER_00: what was your driving question about the brain?
10:27
SPEAKER_01: Yeah.
10:28
SPEAKER_01: So my driving question was trying to understand perception.
10:35
SPEAKER_01: you know,
10:36
SPEAKER_01: particularly like visual perception,
10:38
SPEAKER_01: perceptual decisions,
10:41
SPEAKER_01: memory learning,
10:42
SPEAKER_01: visual learning,
10:43
SPEAKER_01: how are representations.
10:45
SPEAKER_01: But I was always mostly interested at what we call,
10:48
SPEAKER_01: at least my mainstream,
10:49
SPEAKER_01: although I work at other levels,
10:51
SPEAKER_01: has been at this algorithmic or presentational level.
10:54
SPEAKER_01: You know,
10:54
there's this very famous three levels of analysis that David
10:57
Marr proposed that I
10:58
SPEAKER_01: think are still useful.
10:59
SPEAKER_01: There is this idea that to study an intelligent system,
11:03
SPEAKER_01: you have the behavioral or computational level.
11:06
SPEAKER_01: What are these goals, you know?
11:08
SPEAKER_01: So this is,
11:08
SPEAKER_01: let's say for us,
11:09
SPEAKER_01: perception is very important or,
11:11
SPEAKER_01: you know,
11:12
SPEAKER_01: language,
11:13
SPEAKER_01: or,
11:13
SPEAKER_01: you know,
11:13
SPEAKER_01: these are the things that we do.
11:15
SPEAKER_01: And then there is this algorithmic or representation level,
11:18
SPEAKER_01: which is how you see formation represented,
11:21
SPEAKER_01: transformed,
11:22
SPEAKER_01: manipulated in the brain.
11:24
And then there is the implementation level,
11:26
which is like,
11:26
what are the biophysical mechanisms?
11:29
You know,
11:29
we have neurons,
11:30
we have ion channels,
11:31
genes,
11:31
you know,
11:32
the biological details.
11:34
SPEAKER_01: I was always interested to bridge this level.
11:36
SPEAKER_01: I mean,
11:36
SPEAKER_01: ultimately all levels,
11:37
SPEAKER_01: but in particular,
11:38
SPEAKER_01: so I was a graduate student at this representation algorithmic level,
11:42
SPEAKER_01: the perceptual level or the behavioral level.
11:46
SPEAKER_01: But I would say the reason that right now things are very exciting,
11:52
And that was extremely frustrating for me as a graduate student
11:55
and also as opposed
11:56
over the years is that it's only in the last five years or so
12:03
that we have the
12:04
technology to do large scale data collections at the level that
12:08
at least I'm interested,
12:09
SPEAKER_01: which is the single neuron,
12:11
SPEAKER_01: single spike level,
12:12
or looking at like fast time windows with calcium imaging and
12:16
other stuff.
12:17
SPEAKER_01: Because ultimately,
12:19
SPEAKER_01: like,
12:19
SPEAKER_01: you know,
12:20
SPEAKER_01: science,
12:20
SPEAKER_01: if you look at the history of science,
12:22
SPEAKER_01: we need data,
12:23
SPEAKER_01: you know,
12:23
SPEAKER_01: without data,
12:24
SPEAKER_01: you know,
12:24
SPEAKER_01: we can come up with beautiful theories,
12:26
SPEAKER_01: beautiful hypotheses,
12:27
SPEAKER_01: but even Newton,
12:27
SPEAKER_01: you know,
12:28
SPEAKER_01: relied on data to feed his models,
12:31
SPEAKER_01: right?
12:31
SPEAKER_01: So data has always been the driving force.
12:35
SPEAKER_01: And because the brain is a very complex system,
12:38
SPEAKER_01: It's because of a lot of it because of funding from the brain initially.
12:45
SPEAKER_01: They decided to invest in neurotechnology also about 10 years ago,
12:49
SPEAKER_01: at the time that the AI revolution was happening.
12:51
SPEAKER_00: Right, the U.S.
12:52
SPEAKER_00: government and I think the Europeans made these big bets, right?
12:56
SPEAKER_01: Yeah, the Europeans made a different bet that was more very like...
13:00
SPEAKER_01: bottom-up modeling approach,
13:02
like from the biophysical level to recreate intelligence that
13:05
ended up not being
13:07
SPEAKER_01: very successful for reasons I was criticized from the beginning by many,
13:12
SPEAKER_01: and I think rightly so.
13:13
SPEAKER_01: But the United States government decided to invest in technology,
13:18
SPEAKER_01: to build technology,
13:19
SPEAKER_01: and also other institutions,
13:21
SPEAKER_01: private institutions,
13:23
SPEAKER_01: to build technology that will enable us to collect large-scale data.
13:27
SPEAKER_01: And if you think about AI now, it's a data-hungry enterprise, right?
13:32
SPEAKER_01: Like it's all relying on large data.
13:35
So now we have large data in neuroscience and we have tools in
13:38
AI to build models of the brain.
13:41
SPEAKER_01: And that's only possible because of these two things coming together.
13:45
SPEAKER_00: So like, how did we, how did we even get data from the brain?
13:49
SPEAKER_00: I know we have to put a timeline on it,
13:50
SPEAKER_00: call it like in the eighties,
13:53
SPEAKER_00: nineties,
13:53
SPEAKER_00: early two thousands.
13:55
SPEAKER_01: So,
13:56
SPEAKER_01: you know,
13:57
SPEAKER_01: it's,
13:57
SPEAKER_01: you know,
13:57
SPEAKER_01: I,
13:57
SPEAKER_01: I,
13:58
SPEAKER_01: I like to look at the history of a field to sort of understand,
14:02
SPEAKER_01: you know,
14:03
SPEAKER_01: the complications.
14:04
So if you look at like,
14:06
I would say in the last 80 years or so,
14:10
we relied on
14:13
SPEAKER_01: Measuring the activity of individual neurons.
14:16
SPEAKER_01: And if you look at,
14:17
let's say someone that did their PhD thesis in the nineties or
14:22
in the eighties,
14:23
SPEAKER_01: or even early two thousands,
14:25
SPEAKER_01: they were relying on like getting data from a few hundred neurons maybe.
14:30
SPEAKER_01: And then they would describe the system.
14:32
SPEAKER_00: With just like an electrode?
14:33
SPEAKER_01: Yeah, with an electrode.
14:34
SPEAKER_01: Usually the classical thing,
14:35
SPEAKER_01: let's say Hubel and Weasel,
14:37
people that are famous that got Nobel Prizes by understanding
14:41
early computations in
14:42
SPEAKER_01: the visual system,
14:43
SPEAKER_01: they used basically a single electrode.
14:46
SPEAKER_01: So they would lure a single electrode,
14:48
SPEAKER_01: get close to one cell,
14:50
SPEAKER_01: record its spikes or action potentials.
14:52
SPEAKER_01: This is the...
14:54
binary way that neurons communicate information and then try
14:58
and study this and
14:59
SPEAKER_01: they discover stuff that got them a Nobel Prize.
15:01
SPEAKER_01: And there's been a few Nobel prizes like that over the years.
15:03
The most recent one is the grid cells place fields by the
15:06
Mosers and John O'Keefe.
15:07
SPEAKER_01: So these were like mostly based on the analysis of individual neurons.
15:11
And we were lucky enough that even if you recorded from the
15:14
C-few hundred neurons,
15:16
SPEAKER_01: you were starting to understand
15:18
SPEAKER_01: something interpretable about the code,
15:21
SPEAKER_01: because ultimately,
15:21
SPEAKER_01: you know,
15:22
SPEAKER_01: if you think about it,
15:22
what we're trying to do in neuroscience is there's the outside
15:25
world and how we
15:26
SPEAKER_01: interact with it.
15:27
And then there's the inside world,
15:29
which is what's going on inside our brain.
15:32
SPEAKER_01: And we're trying to find correspondences between these two systems.
15:36
And there's no prior reason why we would be able to even find
15:40
any correspondence
15:41
SPEAKER_01: between when you record spikes from individual neurons,
15:44
SPEAKER_01: when there's a hundred billion of them,
15:46
SPEAKER_01: how it will relate to something in the world.
15:49
SPEAKER_01: So that was sort of the paradigm, but it was very slow.
15:53
SPEAKER_01: And also it was,
15:55
SPEAKER_01: a lot of the discoveries were happening by serendipity,
15:57
SPEAKER_01: by accidental discoveries.
15:59
SPEAKER_01: It wasn't like,
15:59
SPEAKER_01: and as much as we like to think,
16:02
SPEAKER_01: oh,
16:02
SPEAKER_01: hypothesis-driven science,
16:03
SPEAKER_01: we all have very smart people who have great ideas.
16:06
And we write papers that sort of,
16:08
often these discoveries were happening accidentally.
16:11
SPEAKER_01: You know,
16:11
SPEAKER_01: and if you read out the history,
16:12
SPEAKER_01: you know,
16:12
SPEAKER_01: the famous scientists will say,
16:14
SPEAKER_01: oh yeah,
16:14
SPEAKER_01: you know,
16:14
SPEAKER_01: I was trying to do this,
16:16
SPEAKER_01: and then one day the experiment went wrong,
16:18
SPEAKER_01: and I noticed this,
16:19
SPEAKER_01: and then I got a Nobel Prize.
16:20
SPEAKER_01: Yeah.
16:21
SPEAKER_01: You know, so we're very data-limited.
16:22
SPEAKER_01: We should all be so lucky.
16:23
SPEAKER_00: Yeah.
16:26
And,
16:26
okay,
16:26
so that's moving along,
16:28
and then you're going to have to help me out here.
16:31
So I've done some reporting on a bit of the modern day and what
16:35
this looks like with,
16:36
SPEAKER_00: in particular,
16:37
SPEAKER_00: this,
16:38
SPEAKER_00: the part that I've covered is called connectomics,
16:40
SPEAKER_00: and I know you overlap with this,
16:42
SPEAKER_00: but...
16:44
SPEAKER_00: So I've been at Harvard at Jeff Lickman's lab.
16:47
SPEAKER_00: And what they do is they take a mouse brain and then,
16:51
or they take a tiny chunk of a human brain that they get from
16:55
like a surgery.
16:56
SPEAKER_00: And then I've seen it.
16:58
SPEAKER_00: I mean, they have this, it's almost like a meat slicer.
17:02
SPEAKER_00: It slices, but it's very precise.
17:04
SPEAKER_00: It slices these tiny, tiny slivers.
17:07
of the brain and then they image these in three dimensions and
17:10
then they stitch
17:11
back all the tiny slivers and create these beautiful images
17:16
where you can you can
17:17
trace the neurons like what else from the era you were
17:21
describing like what are the
17:24
main tools we're using now to get the most resolution yes
17:28
exactly so as i said you
17:30
SPEAKER_01: know there's
17:32
The brain is,
17:33
it's a very interesting,
17:35
neuroscience is a field of research.
17:37
SPEAKER_01: It's not a discipline like biochemistry or chemistry or,
17:41
SPEAKER_01: you know,
17:41
it's an area of research that draws a lot of expertise and
17:46
fields from philosophers
17:47
SPEAKER_01: to molecular biologists,
17:50
SPEAKER_01: to people studying biophysics,
17:52
SPEAKER_01: to...
17:53
SPEAKER_01: And there's different levels to study.
17:58
SPEAKER_01: The connectomics level that you're describing is trying to understand,
18:02
SPEAKER_01: to describe the structure of the brain.
18:04
SPEAKER_01: Like a map.
18:05
SPEAKER_01: Like a map, yeah, at higher resolution.
18:08
SPEAKER_01: So you can, and, you know, previously that field, it's basically anatomy.
18:13
SPEAKER_01: You know, classically in the past, there was anatomy.
18:16
SPEAKER_01: If you go, let's say Cambridge University,
18:19
SPEAKER_01: It's a very old university.
18:20
SPEAKER_01: There's the anatomy department and the physiology department.
18:23
SPEAKER_01: The anatomists were studying structural stuff.
