0:00

Work management platforms. Ugh!

0:02

Endless onboarding, IT bottlenecks,

0:04

admin requests. But what

0:06

if things were different?

0:08

Monday.com is different. No

0:11

lengthy onboarding. Beautiful reports in

0:13

minutes. Custom workflows, you can build

0:15

on your own. Easy to use

0:17

prompt-free AI. Huh! Turns out you

0:20

can love a work management platform.

0:22

Monday. The first work platform, you'll

0:25

love to work platform. You'll love

0:27

to use. McDonald's meets the Minecraft

0:29

universe with one of six collectibles

0:32

and your choice of a Big

0:34

Mac or ten-piece McNuggets with spicy

0:36

netherflame sauce. Now available with a

0:39

Minecraft movie meal. I participate in

0:41

McDonald's for a limited time, a Minecraft

0:43

movie, only in theaters. Since you're new

0:45

to H&R Block, we'll look at your returns from

0:48

the last three years for any money your last

0:50

guy might have missed. For free. I could

0:52

get money back from last year. You

0:54

could. We'll find any mistakes. Could have

0:56

really used that two years ago

0:58

when I dated that mistake for

1:00

five months. Don't leave money on

1:02

the table. Switch to H&R Block

1:04

and get a free second look

1:06

review. Second look is included at

1:08

no additional cost with the purchase

1:10

of tax preparation. Results vary. All

1:13

tax situations are different. Fees apply

1:15

if you have us file an amended

1:17

return. Hello

1:20

and welcome to the new book's network.

1:22

This is your host, Reekarna Orbert. And

1:24

today I am excited to talk

1:26

to Grace Lindsay, author of the book,

1:29

Models of the Mind, how physics,

1:31

engineering and mathematics have

1:33

shaped our understanding of the brain.

1:35

This is a popular science book,

1:37

which doesn't assume any previous neuroscience

1:40

knowledge of the reader, and at

1:42

the same time it has been

1:44

praised by theoretical neuroscientists and described

1:47

as a great. introduction

1:49

to the field. So thanks a lot

1:51

for race and welcome to the

1:53

show. And please start by

1:56

introducing yourself. Yeah, I'm Grace

1:58

Lindsay. I'm Kirk. I'm currently an assistant

2:01

professor of psychology and data science at

2:03

New York University, but my background and

2:05

training is in computational neuroscience. I did

2:07

my PhD at Columbia at the Center

2:10

for the Records Neuroscience. I published this

2:12

book in 2021. In which phase of

2:14

your career were you that time? Yeah,

2:17

so that's when I was doing a

2:19

postdoc. So I finished my PhD in

2:21

around the end of 2017. And then

2:23

I did a postdoc, I ended up

2:26

at the University College London, which also

2:28

has a center for computational neuroscience called

2:30

the Gatsby Unit. And so I was

2:33

working as a postdoctoral researcher there. And

2:35

then I also was working on this

2:37

book kind of on the side a

2:40

little bit while I was, I started

2:42

it before I took that postdoc position

2:44

and then I finished it while I

2:46

was doing them. And what motivated you

2:49

to start with this book? and video

2:51

all your blogging before. Yeah, so I

2:53

did. I started science communication when I

2:56

was a PhD student. It's really, I

2:58

mean, a lot of the science communication

3:00

that I've done, I would say is

3:03

a little bit selfish in the sense

3:05

that I use it as an opportunity

3:07

to explore ideas that aren't directly relevant

3:09

for the research that I'm doing or

3:12

maybe wouldn't come up directly for my

3:14

own research, but things that I want

3:16

to think about and then I want

3:19

to talk to other people about. And

3:21

so yeah, that started with blogging in

3:23

grad school, and then that led to

3:25

some other writing opportunities. I also had

3:28

a podcast for a while in grad

3:30

school with some other PhD students about

3:32

computational neuroscience, and that was another fun

3:35

opportunity for us to read papers and

3:37

discuss it in a group and really

3:39

kind of discuss it in the sense

3:42

of like, we're students here and we're

3:44

trying to figure everything out and why

3:46

are people doing things the way that

3:48

they're doing and that kind of stuff.

3:51

So yeah, a lot of what I

3:53

did was kind of in that vein.

3:55

And then in the course of doing

3:58

this kind of science communication work, I

4:00

was part of a writers group, a

4:02

science writers group called New Right, that

4:05

started at Columbia University, but then they

4:07

also had a chapter in London. And

4:09

so when I moved from my postdoc,

4:11

I joined that chapter as well. And

4:14

throughout my experience with this writers group,

4:16

I was exposed to the process of

4:18

writing a book and what that looks

4:21

like and people worksho workshoping chapters or

4:23

worksho book proposals, you know, how you

4:25

organize a book. how do you market

4:28

a book or get a contract to

4:30

write a book. And so I saw

4:32

that and I thought it was interesting

4:34

and it seemed like a fun challenge

4:37

to write a book, like a, you

4:39

know, do a full book-coing thing. And

4:41

at first I thought it was something

4:44

I would do later in my career.

4:46

But then a publisher that the publisher

4:48

that I used in Bloomsbury Sigma, they

4:50

had worked with some other people that

4:53

I knew from this writers group and

4:55

the editor there said that they're looking

4:57

for more writers and they prefer scientists

5:00

to write rather than writers who write

5:02

about science. And so I first just

5:04

had a meeting to kind of understand

5:07

more specifically how this process works and

5:09

then it became like, okay, what. would

5:11

you want to write a book about?

5:13

And I had a few different ideas.

5:16

And between me and the editor, we

5:18

kind of settled on this basically just

5:20

overview of computational neuroscience because it's not

5:23

well represented in the popular science literature.

5:25

You can find plenty of books about

5:27

the brain. And you could also find

5:30

books about. kind of free hardcore physics.