18:27
SPEAKER_01: For example,
18:27
SPEAKER_01: they were slicing the brain and looking,
18:30
SPEAKER_01: using microscopy or other methods or staining.
18:32
SPEAKER_01: And,
18:32
SPEAKER_01: you know,
18:33
the most famous father of that field is Ramón y Cajal that
18:36
discovered that the
18:37
SPEAKER_01: brain is built from individual neurons.
18:39
SPEAKER_01: So he was like looking at basically,
18:42
SPEAKER_01: you know,
18:42
take a tissue and slice it and look at the staining the neurons
18:46
and look at them
18:47
SPEAKER_01: and describe them.
18:48
SPEAKER_01: But of course,
18:49
SPEAKER_01: the dream of everybody is trying to relay what you see structurally,
18:54
SPEAKER_01: again,
18:54
SPEAKER_01: with what is related to behavior,
18:57
SPEAKER_01: to cognition.
18:58
SPEAKER_01: And then there was another field,
18:59
SPEAKER_01: physiologists,
19:01
SPEAKER_01: who,
19:02
SPEAKER_01: and the father of that field is Adrian,
19:05
SPEAKER_01: an English man at Cambridge.
19:07
SPEAKER_01: And he basically discovered,
19:10
SPEAKER_01: you know,
19:11
SPEAKER_01: people knew there was electricity in their brain,
19:12
SPEAKER_01: and he discovered that
19:14
SPEAKER_01: Again,
19:14
SPEAKER_01: accidentally,
19:16
SPEAKER_01: he thought it was an artifact that when neurons,
19:19
the way neurons communicate with each other is through these
19:22
action potentials or spikes.
19:24
SPEAKER_01: This is like almost like a bit like code,
19:26
SPEAKER_01: like there's,
19:26
SPEAKER_01: they're silent and all of a sudden they say beep and they send a signal.
19:30
SPEAKER_01: And that opened the whole field of physiology in the last 100 years.
19:35
SPEAKER_01: So what I am talking about is that in addition to
19:42
the developments in connectomics and anatomy and molecular
19:46
tools to study the brain.
19:47
SPEAKER_01: Like there's also other things like large scale transcriptomic,
19:50
SPEAKER_01: single cell transcriptomic analysis.
19:53
SPEAKER_01: So you take the neurons and you do, you see all the,
19:57
SPEAKER_01: you know, messenger RNA that they're expressing.
19:59
SPEAKER_01: You do large scale transcriptomics and you find the cell types.
20:02
SPEAKER_01: There's been also a lot of development, exciting work done there.
20:06
Then on the,
20:07
you know,
20:07
that's even at the lower level,
20:09
like at the genetic level.
20:10
SPEAKER_01: Then there is the anatomical level or the connectomics level.
20:14
SPEAKER_01: You know, a lot of people develop techniques to do electron microscopy.
20:18
And now there's other methods like with light microscopy,
20:21
expansion microscopy.
20:23
SPEAKER_01: But then the other level is,
20:25
which I think is in many ways is the most exciting for me
20:29
because it's the level
20:30
SPEAKER_01: that is closest to the computation.
20:33
And it's the level where we can compare brains and machines
20:36
directly right now is the,
20:39
Activity level,
20:41
you know,
20:41
because when you take a brain and you slice it,
20:43
it's dead,
20:44
right?
20:44
SPEAKER_01: So you're studying a structure.
20:47
SPEAKER_01: So it's frozen moment in time.
20:48
SPEAKER_01: Yeah, exactly.
20:49
SPEAKER_01: So it's not like computing anything when you're studying it.
20:52
But now this technology to look at the computation live as it's
20:56
happening in the brain.
20:58
SPEAKER_01: And then you're extracting the activity of the brain.
21:03
SPEAKER_01: And then you're trying to relate these activity patterns to...
21:07
SPEAKER_01: the computation.
21:08
SPEAKER_01: And that's happening,
21:09
SPEAKER_01: you know,
21:09
SPEAKER_01: for example,
21:09
SPEAKER_01: if you go and you,
21:11
SPEAKER_01: you know,
21:11
SPEAKER_01: type a thing on chat GPT or a large language model,
21:14
SPEAKER_01: or you use AI on your iPhone,
21:17
SPEAKER_01: it's running a neural network that also generates activity.
21:20
You know,
21:20
there are,
21:20
as we said,
21:21
individual neurons and they have activity.
21:24
And you can now study that activity that the artificial brain
21:27
generates versus a
21:28
SPEAKER_01: real brain,
21:29
compare them directly at this activation level or what we call
21:33
the representation level.
21:34
SPEAKER_01: So that line comes from the school of physiology, basically.
21:38
SPEAKER_01: Okay.
21:39
SPEAKER_00: Yeah.
21:39
SPEAKER_00: So,
21:40
SPEAKER_00: yeah,
21:40
SPEAKER_00: because as long as I've covered connectomics,
21:44
SPEAKER_00: I know it's fascinating,
21:45
SPEAKER_00: but there's always been naysayers that just say,
21:47
SPEAKER_00: well,
21:47
SPEAKER_00: I mean,
21:48
SPEAKER_00: that's great.
21:48
SPEAKER_00: Yes, it's useful to have...
21:51
this ultra detailed picture of the brain will learn more about,
21:55
about the nature of the brain.
21:57
SPEAKER_00: But like,
21:57
SPEAKER_00: unless you see it working,
22:00
SPEAKER_00: we just don't have anything to correlate what we're seeing too.
22:03
And I know that's really at the heart of what you've been doing
22:07
is to get to this correlation.
22:08
And I know so recently,
22:11
I think it was,
22:12
I mean,
22:13
we're recording this in April.
22:15
SPEAKER_00: I think it was, it was,
22:16
SPEAKER_00: I know there was a nature paper,
22:17
SPEAKER_00: but then I read a New York Times story from the beginning of anything.
22:20
SPEAKER_00: Yeah, it's a series of nature papers that came out together back to back.
22:23
SPEAKER_00: Okay, and then I'll try to set up some framework.
22:27
SPEAKER_00: You tell me where I'm going astray.
22:29
SPEAKER_00: But I think it was like in the 1980s,
22:32
SPEAKER_00: maybe the mid-80s,
22:33
SPEAKER_00: that we did make this map of a worm brain.
22:38
SPEAKER_00: And so this was kind of like this first stab at connectomics.
22:41
SPEAKER_00: And then we've gone all this way until last year.
22:46
SPEAKER_00: We had a full map of the fly brain.
22:49
It was kind of crazy that you go this huge length of time to go
22:53
from the worm to the fly.
22:55
SPEAKER_00: And then the next missions have been the mouse and the human.
22:58
And the thought is that getting a full map of a human brain is
23:05
extremely daunting.
23:07
because of the amount of data and the incredible amount of
23:09
wiring inside the
23:11
SPEAKER_00: complexity of the system.
23:12
SPEAKER_00: And then to start doing stuff where you're watching that brain,
23:17
SPEAKER_00: you know,
23:18
SPEAKER_00: it gets extra daunting as well.
23:21
But so you guys,
23:22
you just came out with this paper where I think it was on mice.
23:25
SPEAKER_00: Yeah.
23:26
SPEAKER_00: And you did the series of papers and you put electrodes into mice brains.
23:31
SPEAKER_00: It was imaging.
23:33
SPEAKER_00: It was all imaging for this part.
23:35
SPEAKER_00: Okay.
23:35
SPEAKER_00: Yeah.
23:35
SPEAKER_01: It was imaging, but it was basically combining.
23:40
SPEAKER_01: So what is unique, this is called the Microns Project.
23:42
SPEAKER_01: It was funded by IARPA as part of the BRAIN Initiative and then NIH, too.
23:46
SPEAKER_00: How does everyone, you guys always get these acronyms, Microns.
23:50
Yeah,
23:50
Microns,
23:51
it was an acronym that came from IARPA when they formed the program.
23:56
And IARPA basically created,
23:58
they wanted to also push connectomics as a technology.
24:01
SPEAKER_01: So they created this program.
24:03
SPEAKER_01: as a way to,
24:04
SPEAKER_01: you know,
24:04
SPEAKER_01: also put funding for connectomics,
24:06
SPEAKER_01: but also combining structure and function.
24:09
SPEAKER_00: I just always wanted...
24:11
SPEAKER_00: How much time gets spent as people shape the acronym?
24:14
SPEAKER_01: Yeah, well, I don't know.
24:15
SPEAKER_01: The acronym wasn't made by us.
24:17
SPEAKER_01: It was made by the program officers at IARPA at the time.
24:21
SPEAKER_00: I'm always impressed.
24:23
SPEAKER_00: Okay, so sorry, sorry.
24:24
SPEAKER_00: So tell me the... And IARPA is like the... It's like DARPA.
24:29
SPEAKER_00: Which is the R&D arm of the Defense Department,
24:31
SPEAKER_00: and IARPA does weird...
24:33
SPEAKER_01: It's more of the Intelligence Department,
24:34
SPEAKER_01: but,
24:35
SPEAKER_01: you know,
24:35
SPEAKER_01: these are...
24:35
SPEAKER_01: They also fund basic science and basic research,
24:38
SPEAKER_01: and this was...
24:39
SPEAKER_01: Part of the basic science funding.
24:43
SPEAKER_01: And I think,
24:43
SPEAKER_01: you know,
24:44
SPEAKER_01: when the Brain Initiative was announced,
24:46
SPEAKER_01: a bunch of organizations,
24:47
SPEAKER_01: NIH,
24:48
SPEAKER_01: IARPA,
24:49
SPEAKER_01: DARPA,
24:50
SPEAKER_01: they contributed to its mission.
24:53
And this was an initial IARPA funded project started by Jacob
24:56
Vogelstein and David
24:57
SPEAKER_01: Markowitz at the time.
24:58
SPEAKER_00: So like the U.S.
24:59
SPEAKER_00: government wants to fund basic research.
25:01
SPEAKER_00: I mean, they want to help us understand the nature of the brain.
25:04
And then maybe they do interesting things with that knowledge
25:07
at some point.
25:08
SPEAKER_00: But we'll give them, you know, there's a purity on one level.
25:12
SPEAKER_01: Yeah, yeah.
25:13
SPEAKER_01: Well, also, you know, talk about funding.
25:15
SPEAKER_01: Like a lot of them,
25:16
SPEAKER_01: you know,
25:18
SPEAKER_01: when we say,
25:18
SPEAKER_01: you know,
25:19
SPEAKER_01: a lot of the stuff that we as society and companies,
25:24
SPEAKER_01: let's say the companies are big tech here in the U.S.
25:26
SPEAKER_01: and around the world,
25:28
SPEAKER_01: benefit from is originally government funding to risking this stuff.
25:33
SPEAKER_01: You know,
25:33
SPEAKER_01: whether this is truck discoveries or it's usually like,
25:36
SPEAKER_01: you know,
25:36
SPEAKER_01: companies,
25:37
SPEAKER_01: they cannot take it too much of a risk because of their,
25:41
SPEAKER_01: you know,
25:41
SPEAKER_01: return on the investment or because of their mission.
25:44
So often like the government,
25:46
federal government research organizations in the U.S.
25:49
and around the world,
25:50
like NIH and others,
25:51
they fund basic research and then
25:55
SPEAKER_01: you know,
25:55
SPEAKER_01: people discovered,
25:56
SPEAKER_01: oh,
25:56
SPEAKER_01: you know,
25:57
SPEAKER_01: there is an oncogene or,
25:58
SPEAKER_01: you know,
25:59
SPEAKER_01: and then it reaches a stage that then in order to be commercialized,
26:03
SPEAKER_01: companies take over.
26:04
So our whole society benefits from funding at this basic
26:09
science research.
26:10
I think this is something that we need to sort of emphasize
26:12
that without that,
26:14
SPEAKER_01: we would stall scientific,
26:16
SPEAKER_01: biomedical,
26:16
SPEAKER_01: engineering progress,
26:18
SPEAKER_01: you know.
26:18
SPEAKER_00: No, you're right.
26:19
SPEAKER_00: And there's like a
26:20
enormous list of massive things that have come out of DARPA in particular,
26:25
you know.
26:26
SPEAKER_01: Yeah, exactly.
26:28
SPEAKER_00: In this particular project, the
26:31
SPEAKER_00: On that timeline that I was talking about, about worm, fly, mice.
26:36
So we're not able to do the whole mouse brain yet because of
26:39
the complexity and the
26:40
SPEAKER_00: amount of data.
26:40
SPEAKER_00: But you guys, if I'm not mistaken, you went after the visual cortex.
26:45
SPEAKER_01: It was the visual cortex.
26:46
And what was very unique about that experiment that was not
26:50
true in the fly brain,
26:53
SPEAKER_01: in the previous one,
26:55
SPEAKER_01: or in other electron microscopy studies,
26:57
SPEAKER_01: let's say the human,
26:59
SPEAKER_01: was that for that same tissue that included visual cortex,
27:03
SPEAKER_01: it was a couple of visual cortical areas.
27:07
SPEAKER_01: We also,
27:08
SPEAKER_01: before,
27:09
SPEAKER_01: you know,
27:10
SPEAKER_01: the Allen Institute sliced the brain to do what Jeff Linkman does,
27:15
SPEAKER_01: you know,
27:15
SPEAKER_01: at the Allen Institute,
27:16
SPEAKER_01: they did the electron microscopy at the Allen Institute.
27:19
and create this beautiful data that then Princeton,
27:22
Sebastian Sangat,
27:23
Princeton like segmented.
27:25
We,
27:26
back then I was still in Houston,
27:28
in Texas at Baylor College of Medicine.
27:30
SPEAKER_01: We did high resolution to photon functional imaging.
27:36
SPEAKER_01: When the mouse was alive and it was watching movies and stuff like that.
27:39
SPEAKER_01: In fact, one of the movies we show was metrics.
27:42
SPEAKER_01: And I think some people are fascinated about that.
27:44
SPEAKER_01: We showed the metrics.
27:46
SPEAKER_00: I want to talk about
27:46
SPEAKER_00: the movie part, but keep going for a second.
27:48
SPEAKER_01: Yeah, so basically we, so we basically took this mouse.
27:51
SPEAKER_01: It was live,
27:53
SPEAKER_01: he was watching movies,
27:54
SPEAKER_01: we're studying his brain,
27:55
SPEAKER_01: you know,
27:56
SPEAKER_01: activity in the brain as he was watching the movies.
27:59
SPEAKER_01: And then for that same brain,
28:02
Then the Allen Institute took it and did the high-resolution electromicroscopy.
28:07
SPEAKER_01: And then we aligned these data together, and that's very unique.
28:10
SPEAKER_01: This is the only... Yeah.
28:12
SPEAKER_00: Okay, wait.
28:13
SPEAKER_00: So there's one group of living mice watching movies.
28:17
SPEAKER_00: Yeah.
28:17
SPEAKER_00: And then there's another where the brain did get frozen in time.
28:21
SPEAKER_01: No, it's one mouse.
28:23
SPEAKER_00: It's the same mouse.
28:24
SPEAKER_01: Ultimately, it was one mouse.
28:25
SPEAKER_01: We imaged, let's say, a few mice.
28:28
SPEAKER_01: And then one of the mice that had the best data...
28:31
SPEAKER_01: was selected because it was a very expensive experiment to slice it.
28:35
SPEAKER_01: Oh, I see.
28:35
SPEAKER_01: So it was exactly that mouse.
28:37
SPEAKER_01: We actually shipped the mouse live to Seattle from Houston.
28:41
SPEAKER_01: This is like a prized mouse.
28:43
SPEAKER_01: Yeah, it was a prized mouse.
28:44
SPEAKER_01: The mouse had, you know, a private jet.
28:47
SPEAKER_01: No, it wasn't a private jet.
28:50
SPEAKER_01: Yeah.
28:50
SPEAKER_01: He flew business.
28:51
SPEAKER_01: No, no, no, no.
28:53
SPEAKER_00: We,
28:53
SPEAKER_00: you know,
28:55
SPEAKER_01: we flew it like,
28:56
SPEAKER_01: you know,
28:57
SPEAKER_01: we took,
28:57
SPEAKER_01: you know,
28:57
SPEAKER_01: we flew it to Seattle and then this mouse was,
29:02
SPEAKER_01: um,
29:03
SPEAKER_01: you know,
29:04
SPEAKER_01: it,
29:05
Basically,
29:05
they performed in that same mouse that we had the functional data.
29:08
SPEAKER_01: I see it.
29:10
SPEAKER_01: They performed the electron microscopy.
29:12
So then it was a matter of like computationally saying this
29:16
neuron that here is how
29:19
its connectivity is or here is where it is in the electron
29:22
microscopy volume.
29:24
SPEAKER_01: In the functional volume, it's here.
29:26
SPEAKER_00: And then we kind of aligned this data together.
29:30
Okay,
29:30
so you gathered all the data with electrodes while the...
29:33
With imaging.
29:33
SPEAKER_00: With imaging, I'm sorry.
29:34
SPEAKER_01: Yeah, but it was measuring activity.
29:37
SPEAKER_00: While it was watching a movie, getting this activity.
29:39
SPEAKER_00: And then the mouse did meet its demise for the last bit of...
29:43
SPEAKER_00: Yeah,
29:44
SPEAKER_01: and then the mouse ultimately went,
29:46
SPEAKER_01: in order to do the anatomy and the connectivity...
29:49
you know,
29:50
the mouse,
29:51
you have to like,
29:52
you know,
29:52
do electron microscopy on the mouse.
29:54
SPEAKER_00: On the brain.
29:55
But then,
29:55
so,
29:55
okay,
29:55
but then you have these two data sets and you try to...
29:57
Exactly.
29:58
SPEAKER_00: Okay.
29:58
SPEAKER_01: So this is very unique.
29:59
SPEAKER_01: In the fly,
30:00
SPEAKER_01: for example,
30:01
SPEAKER_01: the previous studies,
30:02
SPEAKER_01: electron microscopy,
30:03
SPEAKER_01: there was just the anatomical data.
30:05
SPEAKER_01: There was no function.
30:06
SPEAKER_01: Okay.
30:06
So it was very unique because now you can say these neurons or
30:11
this neuron does this,
30:14
SPEAKER_01: How does its anatomy look like?
30:16
SPEAKER_01: What does it look like?
30:17
SPEAKER_01: Who is it talking to?
30:19
SPEAKER_01: How is it connected?
30:20
SPEAKER_00: I see.
30:20
SPEAKER_01: So it's a very unique data set because of this stitching.
30:25
SPEAKER_00: Okay.
30:26
SPEAKER_00: I mean, I'm making light of this, but I find it fascinating.
30:29
SPEAKER_00: And I've seen this in person.
30:31
SPEAKER_00: I went to McKenzie Mathis' lab in Switzerland.
30:35
SPEAKER_00: Yeah.
30:36
SPEAKER_00: I watched mice watching video games and movies.
30:39
I mean,
30:39
and they had sort of a glass panel on where their skull used to
30:45
be to gather this data.
30:48
When you talk about a mouse watching a movie,
30:50
I mean,
30:51
you were saying The Matrix.
30:53
SPEAKER_00: To people who have no idea of this field, what does that even look like?
30:57
SPEAKER_00: You just have a mouse in a little thing in front of a screen?
30:59
SPEAKER_01: Yeah, basically the mouse, you know, the mouse is...
31:04
SPEAKER_01: And you can train it to also play video games kind of once or that.
31:08
SPEAKER_01: That's probably something you say.
31:09
SPEAKER_00: It was like a Mario Kart for mice.
31:11
SPEAKER_00: Yeah, exactly.
31:12
SPEAKER_01: Like they're playing a game.
31:14
SPEAKER_01: So in this case,
31:14
SPEAKER_01: the mouse was,
31:16
SPEAKER_01: you know,
31:16
SPEAKER_01: sitting comfortably on a,
31:18
SPEAKER_01: you know,
31:18
SPEAKER_01: mice like to run on a treadmill.
31:19
SPEAKER_01: So it was like on a treadmill running.
31:22
SPEAKER_01: And then there is,
31:24
SPEAKER_01: as you said,
31:25
SPEAKER_01: a little part of its,
31:27
SPEAKER_01: you know,
31:27
SPEAKER_01: there's a little window on top that we can then put a microscope,
31:31
a very powerful microscope that does not just look at the
31:36
structure of the brain,
31:37
SPEAKER_01: but it looks at the activity,
31:39
SPEAKER_01: particular calcium,
31:40
SPEAKER_01: you know,
31:41
SPEAKER_01: neurons have calcium-like.
31:42
SPEAKER_01: So it's kind of fluorescing.
31:43
SPEAKER_01: Yeah, it fluoresces.
31:44
SPEAKER_01: It's actually like GFP,
31:45
SPEAKER_01: like what jellyfish has,
31:47
but it's a specially modified version of that that is sensitive
31:50
to the amount of activity.
31:52
SPEAKER_01: So,
31:52
SPEAKER_01: you know,
31:52
SPEAKER_01: if the neurons basically,
31:53
SPEAKER_01: they're silent,
31:55
SPEAKER_01: they're like dark,
31:56
SPEAKER_01: and if it's activity,
31:58
SPEAKER_01: they become more gray,
31:59
SPEAKER_01: and we measure that.
32:00
SPEAKER_01: And there are ways to measure thousands of them at the same time.
32:04
SPEAKER_01: And the animal was,
32:05
SPEAKER_01: it didn't just watch one movie,
32:06
it watched many different movies to sample the complexity of
32:10
the world that it
32:11
SPEAKER_01: could experience.
32:12
SPEAKER_00: Like it's watching like outdoor...
32:14
SPEAKER_01: Yeah, like he watched like YouTube movies.
32:18
SPEAKER_01: You know, maybe next time we'll show him your podcast.
32:20
SPEAKER_00: Yeah, next time we'll show your show, yeah.
32:23
SPEAKER_01: Actually,
32:23
SPEAKER_01: no,
32:23
your show with the space race would be good because mine are
32:27
very sensitive to motion.
32:28
SPEAKER_01: So anything has a little of action.
32:30
SPEAKER_01: For example, Mad Max, we chose a few movies that...
32:34
We knew we're having a lot of motion energy in there that will
32:38
be driving the
32:38
SPEAKER_01: activity a lot in mice.
32:41
SPEAKER_00: You know, I'm so fascinated by this.
32:42
SPEAKER_00: So that's what you're looking for is the brain reacting.
32:46
SPEAKER_01: Yeah, the brain reacting.
32:47
SPEAKER_01: You know,
32:48
SPEAKER_01: one thing that is fascinating is that,
32:50
SPEAKER_01: and this was not appreciated,
32:52
SPEAKER_01: and it actually at the heart of the field of AI.
32:56
SPEAKER_01: Vision for us humans, it's very easy, right?
33:00
Like it's very difficult for me to sit down and maybe solve
33:02
some math problem or
33:03
think about science or you like have to like spend time setting
33:07
up a business or
33:08
SPEAKER_01: setting the studio,
33:09
SPEAKER_01: how beautiful it is and blah,
33:10
SPEAKER_01: you know,
33:10
SPEAKER_01: all this stuff.
33:11
SPEAKER_01: But like vision is like,
33:12
SPEAKER_01: you know,
33:12
SPEAKER_01: you came,
33:13
SPEAKER_01: you recognize me and said,
33:14
SPEAKER_01: hey,
33:14
SPEAKER_01: Andreas,
33:14
SPEAKER_01: how are you doing?
33:15
SPEAKER_01: You know, we hacked each other, right?
33:18
SPEAKER_01: But if you think about it and people thought this is easy, right?
33:22
And then there was this famous Dartmouth project in the fifties
33:26
where they hired
33:27
SPEAKER_01: some undergrads,
33:29
SPEAKER_01: you know,
33:29
SPEAKER_01: to basically build an artificial visual system.
33:32
And then when they started doing it,
33:34
they realized,
33:34
oh my God,
33:35
this is a really hard problem.