5:32

Like people are reading popular science books

5:34

about math and physics and people are

5:36

reading books about the brain, but you're

5:39

never seeing the intersection of people using

5:41

math to understand the brain or how

5:43

the study of the brain has been

5:46

influenced by physics and computer science and

5:48

all of that. So it just seemed

5:50

like a topic area that I obviously

5:53

cared about a lot. I kind of

5:55

came to neuroscience as more an experimental

5:57

neuroscientist from a more biology side and

5:59

then learned about mathematical modeling and so

6:02

I came to appreciate what it offers

6:04

the study of the brain and so

6:06

I was happy to be able to

6:09

share that with people and yeah there's

6:11

just a lot of material to cover

6:13

because it isn't that well represented in

6:15

the popular science literature. Yes I absolutely

6:18

agree that this is quite different from

6:20

the popular neuroscience and psychology books and

6:22

in the very first chapter you explore

6:25

this mathematical modeling in biology that are

6:27

benefits and its limitations. And the chapter

6:29

has the funny title spherical cows. So

6:32

can you please explain to us what

6:34

is a spherical cow and how can

6:36

it be useful? Yeah, so this comes

6:38

from like an old joke that kind

6:41

of makes fun of physicists, which is

6:43

a farmer is having trouble with his

6:45

cow, is not producing any milk, and

6:48

so the farmer goes to the local

6:50

university to see if anyone can help

6:52

them figure out why their cow isn't

6:55

producing any milk, and the physicist thinks

6:57

about it, and all of a sudden

6:59

says the farmer, okay, let's assume a

7:01

spherical cow. And so the idea is

7:04

physicists take... a complex thing like a

7:06

cow and its ability to produce milk,

7:08

and they simplify it. They say, okay,

7:11

well, we're not going to worry about

7:13

all the cow-shaped parts of the cow,

7:15

we're going to assume this is here,

7:18

and then we're going to think about

7:20

it from there. And that is, you

7:22

know, the essence of mathematical modeling in

7:24

the biological sciences that you start with

7:27

this complex thing and this, you know,

7:29

biology that comes with all these details

7:31

and we don't know what they're trying

7:34

to solve. And so this can be

7:36

done, you know, a little bit to

7:38

the extreme, you know, maybe assuming a

7:40

spherical cow isn't the best way to

7:43

solve a problem in the production. But

7:45

when it's done correctly, it can be

7:47

really beneficial because it allows you to

7:50

think through how something works without getting

7:52

distracted by maybe details that don't matter

7:54

or don't matter right now or don't

7:57

matter for the specific research question that

7:59

you have. So yeah, so that title and

8:01

that chapter is really about this idea

8:03

of there's this constant push and pull

8:06

between mathematical models trying to simplify

8:08

things and trying to make them

8:10

kind of more elegant and more

8:12

suited for mathematical description and then

8:14

also the biology kind of worrying

8:16

back and saying, hey, actually you

8:18

can't solve this problem if you

8:20

don't think about this detail in

8:22

that detail and just this balance

8:24

that you have to have. when you're

8:27

trying to build mathematical models

8:29

of something from biology, you

8:31

have to think about really, it's a really

8:33

hard problem. It's a core problem

8:36

of mathematical modeling is what matters.

8:38

How can you represent it in

8:40

equations? How can you decide what

8:42

matters and what doesn't matter and

8:44

convince other people of the same

8:46

thing? That's kind of core to

8:48

the endeavor of mathematical modeling

8:50

in the Biosaio. It's this idea comes

8:52

up a lot of times through the

8:55

book that these models. can be useful

8:57

even if they are obviously

8:59

not true and at the

9:01

same time it's necessary to

9:03

be aware of their limitations.

9:05

One chapter where it is

9:07

phrased very explicitly is this

9:09

cracking the neural code where

9:11

you explore how information theory

9:13

can be applied in neuroscience

9:15

and the end you write

9:17

that it has yielded many

9:19

insights and ideas that's too

9:21

long and the cracks in

9:24

the analogy become visible. Can

9:26

you tell more about this?

9:28

Yeah, as you said, this is a

9:30

theme that shows up throughout the

9:32

book is that in every example

9:34

of math being applied to biology,

9:36

you see the progress that's made

9:39

by simplifying, and then comes the,

9:41

not quite backlash, but comes the

9:43

second round of trying to refine

9:45

and get more details. And in

9:47

information theory in particular, the kind

9:50

of analogy that's used is to

9:52

think of the nerves in the

9:54

nervous system or neurons in the

9:56

brain as being kind of like

9:58

telephone wires like these. communication channels where

10:01

the main goal is to just

10:03

transmit a message and then the

10:05

question is what is the best

10:07

way to transmit that message? What

10:09

can you use in terms of

10:12

how you encode the information and

10:14

how will it be able to

10:16

be transmitted in the most robust

10:18

possible way? And it will be

10:20

able to be decoded accurately. And

10:22

so you can see how that

10:25

can be helpful, that analogy and

10:27

like the math that comes along

10:29

with it can be helpful because

10:31

nerves do need to transmit signals

10:33

and carry information. far you're kind

10:36

of losing the essence of what

10:38

the brain does which is computation

10:40

as well we don't just want

10:42

to transmit signals we want to

10:44

do computation by them we're going

10:47

to change the nature of the

10:49

information so that we can turn

10:51

inputs into notably different outputs and

10:53

so in that case in the

10:55

information theory case the math only

10:58

gets you so far because the

11:00

brain isn't a bunch of telephone

11:02

lines that are just trying to

11:04

transmit a message faithfully. They're trying

11:06

to do other things as well.