33:37
SPEAKER_01: And then over the years from the fifties until now, we realized that
33:43
SPEAKER_01: In primates,
33:44
SPEAKER_01: and we're very visual creatures,
33:46
SPEAKER_01: like vision is our dominant sense,
33:48
SPEAKER_01: unlike mice and unlike other where there's biblical function and stuff,
33:52
SPEAKER_01: about 30 to 50% of our cortex,
33:55
SPEAKER_01: which is the most
33:57
you know,
33:57
evolved part of our brain is actually processing visual information.
34:01
SPEAKER_01: So it's a very difficult computational problem.
34:05
SPEAKER_01: And nowadays,
34:06
SPEAKER_01: you know,
34:07
SPEAKER_01: if AI excels,
34:08
SPEAKER_01: let's say at large language or even solving math problems,
34:11
SPEAKER_01: but it's still vision recognition problems that it lacks humans behind.
34:18
SPEAKER_00: Just this complexity.
34:19
SPEAKER_00: Because of the complexity.
34:20
SPEAKER_01: And it's a very difficult problem.
34:22
SPEAKER_01: Like,
34:22
SPEAKER_01: you know,
34:23
we do it effortlessly because we've evolved to not think about
34:26
it because we
34:27
SPEAKER_01: wouldn't survive in the savannah if we had to think about it.
34:30
SPEAKER_01: So when mice sit in front of the computer screen,
34:33
SPEAKER_01: they watch movies and these movies are composed of motion,
34:37
SPEAKER_01: objects,
34:38
SPEAKER_01: things moving around.
34:38
And we're trying to understand how is it,
34:41
how are they essentially extracting?
34:44
SPEAKER_01: Because if you think about it, what is the goal of vision?
34:47
SPEAKER_01: It's just photons hitting our retina.
34:49
So you can think of it as some noisy photons hitting the retina
34:53
and the brain has
34:54
SPEAKER_01: to infer or make a prediction of what is the causes of these photons.
35:00
SPEAKER_01: So it's not like,
35:00
SPEAKER_01: you know,
35:00
SPEAKER_01: we open our eyes and we,
35:02
SPEAKER_01: you know,
35:03
SPEAKER_01: we see what's there,
35:04
SPEAKER_01: but the brain has to compute it.
35:06
SPEAKER_01: It's not like it's on the camera.
35:08
SPEAKER_01: Like a translation machine.
35:09
SPEAKER_00: Exactly.
35:10
SPEAKER_01: It's not like,
35:10
SPEAKER_01: you know,
35:11
SPEAKER_01: so basically it's like a noisy signal comes in,
35:14
the brain has to figure out what has caused it and that creates
35:17
our presence.
35:18
SPEAKER_01: And that's how we can also dream, right?
35:20
We can,
35:20
the brain can,
35:21
without an input,
35:22
generate things that look sort of realistic.
35:26
And when you're gathering this data,
35:27
I mean,
35:28
how specific does it have to be?
35:29
Has somebody watched the matrix and cataloged what's happening
35:34
second by second and
35:35
SPEAKER_00: matching that to the data?
35:36
SPEAKER_00: Yeah.
35:36
SPEAKER_01: Basically the way you do it is the,
35:39
SPEAKER_01: you know,
35:40
SPEAKER_01: when you build this equipment to do the recordings,
35:43
SPEAKER_01: everything is synchronized.
35:44
SPEAKER_01: Like, so we, when we show something on the monitor.
35:48
The data acquisition that we're acquiring the data with the
35:51
imaging system,
35:52
SPEAKER_01: they have to be synchronized extremely precisely.
35:55
SPEAKER_01: So we know when,
35:57
SPEAKER_01: let's say this scene appeared or this frame on the movie,
36:03
SPEAKER_01: it corresponded to this activity pattern.
36:05
SPEAKER_01: When it changed to this, it corresponded to this.
36:07
SPEAKER_01: So these things have to be perfectly synchronized.
36:09
SPEAKER_01: and the behavior of the animal.
36:11
SPEAKER_01: So,
36:11
SPEAKER_01: for example,
36:12
if this scene appeared and the animal started moving its eyes
36:16
or it started running,
36:17
we need to have everything measured very precisely to make
36:21
these correlations
36:22
SPEAKER_01: between the stimulus,
36:23
SPEAKER_01: the behavior,
36:24
SPEAKER_01: and the activity in the brain.
36:26
SPEAKER_00: And the mouse is kind of,
36:27
SPEAKER_00: okay,
36:28
SPEAKER_00: so it's on this treadmill,
36:29
so it's a little bit sort of fixed in place with this screen in
36:32
front of it and
36:33
SPEAKER_00: then a microscope.
36:34
SPEAKER_00: Yeah.
36:34
SPEAKER_00: Right over its head.
36:35
And then it might watch,
36:38
so this one special mouse,
36:41
like how many movies did it watch?
36:42
SPEAKER_01: He watched, basically the special mouse watched
36:46
You know,
36:47
it was imaged for two to three hours a day for a couple of weeks.
36:52
SPEAKER_01: So it watched like a lot of different movies.
36:56
Like when you're calling out,
36:58
I mean,
36:59
you said Mad Max because of the motion.
37:01
SPEAKER_01: Yeah.
37:02
SPEAKER_00: I mean, could you tell that it likes any kind of movie?
37:05
SPEAKER_01: Well, we don't know what it likes.
37:06
You know,
37:08
we know that the brain of the mouse likes certain movies versus others.
37:12
SPEAKER_01: Okay.
37:13
SPEAKER_01: And we kind of have an idea why,
37:14
SPEAKER_01: you know,
37:15
SPEAKER_01: mice have a,
37:16
SPEAKER_01: unlike us,
37:17
SPEAKER_01: they have a low spatial resolution visual system.
37:20
It's almost like can think they look at the world with a very
37:23
blurred filter.
37:25
SPEAKER_01: So,
37:25
SPEAKER_01: and there are more,
37:26
SPEAKER_01: you know,
37:26
SPEAKER_01: if you think,
37:27
SPEAKER_01: you know,
37:27
SPEAKER_01: we,
37:28
SPEAKER_01: we are very visually strong features,
37:32
SPEAKER_01: you know,
37:33
SPEAKER_01: basically our vision is our strong sense,
37:35
SPEAKER_01: right?
37:35
SPEAKER_01: We see the world through our visual system, mostly.
37:40
SPEAKER_01: And mice,
37:42
SPEAKER_01: you know,
37:42
SPEAKER_01: they use their vision for specific things like to avoid predators or,
37:47
SPEAKER_01: you know,
37:47
SPEAKER_01: and they're very,
37:48
SPEAKER_01: you know,
37:48
SPEAKER_01: as I said,
37:50
SPEAKER_01: they have a low spatial resolution,
37:52
SPEAKER_01: so they cannot like discriminate details,
37:55
SPEAKER_01: but then they are more sensitive to movement.
37:57
So,
37:57
so movies that have a lot of movement and changes in spatial
38:01
temporal dynamics,
38:03
SPEAKER_01: you know, the mouse will respond more to those movies.
38:08
So this was one of the type of movies we showed,
38:11
basically trying to extract these movies.
38:13
SPEAKER_00: Okay, okay.
38:14
SPEAKER_01: Did the special mouse have a name?
38:18
It was called,
38:19
you know,
38:20
different people in the lab gave it different names.
38:22
SPEAKER_01: We used to call it the platinum mouse.
38:24
SPEAKER_01: Okay.
38:25
SPEAKER_01: I think the Allen Institute of Bristol called it the mini mouse, mini.
38:30
SPEAKER_01: And it was just special because it was on task and...
38:34
SPEAKER_00: Well,
38:34
SPEAKER_01: yeah,
38:34
SPEAKER_01: I mean,
38:35
SPEAKER_01: it was basically,
38:36
SPEAKER_01: you know,
38:37
SPEAKER_01: some mice will fall asleep,
38:39
SPEAKER_00: others are like,
38:39
SPEAKER_00: you know,
38:41
SPEAKER_01: this mouse was like,
38:42
SPEAKER_01: it gave very good data,
38:44
SPEAKER_01: you know,
38:44
SPEAKER_01: it was a very alert mouse,
38:46
SPEAKER_01: let's say,
38:47
SPEAKER_01: and it liked the movies maybe or something.
38:50
SPEAKER_01: So,
38:50
SPEAKER_01: and then it was also when the mice went to the Allen,
38:54
there was a combination we had to collectively with the Allen
38:57
Institute and us
38:59
select one mouse that was the winner mouse that would go
39:02
through their electron
39:04
SPEAKER_01: microscopy pipeline because that was...
39:06
SPEAKER_01: Very heroic effort.
39:07
SPEAKER_01: It took many, many months to get the data.
39:09
SPEAKER_01: It was a very expensive project.
39:11
SPEAKER_01: So we had to select one mouse to get to this,
39:14
SPEAKER_01: as you've seen in Jeff's lab,
39:16
SPEAKER_01: to do this thin slicing.
39:18
SPEAKER_00: Okay.
39:19
SPEAKER_00: Okay.
39:19
SPEAKER_00: And I was reading this New York Times...
39:22
story about all this and last night now i've i'm trying to
39:26
refresh myself on on um
39:29
some of the information in there but what was our what was our
39:32
biggest takeaway
39:34
from this yeah what did we learn yeah so there's a it is a
39:37
number of things okay so
39:40
SPEAKER_01: I would categorize them into three main things.
39:44
SPEAKER_01: Okay,
39:44
SPEAKER_01: the first,
39:45
SPEAKER_01: and this was mostly work led by the Allen Institute,
39:49
SPEAKER_01: and it was on the anatomical side.
39:53
It was trying to figure out,
39:54
you know,
39:55
the brain is composed of many cell types.
39:59
SPEAKER_01: And,
39:59
SPEAKER_01: you know,
39:59
SPEAKER_01: for example,
40:00
SPEAKER_01: our blood has red and white blood cells,
40:02
SPEAKER_01: and we know what they are sort of doing.
40:04
The brain,
40:05
we still are trying to figure out what is the logic behind the
40:08
cell type diversity.
40:10
SPEAKER_01: So, I mean, different types of neurons.
40:13
SPEAKER_01: Different types of neurons, yeah.
40:14
Like there are neurons that,
40:15
for example,
40:16
we know there's major two classes.
40:17
There's the excitatory,
40:19
ones that excite each other and ones that inhibit each other.
40:23
And then within each of these big families,
40:25
there's many different types.
40:27
SPEAKER_01: And we knew this since the days of Cajal,
40:30
SPEAKER_01: like 100 years ago,
40:31
SPEAKER_01: they morphologically look different.
40:33
SPEAKER_01: You know,
40:34
SPEAKER_01: neurons have basically the area where they receive information,
40:37
SPEAKER_01: like their antenna is called the dendrite,
40:39
and then they have the region where they send information out
40:41
called the axon.
40:42
SPEAKER_01: And all of these neurons, if you look at them, they're like trees.
40:45
You know,
40:45
a redwood tree looks different than an olive tree or a banana tree.
40:49
SPEAKER_01: They look like that.
40:50
You know,
40:50
like you look at them,
40:51
in fact,
40:51
they do look like trees,
40:53
a lot of them.
40:54
SPEAKER_01: And they look differently and you can classify them.
40:56
So the electron microscopy allows you to do this morphological
40:59
level characterization.
41:02
SPEAKER_01: And then one of the discoveries in this study was like the way,
41:08
SPEAKER_01: you know,
41:08
SPEAKER_01: they're special,
41:09
SPEAKER_01: the way they connect to each other,
41:10
they're not just like randomly connected because you could
41:13
think of them like
41:14
SPEAKER_01: they're different types,
41:15
SPEAKER_01: but maybe it's just random,
41:16
SPEAKER_01: the connectivity.
41:18
SPEAKER_01: Right?
41:18
SPEAKER_01: So what we found,
41:20
SPEAKER_01: or what came out of this project,
41:22
SPEAKER_01: mostly work done at the Allen Institute,
41:24
was that the special rules of who connects to whom based on
41:29
their morphological characteristics.
41:31
SPEAKER_01: And of course,
41:31
SPEAKER_01: this was known before from other studies,
41:33
SPEAKER_01: but it learned new rules and developed a much more detailed,
41:37
SPEAKER_01: let's say,
41:38
SPEAKER_01: code book,
41:38
SPEAKER_01: if you want,
41:39
SPEAKER_01: of who is going to connect to whom at this morphological level.
41:43
SPEAKER_01: So this was one of the discoveries done at the anatomical level.
41:47
SPEAKER_01: And you need that large-scale data to do this.
41:51
Another discovery that was part of the PyGauge that was done
41:56
actually with
41:56
SPEAKER_01: collaboration with the University of Gettyget,
41:59
SPEAKER_01: my collaborator,
42:00
SPEAKER_01: Alex Ecker,
42:01
SPEAKER_01: is these excitatory cells,
42:05
SPEAKER_01: you know,
42:05
SPEAKER_01: the cortex,
42:06
SPEAKER_01: the neocortex,
42:07
SPEAKER_01: has basically six layers,
42:10
SPEAKER_01: and there's five of these layers,
42:11
SPEAKER_01: layer two to six,
42:13
SPEAKER_01: they have excitatory cells.
42:15
SPEAKER_01: There's no excitatory cells in layer one.
42:17
And the idea is like,
42:20
are these distinct types or is it more like a continuum?
42:26
SPEAKER_01: You know,
42:26
SPEAKER_01: because we don't have enough data,
42:28
SPEAKER_01: you know,
42:29
you could think of it like if you are studying redwood trees
42:31
and you go into a
42:32
SPEAKER_01: forest and you see some are very big and some are very small.
42:37
SPEAKER_01: But you say,
42:37
SPEAKER_01: oh,
42:37
SPEAKER_01: you know,
42:37
SPEAKER_01: these two classes,
42:38
SPEAKER_01: these big and small,
42:39
SPEAKER_01: but now it's like,
42:40
SPEAKER_01: well,
42:40
SPEAKER_01: let me like go in to many parts of the country.
42:43
SPEAKER_01: I don't know,
42:44
SPEAKER_01: go to Myrwood,
42:45
SPEAKER_01: go to,
42:46
SPEAKER_01: you know,
42:46
SPEAKER_01: different places in California where I have redwood trees.
42:49
SPEAKER_01: And I'm just collect like hundreds of thousands of them.
42:52
SPEAKER_01: And then if I do some statistical analysis, maybe it's a continuum.
42:57
SPEAKER_01: Maybe it was my sampling that made me think.
42:59
So what Alex and his students found was that if you look at
43:03
this larger scale data,
43:05
SPEAKER_01: Maybe it's more of a continuum.
43:07
SPEAKER_01: And that may change the lot,
43:09
SPEAKER_01: because as I said before,
43:12
SPEAKER_01: you know,
43:12
SPEAKER_01: a lot of the systems like the heart,
43:14
SPEAKER_01: the liver,
43:14
SPEAKER_01: we know the types of cells and we understand the function somehow.
43:18
But here we still don't even understand the logic or the
43:21
categorization of these cells.
43:22
So this was,
43:23
what he found was that maybe it's more of a continuum with
43:26
these excitatory cells.
43:27
SPEAKER_01: And it still needs more research in this direction.
43:31
SPEAKER_01: So this was another anatomical study.
43:33
SPEAKER_01: The other studies,
43:35
SPEAKER_01: And this is what is truly unique about this data, right?
43:38
SPEAKER_01: Again,
43:39
SPEAKER_01: this other data that did electron microscopy in different spaces,
43:42
SPEAKER_01: including humans,
43:42
SPEAKER_01: as you know,
43:43
SPEAKER_01: by jet,
43:43
SPEAKER_01: but this data had function.
43:46
SPEAKER_01: Okay.
43:48
SPEAKER_01: And this is where, you know, the gold is, right?
43:51
SPEAKER_01: This is where the money is, I think.
43:53
SPEAKER_01: It's the function because that's where the computation is happening.
43:56
SPEAKER_01: It's a live tissue that computes.
43:59
SPEAKER_01: So in this mouse, we had both structure and function.
44:02
SPEAKER_01: And for the first time,
44:03
SPEAKER_01: we could study,
44:05
try and figure out what is the relationship between structure
44:10
and function at the
44:11
SPEAKER_01: large scale.
44:12
SPEAKER_01: There's many heroic studies that were done before.
44:15
SPEAKER_01: A lot of them were done by doing very...
44:19
SPEAKER_01: complex and heroic experiments where people would,
44:21
SPEAKER_01: let's say,
44:22
SPEAKER_01: go and use fine glass pipettes and patch two or three cells,
44:28
SPEAKER_01: study its function,
44:30
and then at the same time to slice physiology and figure out if
44:33
they connected or not.
44:35
But here for the first time,
44:36
we had both the anatomy and the physiology in the same mouse.
44:41
SPEAKER_01: So I'll tell you first what we discovered.
44:43
SPEAKER_01: They'll tell you,
44:43
SPEAKER_01: maybe you're interested how this discovery happened,
44:46
SPEAKER_01: which is by building this digital twin of the mouse.
44:49
SPEAKER_01: No, I do want to go into that.
44:51
SPEAKER_01: So basically,
44:53
SPEAKER_01: let's say I'm a neuron and I have some preference,
44:58
SPEAKER_01: what I like in the world to fire.
45:01
In vision,
45:03
you can think of its preference as broken down to what part of
45:06
space I care about.
45:08
SPEAKER_01: Because that's called the receptive field of the neuron.
45:11
SPEAKER_01: So for example,
45:12
SPEAKER_01: if I'm looking straight at you,
45:14
there may be a neuron that likes the upper left corner of my
45:17
visual field.
45:18
SPEAKER_01: So if there's something there,
45:20
SPEAKER_01: it cares about,
45:21
SPEAKER_01: but if there's nothing there,
45:22
SPEAKER_01: it doesn't do anything.
45:24
SPEAKER_01: And then what is that thing that is there that it cares about?
45:27
SPEAKER_01: And just to make it simple, though, mice don't see colors the way we do.
45:31
SPEAKER_01: Let's say it likes red in the upper left visual field,
45:35
but it doesn't like if green is there or yellow is there or
45:39
some other color.
45:40
SPEAKER_01: It just likes something red.
45:41
SPEAKER_01: Okay.
45:43
SPEAKER_01: So the question is, if I am a neuron, who do I connect to?
45:48
Do I connect to neurons that are similar to me,
45:52
completely different than me?
45:54
SPEAKER_01: What is the rule?
45:55
SPEAKER_01: Like, who do I talk to?
45:57
SPEAKER_01: And there is,
45:58
SPEAKER_01: we know from previous studies,
46:00
SPEAKER_01: there's a very,
46:01
SPEAKER_01: you know,
46:01
SPEAKER_01: very well known rule called like to like.
46:04
If I like something,
46:05
I'm more likely to connect with other people that like.
46:07
SPEAKER_01: It's kind of like talking to people, like people have similar interests.
46:11
SPEAKER_01: They're more likely to talk to each other.
46:13
SPEAKER_01: Right.
46:13
SPEAKER_01: So it's kind of the same rule.
46:16
SPEAKER_01: But what this data allowed us to do was to ask the following question.
46:24
There is a very famous rule in neuroscience called Peter's Rule
46:31
that says that
46:33
SPEAKER_01: If I'm a neuron,
46:34
SPEAKER_01: you have to imagine now that these neurons,
46:37
SPEAKER_01: as you've seen in Jeff's lab,
46:39
SPEAKER_01: it's like,
46:40
SPEAKER_01: think of it like you're in a forest and it's very dense.
46:43
Each branch is like,
46:44
it's much denser than any forest you can ever imagine.
46:47
SPEAKER_01: The images are crazy.
46:49
SPEAKER_01: Yeah, exactly.
46:50
SPEAKER_01: It's like,
46:50
SPEAKER_01: it's like each branch is like,
46:52
SPEAKER_01: and each,
46:53
SPEAKER_01: you know,
46:53
SPEAKER_01: leaf is like next to another leaf and it's like,
46:56
SPEAKER_01: there's no,
46:56
SPEAKER_01: you cannot move through that forest.
46:58
SPEAKER_01: It's like, it's like a jungle, like, you know,
47:02
SPEAKER_01: And then the neurons,
47:03
SPEAKER_01: they send,
47:04
SPEAKER_01: you know,
47:05
SPEAKER_01: as I said,
47:05
they have the antenna and then they have this broadcasting area
47:09
called the axon.
47:10
SPEAKER_01: And these axons are very complex.
47:12
And Peter's rule says you send an axon and then if the antenna
47:18
of another neuron is
47:19
SPEAKER_01: close by,
47:20
SPEAKER_01: you flip a coin and you connect with it.
47:23
SPEAKER_01: It's a simple rule, right?
47:24
You don't need a lot of complexity for this because let's say
47:27
you need complexity,
47:29
SPEAKER_01: but you don't need to select who you talk to,
47:33
like based on what maybe they are doing or based on some more
47:37
complicated rules.
47:38
SPEAKER_01: It's sort of like randomly flip a coin.
47:40
SPEAKER_01: There's no method to the madness.
47:41
SPEAKER_00: Yeah,
47:41
SPEAKER_00: it's like,
47:42
and in some ways it's an important rule because maybe you
47:46
create complex networks
47:47
SPEAKER_01: like that,
47:48
SPEAKER_01: or maybe if you want to build AI systems,
47:50
SPEAKER_01: maybe you'll use that rule or something.
47:51
SPEAKER_01: Okay.
47:52
SPEAKER_01: So it is called Peter's rule.
47:55
SPEAKER_01: And what we found,
47:58
SPEAKER_01: because for the first time we had this high resolution anatomical data,
48:04
SPEAKER_01: we could ask the question,
48:05
SPEAKER_01: if I'm a neuron and I like red in the upper left visual field,
48:10
SPEAKER_01: and I send my axon into an area,
48:12
SPEAKER_01: do I randomly flip the coin?
48:16
SPEAKER_01: Because in that area,
48:17
there are also going to be neurons that happen to like the
48:21
upper left visual field,
48:23
SPEAKER_01: but there's some variability.
48:24
SPEAKER_01: Some,
48:25
SPEAKER_01: let's say most of them will be upper left visual field,
48:28
SPEAKER_01: upper left visual field with liking red,
48:33
SPEAKER_01: but there will be some exceptions.
48:34
SPEAKER_01: Some will like green,
48:35
SPEAKER_01: some will be a little bit further down,
48:37
SPEAKER_01: some will like yellow,
48:39
SPEAKER_01: some will like maybe red and green or a combination.
48:42
SPEAKER_01: So what we found was that this Peters rule,
48:45
SPEAKER_01: this flipping of the coin,
48:47
SPEAKER_01: connects neurons based on their spatial relationships.
48:51
SPEAKER_01: And if anything,
48:52
SPEAKER_01: we even found a little bit of a repulsion,
48:54
SPEAKER_01: like they had perfectly overlapping visual field.
48:59
SPEAKER_01: For a very small probability, they will almost avoid connecting.
49:03
SPEAKER_01: But the feature they like, whether they like red, green, I don't know.
49:08
SPEAKER_01: Of course, mice don't like this, but I'm just making this as an example.
49:11
If they like this cup or the Coke,
49:14
they would not do this flipping of the coin.
49:16
There was a probability where they would select the one that
49:20
was more feature-like.
49:22
SPEAKER_01: And, you know, and this may have implications on how brains wire up.
49:28
SPEAKER_01: You know,
49:28
SPEAKER_01: we know that a lot of diseases,
49:29
SPEAKER_01: for example,
49:30
SPEAKER_01: or we think,
49:32
SPEAKER_01: for example,
49:32
SPEAKER_01: a lot of diseases like autism or schizophrenia or,
49:36
SPEAKER_01: you know,
49:39
SPEAKER_01: basically mental disorders, right?
49:41
SPEAKER_00: Like depression.