11:08

And so you get a little

11:11

bit from that kind of math,

11:13

but then you need to obviously

11:15

combine it with different types of

11:17

math and different approaches to get

11:19

a much fuller picture or to

11:22

answer very different research questions that

11:24

people have. And I would say

11:26

that most readers probably aren't surprised

11:28

if they see a chapter about

11:30

information theory. in a neuroscience book,

11:33

but what might be less obvious

11:35

that by the second chapter, immediately

11:37

after the introduction, is about electricity,

11:39

and this plays a huge part

11:41

in the neurobiology of the brain,

11:44

so can you tell us a

11:46

bit about it? Yeah, so... electricity

11:48

is kind of the way that

11:50

neurons send signals within the cell

11:52

itself and the study of electricity

11:54

in the early study of physiology

11:57

and biology are very intertwined and

11:59

so people were you know discovering

12:01

principles of electricity how to control

12:03

electricity and at the same time

12:05

they're discovering principles of biology and

12:08

how organisms work. And so it's

12:10

also one of the earlier examples

12:12

of math being used to model

12:14

and understand the nervous system. And

12:16

so in the second chapter of

12:19

the book, I talk about the

12:21

leaky integrated fire model, which is

12:23

a kind of just direct analogy

12:25

where people started to think of

12:27

a neuron as equivalent to a

12:30

little electrical circuit, a little RC

12:32

circuit. And that you know, that

12:34

ability to make that connection is

12:36

really fruitful because now you have

12:38

this little thing that you can

12:40

describe mathematically because people had developed

12:43

mathematics to understand electrical circuits, and

12:45

you can use those same terms

12:47

and the same kind of parameters

12:49

to characterize what a neuron is

12:51

doing. And so this was, you

12:54

know, work that got started back

12:56

in the 1700s where people were

12:58

understanding how electricity could create movement

13:00

in animals and be able to

13:02

observe. electricity actually at play in

13:05

the nervous system and then it

13:07

got combined with the advances that

13:09

were happening in understanding the mathematics

13:11

of electricity from physicists and engineers

13:13

and then this was able to

13:15

come together to create these mathematical

13:18

models of neurons that still form

13:20

the basis for how we model

13:22

individual neurons today. We used maybe

13:24

a little bit more elaborate. circuitry

13:26

and we have more data to

13:29

make more accurate models. But really,

13:31

if you want to model the

13:33

way that an individual neuron sends

13:35

its own signal, its spike, you're

13:37

just using these kinds of equivalent

13:40

circuit models, as they're known, where

13:42

you're using the language of electrical

13:44

engineering, basically, to describe the function

13:46

of a neuron, how it takes

13:48

in electrical inputs and produces electrical

13:51

changes in its membrane as a

13:53

result. In this chapter you also

13:55

describe how this regarding the brain

13:57

as an electric circuit helped to

13:59

discover the various ion channels and

14:01

how but the role of different

14:04

ions is in this insigna processing.

14:06

Can you talk a bit about

14:08

this? Yeah, so the way

14:10

that a neuron works in terms

14:12

of kind of electrical signaling is

14:14

that the membrane is able to

14:17

keep positive and negative ions apart.

14:19

And so you get this kind

14:21

of imbalance across the membrane, this

14:23

difference. And basically when a neuron

14:25

gets enough inputs, usually from other

14:27

neurons, that kind of balance gets

14:29

broken down and then restored in

14:31

this really rapid thing called an

14:33

action potential. And so you have

14:35

these ion channels which are able

14:37

to open and close depending on

14:39

a variety of things including if

14:41

another neuron is trying to send

14:43

a signal through neurotransmitters or based

14:45

on the voltage of the neuron

14:47

itself. And so you get this

14:50

pretty elaborate kind of dynamical change

14:52

in ion channels opening which are

14:54

letting different ions with different charges

14:56

in and out of the cell,

14:58

and then they close later on,

15:00

and so you get this big

15:02

kind of swoop of a change

15:04

in the membrane potential, and then

15:06

it goes back to its resting

15:08

state so that the neuron can

15:10

get more input and potentially do

15:12

it all again in order to

15:14

send a signal. I think that

15:16

this is, you know, it's a

15:18

good example especially because it is

15:20

this. multifaceted thing that happens and

15:23

to understand that you need to

15:25

be looking about the dynamics of

15:27

all of these individual ion channels

15:29

and the different things that they're

15:31

doing at different times and the

15:33

way that it impacts the cell

15:35

and how that loops back and

15:37

impacts the ion channels. It's this

15:39

big connected system of all these

15:41

moving parts. And so without the

15:43

mathematics of these electrical circuits and

15:45

our understanding of electrical circuits, it

15:47

seemed like it would be very

15:49

difficult to understand what's happening in

15:51

a neural. It requires... that we

15:53

have these mathematical tools that let

15:56

us describe the system and describe

15:58

it with this level of precision.

16:00

and this ability to precisely state,

16:02

okay, this is how these different

16:04

components interact, and when I piece

16:06

them all together, a lot of

16:08

stuff happens. It gets really crazy,

16:10

and it gets crazy in this

16:12

beautiful way that creates this action

16:14

potential. But yeah, it would be

16:16

really difficult to have to just

16:18

think through that without the mathematical

16:20

tools that support a very fine-grained

16:22

description of what's happening. Exhama isn't

16:24

always obvious, but it's real, and

16:26

so is the relief. from Edgliss.