49:42
SPEAKER_01: Yeah, depression.
49:43
SPEAKER_01: They may be wiring issues, right?
49:46
SPEAKER_01: And it may be these subtle changes in the wiring.
49:49
SPEAKER_01: This rule,
49:50
SPEAKER_01: okay,
49:50
SPEAKER_01: this is a hypothesis,
49:51
SPEAKER_01: a speculation,
49:52
SPEAKER_01: but for example,
49:53
SPEAKER_01: rules like this,
49:54
SPEAKER_01: because this is like a simple rule,
49:56
SPEAKER_01: right?
49:56
SPEAKER_01: You say, I'm going to connect neurons,
49:59
SPEAKER_01: randomly based on what part of space they like,
50:03
SPEAKER_01: but then I'm going to be more selective in what features they like.
50:06
Now you could think maybe that's like more complex rules like
50:09
that or rules like
50:10
SPEAKER_01: that for other domains of cognition.
50:13
SPEAKER_01: So,
50:14
This is one of the studies that tries to understand the wiring
50:18
of the brain and how
50:19
SPEAKER_01: it's related to function at this level of precision.
50:22
And this would not have been possible if we only had the anatomy,
50:25
we couldn't do it.
50:26
SPEAKER_01: If we only had the function, we couldn't do it.
50:28
It was only because this Microns project both put function and
50:33
structure together
50:34
SPEAKER_01: to enable us to study this rule.
50:37
SPEAKER_00: Okay, and this is fascinating.
50:40
SPEAKER_00: I'm keeping an eye on time because I know you have meetings.
50:42
SPEAKER_00: Yeah.
50:44
But I'm going to go off what you said into some,
50:48
I think,
50:49
interesting areas.
50:49
SPEAKER_00: You mentioned digital.
50:50
SPEAKER_00: Yeah.
50:51
SPEAKER_00: So wait for a second.
50:51
SPEAKER_00: But okay, so here we are.
50:53
SPEAKER_00: We're on this moment where...
50:56
On one hand,
50:57
we've got,
50:58
okay,
50:58
so we have all these fascinating microimaging techniques,
51:02
right?
51:02
SPEAKER_00: We have,
51:03
there's companies like E11 that are trying different techniques
51:08
to sort of map out
51:10
SPEAKER_00: the brain with viruses and different things.
51:12
SPEAKER_00: We've got...
51:14
SPEAKER_00: I don't know what,
51:15
like 10 BCI startups putting electrodes into actual humans and
51:21
just funneling out data.
51:22
And then we've got obviously AI able to make sense of patterns
51:28
of data and do all
51:29
SPEAKER_00: this enormous computation.
51:31
SPEAKER_00: So,
51:31
SPEAKER_00: okay,
51:32
SPEAKER_00: so we're kind of like learning a lot about the brain on one hand,
51:36
SPEAKER_00: making artificial brains over here.
51:39
SPEAKER_00: are complementary and overlap.
51:42
When you talked about digital twin,
51:43
I'm not 100% sure if this is where you were going.
51:45
SPEAKER_00: But I mean,
51:47
SPEAKER_00: obviously,
51:47
SPEAKER_00: the holy,
51:49
a holy krail would be some kind of complete representation of
51:55
how the brain works.
51:58
SPEAKER_00: And then some kind of union with computers, maybe.
52:04
SPEAKER_00: I mean, if you're into that, right?
52:05
SPEAKER_00: Or some sort of...
52:07
SPEAKER_00: blonde between these intelligences.
52:09
I mean,
52:09
is that,
52:09
is that like,
52:12
is that where you were going a little bit or is that,
52:14
am I crazy?
52:16
SPEAKER_00: Is this where everybody's coming?
52:17
SPEAKER_00: No, no, no.
52:18
SPEAKER_01: I think in the future we'll be seeing things like that,
52:20
SPEAKER_01: but specifically what we did,
52:23
SPEAKER_01: you know,
52:24
SPEAKER_01: let's say one important stepping stone is the following.
52:29
SPEAKER_01: If you...
52:30
SPEAKER_01: If you could, if you want to study a system, let's see this mouse.
52:35
SPEAKER_01: I said, I don't know, we did experiments.
52:37
SPEAKER_01: I don't know.
52:37
Let's say we show it 10 hours worth of movies or 20 hours worth
52:40
of movies.
52:42
SPEAKER_01: And there's like millions hours worth of movies.
52:44
SPEAKER_01: Right.
52:46
SPEAKER_01: And this mouse,
52:48
SPEAKER_01: you know,
52:48
SPEAKER_01: even if we did experiments,
52:50
SPEAKER_01: even if I did experiments every day of my life,
52:53
there's so many experiments I can do because I'm only going to
52:54
live a certain
52:55
SPEAKER_01: amount of years.
52:56
SPEAKER_01: Right.
52:56
SPEAKER_01: Right.
52:57
SPEAKER_01: Now,
52:58
SPEAKER_01: and because we're doing something in the physical world,
53:01
SPEAKER_01: we're limited by,
53:03
SPEAKER_01: you know,
53:04
SPEAKER_01: our lifetime ultimately,
53:06
SPEAKER_01: right?
53:06
SPEAKER_01: How many experiments we can do.
53:09
SPEAKER_01: But now imagine that if I had a way to create a model,
53:14
SPEAKER_01: a functional model,
53:15
SPEAKER_01: a digital twin of that system.
53:18
Then if I have enough resources and I buy enough GPUs or enough computers,
53:24
SPEAKER_01: I could run what would take 10,000 years.
53:26
SPEAKER_01: I can run it in an hour.
53:28
SPEAKER_01: Simulations.
53:29
SPEAKER_01: Because yeah, exactly.
53:30
SPEAKER_01: I can just download all the video on the internet.
53:33
SPEAKER_01: or, you know, and then just show it to these digital tweets.
53:36
SPEAKER_00: It's like a,
53:37
it's a digital representation of the mouse's visual cortex
53:41
that's watching the movie.
53:42
SPEAKER_00: Exactly.
53:43
SPEAKER_01: So this move,
53:44
SPEAKER_01: basically what we created,
53:45
SPEAKER_01: and this used deep learning,
53:48
SPEAKER_01: it used similar architectures like CharGPT,
53:52
SPEAKER_01: like large scale data and neural networks.
53:55
So what we did is we collected this data and then we trained
53:59
the model,
54:01
not to do
54:03
SPEAKER_01: to be a biophysical model,
54:05
SPEAKER_01: like the human brain project in the European project,
54:07
SPEAKER_01: which,
54:08
SPEAKER_01: you know,
54:09
SPEAKER_01: was criticized because of that level of analysis.
54:12
SPEAKER_01: But the functional model,
54:13
SPEAKER_01: the goal of this model was to say,
54:15
SPEAKER_01: can you predict,
54:17
SPEAKER_01: it was 75,000 neurons in this mouse.
54:20
SPEAKER_01: Can you predict, how well can you predict
54:25
SPEAKER_01: each one of the 75,000 neurons,
54:28
how will they respond to a new movie that I never showed to the
54:34
model to train it with?
54:36
SPEAKER_01: But we did show it to the animal,
54:38
SPEAKER_01: a lot of them,
54:38
SPEAKER_01: so we knew the ground truth,
54:40
SPEAKER_01: so then we could ask the question,
54:41
SPEAKER_01: how well could this model predict?
54:44
SPEAKER_01: And once we built this,
54:46
SPEAKER_01: and it was built using a foundation model approach,
54:49
SPEAKER_01: then we created a good enough model that then we could
54:53
SPEAKER_01: basically showing all kinds of stimuli.
54:55
SPEAKER_01: The mouse is dead now,
54:56
but one could download his digital twin and still run
55:00
experiments and then go back
55:02
SPEAKER_01: to the anatomy and maybe make a discovery.
55:05
SPEAKER_01: Of course, this is still in the digital twin.
55:08
You have to be careful because like CharGPT or other models can hallucinate,
55:12
SPEAKER_01: the digital twin can hallucinate and tell you stuff that may not be true.
55:15
SPEAKER_01: So you have to...
55:16
SPEAKER_01: develop hypotheses and then to maybe new experiments to test it.
55:20
SPEAKER_01: But this is the idea of the digital tweet.
55:23
SPEAKER_01: And it's not, you know, we're not the first to use this term.
55:25
SPEAKER_01: I mean,
55:25
SPEAKER_01: it's used,
55:26
SPEAKER_01: for example,
55:26
SPEAKER_01: in the aircraft industry,
55:28
SPEAKER_01: you know,
55:29
SPEAKER_01: when,
55:29
SPEAKER_01: you know,
55:29
SPEAKER_01: they want to,
55:31
SPEAKER_01: or engineers,
55:32
SPEAKER_01: when they want to optimize the system,
55:34
Often they will create a digital twin of it and they will run
55:36
the optimization in a
55:37
SPEAKER_01: digital twin because they can do it cheaper and not scale.
55:41
SPEAKER_00: Yeah.
55:41
You mean when they take like the,
55:42
all the electronics and just,
55:43
just kind of lay.
55:45
SPEAKER_01: Yeah.
55:46
Or they create a simulation of the,
55:48
of the,
55:48
of the airplane or the wing of the airplane.
55:51
SPEAKER_01: Yeah.
55:51
SPEAKER_01: And then in the simulation,
55:53
SPEAKER_01: they can say,
55:53
SPEAKER_01: okay,
55:53
SPEAKER_01: I'm going to change the wing in this way and see,
55:56
SPEAKER_01: do I get,
55:57
SPEAKER_01: you know,
55:58
SPEAKER_01: so of course,
55:59
SPEAKER_01: ultimately you have to test it in the,
56:01
SPEAKER_01: in the physical system or the biological system.
56:04
SPEAKER_01: And this process of collecting data,
56:06
SPEAKER_01: building digital tweets,
56:08
SPEAKER_01: and then testing the bug is very powerful.
56:10
And this is one of the ways that we're seeing accelerated
56:13
progress of AI being used in science.
56:16
Possibly incredibly stupid question,
56:19
but how do you show a digital twin brain a movie?
56:24
SPEAKER_00: It's just seeing the code?
56:26
SPEAKER_00: Yeah,
56:26
SPEAKER_01: it's basically the input to the digital twin is trained,
56:30
SPEAKER_01: you know,
56:31
SPEAKER_01: like it's a neural network.
56:33
And right now,
56:34
if you open,
56:35
let's say you are a chat box,
56:38
you just type code,
56:39
right?
56:40
SPEAKER_01: But you can also download an image, right?
56:42
SPEAKER_01: And say, what do you see?
56:43
SPEAKER_01: Yeah.
56:44
SPEAKER_01: It's the same type of code.
56:46
SPEAKER_00: It's looking for all the edges.
56:47
SPEAKER_00: Yeah,
56:47
SPEAKER_01: if we download the code and we show it an image,
56:50
SPEAKER_01: but instead of predicting what is the image,
56:53
SPEAKER_01: that's what,
56:54
SPEAKER_01: let's say,
56:54
SPEAKER_01: large language model or a multi-model model does,
56:57
SPEAKER_01: predicts neural activity.
56:58
SPEAKER_01: Okay.
56:59
So it's exactly the same idea,
57:01
but you can think of it,
57:02
its input is visual movies.
57:05
SPEAKER_01: It's movies,
57:06
SPEAKER_01: basically,
57:07
SPEAKER_01: and its output is neural activity instead of being,
57:10
SPEAKER_01: oh,
57:10
SPEAKER_01: this is a carrot or a cucumber or whatever.
57:13
SPEAKER_00: All right.
57:14
SPEAKER_00: This might also sound stupid, but it's always kind of fascinating me.
57:19
SPEAKER_00: I mean,
57:19
SPEAKER_00: so basically the brain is our most,
57:21
SPEAKER_00: I guess all our,
57:22
SPEAKER_00: most of our organs are important and you need them to be alive.
57:26
SPEAKER_00: But I mean, you know, it obviously shapes everything about our organs.
57:30
reality and I think it's sort of safe to argue it's maybe the
57:35
organ we know the least about.
57:37
SPEAKER_00: Yeah, definitely I would say it's safe to say.
57:39
SPEAKER_00: Which is crazy, right?
57:41
I mean,
57:41
but I guess this is because it's our most complex and hardest
57:45
to study for.
57:47
SPEAKER_00: Not just because of its complexity,
57:48
but because it's harder to get access to a lot of brains to do
57:53
this kind of work.
57:54
When you see what's happening in the world now with a lot of
57:59
the technology we just mentioned,
58:01
SPEAKER_00: plus...
58:03
SPEAKER_00: AI, I mean, what do you, what do you think?
58:07
SPEAKER_00: I mean,
58:07
do you think we get to a much deeper fundamental understanding
58:14
of how our mind
58:15
SPEAKER_00: works in,
58:16
SPEAKER_00: in what five,
58:17
SPEAKER_00: 10,
58:17
SPEAKER_00: 15,
58:17
SPEAKER_00: a hundred years?
58:19
SPEAKER_00: That's a really good question.
58:20
I think my hope,
58:23
you know,
58:24
is that I would see in the next,
58:29
you know,
58:30
five to 10 years,
58:33
and maybe even earlier because,
58:35
you know,
58:35
there's early indication for that.
58:38
We're going to start what we call cracking the neural code in a
58:43
way that it would
58:44
be an accelerated cracking of the code because we already like
58:49
know a lot of stuff
58:50
SPEAKER_01: about this code.
58:51
SPEAKER_01: And as I mentioned,
58:52
SPEAKER_01: a lot of people got Nobel prizes because they understand,
58:54
SPEAKER_01: let's say,
58:54
SPEAKER_01: how edges are represented in the world.
58:56
SPEAKER_01: or maybe how spots of light in the retina,
59:00
SPEAKER_01: or how a rat is encoding its location in space,
59:04
SPEAKER_01: right?
59:05
SPEAKER_01: But most of the stuff we really don't know.
59:08
SPEAKER_01: You know,
59:08
SPEAKER_01: how is it,
59:09
SPEAKER_01: as I said,
59:10
SPEAKER_01: again,
59:10
SPEAKER_01: it's the correspondences between the outside world,
59:13
SPEAKER_01: our actions,
59:13
SPEAKER_01: our feelings,
59:14
SPEAKER_01: our thoughts,
59:15
SPEAKER_01: and what's happening inside the brain.
59:18
SPEAKER_01: And I think that in the next, I would say three to 10 years,
59:23
SPEAKER_01: We're going to start seeing an accelerated rate of discovery.
59:28
And this often happens if you look at the history of science or engineering,
59:31
SPEAKER_01: you know,
59:31
SPEAKER_01: it's like you have incremental and then you have an accelerated,
59:34
SPEAKER_01: almost an exponential increase of knowledge.
59:37
And I think a lot of it is going to be done at this information
59:40
processing level.
59:42
SPEAKER_01: And it's going to be a combination of large scale data collection and AI.
59:47
SPEAKER_01: And it's going to help us building these like digital twins,
59:51
SPEAKER_01: accelerate the progress of discovery and understanding the neural code.
59:55
SPEAKER_01: And why, what do I mean by the neural code?
59:57
SPEAKER_01: Just to be clear,
59:59
SPEAKER_01: as one of my colleagues,
1:00:00
SPEAKER_01: Sophia Sandberg likes to call it,
1:00:01
SPEAKER_01: it's like a Rosetta Stone where we're relating,
1:00:05
SPEAKER_01: it's like a translation where we can relate to an image,
1:00:09
SPEAKER_01: to an activity pattern.
1:00:11
SPEAKER_01: And if we can understand this translator,
1:00:15
SPEAKER_01: That means we understand something mechanistically about that code.
1:00:20
SPEAKER_01: You know, it has to do with the interpretability of that code.
1:00:23
SPEAKER_01: And it's also related to things like,
1:00:25
SPEAKER_01: you know,
1:00:27
SPEAKER_01: you hear often this word that the brain,
1:00:30
SPEAKER_01: the artificial neural networks are like a black box.
1:00:32
SPEAKER_01: We don't understand that's why we don't trust them.
1:00:33
And there's AI safety and stuff that,
1:00:35
you know,
1:00:36
we've talked about before,
1:00:37
right?
1:00:38
SPEAKER_01: So if the code is interpretable,
1:00:40
SPEAKER_01: for example,
1:00:41
SPEAKER_01: if you understand how the brain says the truth versus it lies,
1:00:45
SPEAKER_01: I'm just making this as an exaggerated example,
1:00:48
SPEAKER_01: but just to give you.
1:00:49
SPEAKER_01: And if you could build something like an artificial system,
1:00:52
SPEAKER_01: then you could not do jailbreaking,
1:00:54
SPEAKER_01: for example,
1:00:55
SPEAKER_01: because you'll be able to like see it with humans or other AI systems,
1:01:00
SPEAKER_01: look at it and understand the inner workings of it.
1:01:03
SPEAKER_01: So basically,
1:01:04
SPEAKER_01: this is a fundamental thing,
1:01:05
SPEAKER_01: and I think that is one of the things that's going to happen.
1:01:08
SPEAKER_01: Now,
1:01:09
SPEAKER_01: if that happens,
1:01:11
SPEAKER_01: then the next step,
1:01:13
SPEAKER_01: what you said,
1:01:14
SPEAKER_01: is all these PMI companies,
1:01:17
SPEAKER_01: if you think about it,
1:01:18
SPEAKER_01: for us to build an interface to the brain,
1:01:22
We can optimize that interface if we understand the code,
1:01:26
because it's like language.
1:01:27
SPEAKER_01: Like if I want to like,
1:01:28
SPEAKER_01: if I go to a new country and I want to communicate,
1:01:31
SPEAKER_01: if I understand,
1:01:32
SPEAKER_01: if I have a translator,
1:01:34
SPEAKER_01: it'll be easier for me to communicate.
1:01:35
SPEAKER_01: Now,
1:01:35
there are ways I can communicate maybe by hand waving and doing
1:01:39
all kinds of stuff,
1:01:40
SPEAKER_01: but you know,
1:01:41
SPEAKER_01: if I go,
1:01:42
let's say to Tibet and I have someone that speaks Tibetan and
1:01:46
also English or Greek
1:01:48
SPEAKER_01: in my case,
1:01:50
SPEAKER_01: You know, I'll be able to like communicate better, right?
1:01:54
SPEAKER_01: So by cracking this code of communication,
1:01:59
SPEAKER_01: because between behavior and perception and cognition and activity,
1:02:06
SPEAKER_01: Then we can interface a device that's going to have,
1:02:10
SPEAKER_01: let's say,
1:02:11
what I like to call the exocortex or artificial neural network
1:02:14
communicate with it,
1:02:16
whether this is to write information for blindness or to read
1:02:21
out information for,
1:02:23
SPEAKER_01: you know,
1:02:24
SPEAKER_01: driving a robotic arm or ultimately maybe to enhance our cognition,
1:02:27
SPEAKER_01: right?
1:02:28
SPEAKER_01: To hook information.
1:02:29
SPEAKER_01: me up and give me like,
1:02:30
SPEAKER_01: I don't know,
1:02:31
SPEAKER_01: a very large bandwidth of memory that I don't normally have in my brain.
1:02:36
SPEAKER_01: But I think there's, you know, I think ultimately
1:02:41
SPEAKER_01: You know,
1:02:42
if you look at the history of scientific progress and
1:02:44
industrial progress,
1:02:46
we always had big revolutions when we had a scientific
1:02:50
understanding and not a
1:02:51
SPEAKER_01: brute force engineering approach.
1:02:55
You know,
1:02:55
and I think right now we may be more brute force engineering it.
1:02:59
SPEAKER_01: And that's why,
1:03:00
SPEAKER_01: you know,
1:03:00
and I think we'll see a stronger impact if we have a true
1:03:04
scientific understanding
1:03:05
SPEAKER_01: of intelligence.
1:03:06
And I think that's where we'll see a lot of progress in the
1:03:08
next three to 10 years.
1:03:10
SPEAKER_00: All right, last question, just because I would ask you a million more.
1:03:14
We didn't even get into like neuropixels and stuff,
1:03:16
which I really want to.
1:03:17
SPEAKER_00: People can go look that up on the internet, or we'll have you come back.
1:03:21
SPEAKER_00: Yeah, I'll come back for that.
1:03:22
SPEAKER_00: Okay,
1:03:23
SPEAKER_00: so if you're saying three to ten years,
1:03:25
SPEAKER_00: hopefully,
1:03:26
SPEAKER_00: to get this new level of insight,
1:03:29
SPEAKER_00: develop some of this language,
1:03:31
SPEAKER_00: develop some of these bonds between...
1:03:34
SPEAKER_00: our flesh brain and our computer brains.
1:03:37
SPEAKER_00: I mean, you already alluded to it.
1:03:40
I mean,
1:03:41
my sense from you is you do see a cyborg-type future and that
1:03:46
you're not horrified.
1:03:49
SPEAKER_00: Is that true?
1:03:49
SPEAKER_00: What does the world look like in 25 years?
1:03:52
SPEAKER_00: I think...
1:03:54
SPEAKER_01: You know,
1:03:55
it's hard for me to speculate what it would look like in 25
1:03:59
years because,
1:03:59
SPEAKER_01: you know,
1:04:00
SPEAKER_01: like there's all kinds of,
1:04:02
SPEAKER_01: you know,
1:04:02
examples historically where we underestimate and overestimate
1:04:05
at the same time technology.
1:04:07
But I do think that,
1:04:10
you know,
1:04:10
I mean,
1:04:11
in some ways we're already like having iPhone devices.
1:04:14
SPEAKER_01: I mean, if you think about it,
1:04:16
The fact that we communicate with AI and Chagipitina or,
1:04:20
you know,
1:04:20
through our vision.
1:04:21
SPEAKER_01: I mean, our vision is a highly...
1:04:25
SPEAKER_01: camera input is a BMI to drive the brain, right?
1:04:28
SPEAKER_01: So that already exists, the fact we have hands.
1:04:31
So,
1:04:31
you know,
1:04:32
we've evolved to have,
1:04:34
you know,
1:04:34
interfaces to communicate to our neural network.
1:04:37
SPEAKER_01: And now we build external devices like phones and computers,
1:04:41
SPEAKER_01: you know,
1:04:41
SPEAKER_01: so that's gonna accelerate our way to do science.
1:04:46
SPEAKER_01: You know, for example, let's say,
1:04:49
SPEAKER_01: you're going to design this room and optimize it so people feel better.
1:04:53
Or let's say every person that comes on the podcast,
1:04:56
you optimize it for their happiness,
1:04:59
right?
1:05:00
SPEAKER_01: Well, I mean, maybe, you know, you can have a computer help you do that.
1:05:04
SPEAKER_01: And in some ways that's a...
1:05:05
SPEAKER_01: a cyborg, right?
1:05:06
SPEAKER_01: Like you have a human and a machine working side by side.
1:05:11
And then maybe they're committing through language or through a camera,
1:05:14
right?
1:05:15
SPEAKER_01: Now,
1:05:15
SPEAKER_01: whether we're going to have a direct interface to the brain,
1:05:17
SPEAKER_01: that's also going to happen in the future,
1:05:19
SPEAKER_01: I hope.
1:05:22
SPEAKER_01: You do hope.
1:05:22
SPEAKER_01: Yeah, I hope.
1:05:23
SPEAKER_01: And I'm not worried about it,
1:05:25
but I don't think it's something that necessarily will happen
1:05:31
in the next five to
1:05:33
SPEAKER_01: 10 years.
1:05:34
SPEAKER_01: And the reason is because
1:05:36
you know,
1:05:36
there's a lot of human,
1:05:38
you know,
1:05:39
unless,
1:05:39
I mean,
1:05:40
it depends on the device.
1:05:41
SPEAKER_01: The more invasive a device is, it's a trade-off, right?
1:05:44
SPEAKER_01: If there is a non-invasive device that we can all try,
1:05:47
SPEAKER_01: and there are people trying these things,
1:05:48
SPEAKER_01: but they don't work very well right now in a way that is useful.
1:05:51
SPEAKER_01: You know,
1:05:52
SPEAKER_01: people try this EEG to,
1:05:54
SPEAKER_01: I don't know,
1:05:55
SPEAKER_01: control something,
1:05:55
SPEAKER_01: but,
1:05:55
SPEAKER_01: you know,
1:05:56
SPEAKER_01: it's not like worth it to be carrying this thing on your head because,
1:05:59
SPEAKER_01: but,
1:05:59
SPEAKER_01: you know,
1:06:00
SPEAKER_01: so,
1:06:01
SPEAKER_01: but if it's something that is,
1:06:02
SPEAKER_01: the more invasive it is,
1:06:03
SPEAKER_01: the more,
1:06:05
SPEAKER_01: People are not gonna adopted.
1:06:07
SPEAKER_01: when I can just read it on my phone, right?
1:06:10
SPEAKER_01: Like why?
1:06:11
SPEAKER_01: People aren't going to have brain surgery.
1:06:12
SPEAKER_00: Yeah,
1:06:13
they're going to have brain surgery because it's going to take
1:06:15
me a little bit
1:06:15
SPEAKER_01: faster to like,
1:06:17
SPEAKER_01: you know,
1:06:18
SPEAKER_01: read a book.
1:06:19
SPEAKER_01: You know,
1:06:20
SPEAKER_01: because I can have,
1:06:21
SPEAKER_01: if I want to do it faster,
1:06:23
I'll have an agent read the book for me and then summarize it
1:06:26
to me or I'll ask it questions,
1:06:28
SPEAKER_01: right?
1:06:29
SPEAKER_01: But if you could download Kung Fu.
1:06:31
SPEAKER_00: Yeah, exactly.
1:06:32
SPEAKER_00: Then maybe.
1:06:33
SPEAKER_00: Yeah, if I could learn skills and...
1:06:36
SPEAKER_00: So, okay, so you're not horrified by that future.
1:06:39
And I'm not trying to put a date on things because I know all
1:06:45
this stuff is really hard,
1:06:45
SPEAKER_00: but it sounded like you were saying not horrified by it,
1:06:49
SPEAKER_00: kind of like maybe in the next 10 years,
1:06:52
SPEAKER_00: there's like a universe where that happens.
1:06:54
Yeah,
1:06:54
I think we'll see patients,
1:06:57
an accelerated form of brain machine interfaces.
1:06:59
SPEAKER_01: I hope we do.
1:07:01
SPEAKER_01: You know,
1:07:01
SPEAKER_01: because there's been a lot of investment in new technologies,
1:07:05
SPEAKER_01: you know,
1:07:05
SPEAKER_01: to interface to the brain.
1:07:07
SPEAKER_01: And I hope that,
1:07:08
SPEAKER_01: you know,
1:07:09
SPEAKER_01: right now,
1:07:09
let's say the state of the art in many ways that is more
1:07:13
standardized and used,
1:07:15
SPEAKER_01: you know,
1:07:17
SPEAKER_01: You know,
1:07:17
SPEAKER_01: it's like things,
1:07:18
SPEAKER_01: technologies from the 80s,
1:07:20
SPEAKER_01: basically,
1:07:21
SPEAKER_01: including things like debris brain stimulation or the UTI,
1:07:25
SPEAKER_01: you know,
1:07:26
SPEAKER_01: we'll see new forms.
1:07:28
SPEAKER_01: There's been a lot of investment,
1:07:31
SPEAKER_01: not just in industry,
1:07:31
SPEAKER_01: but also,
1:07:32
SPEAKER_01: again,
1:07:32
SPEAKER_01: like DARPA,
1:07:33
SPEAKER_01: NIH has put a lot of investment in this.
1:07:36
SPEAKER_01: So there's a lot of technologies that I think...
1:07:39
SPEAKER_01: they could be optimized and be deployed to human patients,
1:07:43
SPEAKER_01: like from blindness,
1:07:44
SPEAKER_01: for other kinds of prosthetics.
1:07:46
SPEAKER_01: So I hope that in the next 10 years, we'll see a lot of that.
1:07:49
SPEAKER_01: But I think to get an invasive device inside our normal people's brain,
1:07:56
SPEAKER_01: to give them these extra capabilities,
1:07:58
SPEAKER_01: I think that's gonna be longer.
1:08:00
SPEAKER_01: It won't,
1:08:01
SPEAKER_01: you know,
1:08:01
SPEAKER_01: unless there's something traumatic happens,
1:08:04
SPEAKER_01: because,
1:08:04
SPEAKER_01: you know,
1:08:04
SPEAKER_01: like 10 years,
1:08:06
SPEAKER_01: I mean,
1:08:06
SPEAKER_01: if you have to think the last 10 years,
1:08:08
we still have these technologies and they still haven't been
1:08:12
deployed at scale yet.
1:08:13
But I think that's what we're going to see next,
1:08:15
you know,
1:08:16
in patients at the beginning.
1:08:17
SPEAKER_00: Okay.
1:08:18
SPEAKER_00: Can I sneak one last question?
1:08:19
SPEAKER_00: Yeah, sure.
1:08:20
SPEAKER_00: There is a danger that this gets expansive.
1:08:23
SPEAKER_00: So try to keep it.
1:08:24
SPEAKER_00: I'm mindful of your time.
1:08:25
SPEAKER_00: Okay.
1:08:26
SPEAKER_00: But if we learn all the mechanics,
1:08:27
SPEAKER_00: if we learn more of the mechanics of the brain,
1:08:30
SPEAKER_00: oh God,
1:08:31
SPEAKER_00: I could be setting you up for trouble on this one.
1:08:33
SPEAKER_00: Yeah.
1:08:35
do we find things like consciousness you think i mean i know
1:08:40
this could be like a
1:08:41
really abstract thing like do we just do we just find the yeah
1:08:45
like what level of
1:08:46
SPEAKER_00: mechanics of how the brain functions do you think we end up uncovering
1:08:51
SPEAKER_01: Yeah,
1:08:52
SPEAKER_01: I think that,
1:08:53
SPEAKER_01: I mean,
1:08:55
SPEAKER_01: you know,
1:08:55
SPEAKER_01: we all,
1:08:56
SPEAKER_01: as neuroscientists,
1:08:56
SPEAKER_01: we all like to think about consciousness.
1:08:59
SPEAKER_01: But frankly speaking, for me, I never really could.
1:09:04
SPEAKER_01: I mean,
1:09:05
SPEAKER_01: there are interesting experiments and,
1:09:08
SPEAKER_01: you know,
1:09:09
SPEAKER_01: even like stuff that when I was a postdoc,
1:09:11
person I was working with did these very cool experiments with
1:09:15
things like by
1:09:15
SPEAKER_01: stable perception.
1:09:16
I don't know if you know,
1:09:17
like the Necker cube,
1:09:18
like,
1:09:19
or the old young woman.
1:09:21
SPEAKER_01: Like if you show an image,
1:09:23
SPEAKER_01: the same image,
1:09:23
but your perception flips and recorded from neurons in the
1:09:26
brain and found that
1:09:27
SPEAKER_01: some neurons were like,
1:09:29
SPEAKER_01: flipping with the percept of a monkey, but others were like not.
1:09:33
So maybe you can see,
1:09:34
and then Francis Crick at the time was very excited about this.
1:09:37
You know,
1:09:38
so there's very cool experiments,
1:09:39
but I do think that consciousness,
1:09:42
to me,
1:09:44
You know,
1:09:44
I've never,
1:09:45
I think there's very interesting philosophical ideas.
1:09:48
SPEAKER_00: I know, I'm sorry.
1:09:49
SPEAKER_01: Yeah,
1:09:49
SPEAKER_01: no,
1:09:49
SPEAKER_01: but it's never,
1:09:51
SPEAKER_01: you know,
1:09:52
SPEAKER_01: and I'm not sure,
1:09:54
SPEAKER_01: for example,
1:09:54
if you see what is the computational advantage of being
1:09:57
conscious versus a zombie,
1:10:00
I don't think there's even a good argument why we need
1:10:03
consciousness to build intelligence.
1:10:05
SPEAKER_01: So maybe it will just emerge or,
1:10:07
SPEAKER_01: you know,
1:10:08
SPEAKER_01: it will be some sort of emergent property,
1:10:10
SPEAKER_01: or maybe there's something very different about it,
1:10:12
SPEAKER_01: like what Penrose says,
1:10:13
SPEAKER_01: that it's something that we cannot right now even experimentally study.
1:10:18
SPEAKER_01: I don't know, but I'm not as, I'm not worried about it.
1:10:21
SPEAKER_01: You know, I don't think it's a priority issue.
1:10:25
SPEAKER_01: as a human species to understand consciousness.
1:10:29
I think it's a priority to understand intelligence behavior so
1:10:34
that we can
1:10:34
SPEAKER_01: understand the brain and fix it for mental disorders and stuff like that.
1:10:39
SPEAKER_01: then we can build AI systems that are more aligned with us.
1:10:42
SPEAKER_01: They don't end up killing us all eventually.
1:10:46
SPEAKER_01: But I don't worry as much about consciousness right now.
1:10:49
SPEAKER_01: I think about it, but I'm not worried about it.
1:10:51
SPEAKER_00: It doesn't keep me up at night.
1:10:53
SPEAKER_00: This would probably be another multi-hour episode to fully explore that.
1:10:57
SPEAKER_00: Andreas, well, we didn't talk much.
1:11:02
SPEAKER_00: At Stanford, you're working on this Enigma project.
1:11:05
SPEAKER_00: It's a little bit...
1:11:07
SPEAKER_00: We'll come back to that.
1:11:08
SPEAKER_00: Yeah, we'll come back to that.
1:11:09
SPEAKER_00: In the future.
1:11:10
SPEAKER_01: We'll keep it enigmatic right now.
1:11:12
SPEAKER_00: So we'll have you back on the show.
1:11:13
SPEAKER_00: But you've got more research that you're pushing on.
1:11:18
SPEAKER_00: And I mean, I'm so fascinated by your work.
1:11:20
SPEAKER_00: And thank you so much.
1:11:22
for coming to visit,
1:11:23
and yeah,
1:11:25
we'll update all our listeners as you march along.
1:11:29
SPEAKER_00: Yeah, yeah, thank you.
1:11:30
SPEAKER_01: I think it's been great,
1:11:32
SPEAKER_01: you know,
1:11:33
SPEAKER_01: seeing you finally in your new studio,
1:11:35
SPEAKER_01: and it's an honor to be the first guest.
1:11:37
SPEAKER_01: Thank you so much for having me.
1:11:38
Maybe this is when the AIs light up and they can appreciate
1:11:44
this moment we were
1:11:45
SPEAKER_00: getting into all this stuff.
1:11:46
SPEAKER_00: So thank you very much.
1:11:47
SPEAKER_00: Thank you.
1:11:48
SPEAKER_00: And we will keep everyone updated on your work.
1:11:52
SPEAKER_00: Thank you.
1:11:52
SPEAKER_00: Thank you.
1:11:52
SPEAKER_00: All right.
1:11:53
SPEAKER_00: Thanks, Andreas.
1:11:53
SPEAKER_00: Thanks.
1:11:57
SPEAKER_00: I hope everyone enjoyed that episode.
1:12:00
SPEAKER_00: It was sponsored by E1 Ventures.
1:12:04
SPEAKER_00: Very kind people.
1:12:05
SPEAKER_00: Turn to E1 Ventures for all your venture capital needs.
1:12:08
SPEAKER_00: The show was hosted by me, Ashley Vance.
1:12:11
SPEAKER_00: Our guest today was Andreas Tolias.
1:12:15
SPEAKER_00: The show is also produced by me and David Nicholson.
1:12:19
SPEAKER_00: Our theme song is by James Mercer and John Sortwood.
1:12:23
SPEAKER_00: Core memory is edited by the percussive John Sortwood.
1:12:29
SPEAKER_00: See you on the next episode.