16:29

After an initial dosing phase of

16:31

16 weeks, about four in 10

16:33

people taking Epgliss achieved its relief

16:35

and clear or almost clear skin,

16:37

and most of those people maintain

16:39

skin that's still more clear at

16:41

one year with monthly dosing. Ebgliss,

16:43

library kizumab, LBKZ, a 250 milligrams,

16:45

a 250 milligram per 2 milliliter

16:47

injection, is a prescription medicine used

16:49

to treat adults and children 12

16:51

years of age and older who

16:53

weigh at least 88. can be

16:55

used with or without topical corticosteroids.

16:57

Don't use if you're allergic to

16:59

epglis. Allergic reactions can occur that

17:02

can be severe. Eye problems can

17:04

occur. Tell your doctor if you

17:06

have new or worsening eye problems.

17:08

You should not receive a live

17:10

vaccine when treated with epglis. Before

17:12

starting, epglis, tell your doctor if

17:14

you have a parasitic infection. Searching

17:16

for real relief. Ask your doctor

17:18

about epglis and visit epglis.lis.l and

17:20

visit epglis.lis.lis.l.lis.lis.com or 1.com or 1.

17:22

Or 1. Or 1. Or 1.

17:24

Hi, this is Javon your blinds.com

17:26

design consultant. Oh wow, a real

17:28

person. Yep, I'm here to help

17:30

with everything from selecting the perfect

17:32

window treatments to... Well, I've got

17:35

a complicated project. No problem. I

17:37

can even help schedule a professional

17:39

measure and install. We can also

17:41

send your samples fast and free.

17:43

Hmm, I just might have to

17:45

do more. Whatever you need. So

17:47

the first room we're looking at

17:49

is for... Shup blinds.com now and

17:51

save up to 45% site wide.

17:54

At one point you

17:56

also say that for

17:58

this kind of complex

18:01

systems we usually don't

18:03

have a good intuition

18:05

and that's right we

18:07

made mathematical models. Do

18:09

you have some threshold on it

18:12

that where does a system

18:14

start to get the E2

18:16

complex? I think it's surprisingly low.

18:18

Things can get complex surprisingly quickly. So

18:21

one example that I always like to

18:23

bring up is just an example of

18:25

two neurons interacting. If one neuron sends

18:27

a positive connection to another and that

18:30

neuron sends a negative or inhibitory connection

18:32

back to it, it can get really

18:34

confusing because it's like, okay, if we...

18:37

give input to the first neuron, it's

18:39

going to drive the second neuron, but

18:41

that neuron is going to shut down

18:43

the first neuron, but then that might

18:46

allow that neuron, you know, it's going

18:48

to stop inhibiting itself, and it's just like

18:50

you're already confused. So basically any time, especially

18:52

there's loops in the system. If there's, you

18:54

know, A impacts B, but then through some

18:56

mechanism B can impact A again, then you're

18:59

already in a situation where it's confusing, I

19:01

think, or where if you're not precise, you

19:03

might end up coming to the wrong conclusion.

19:06

So I think any system with any kind

19:08

of loops or recurrence is going to give

19:10

you problems and then any system where there's

19:12

a lot of parts even if they don't

19:15

have a lot of recurrence that there's just

19:17

a lot of things that you know go

19:19

into a system it's very hard to know

19:21

what's going to happen if you change any

19:24

one of those things. So I do

19:26

think we very quickly get to a

19:28

realm where mathematical models are necessary especially

19:30

in the brain which is full of

19:32

loops, almost exclusively loops, and is very,

19:34

as we would say, high dimensional, it's

19:37

got a lot of parts. It's a

19:39

lot of neurons, and even all of

19:41

those neurons are made up of a

19:43

lot of different sub components, and so

19:45

it just quickly becomes too many parts,

19:47

and they're connected in far too complex

19:49

a way to have a hope of

19:51

really understanding what's going on. If we

19:54

can't be precise, and as soon

19:56

as you're getting precise, then you're

19:58

basically doing mathematics. systems

20:01

getting very complex. Can you tell

20:03

us a bit about the neurocircuit

20:05

that guides the digestion in a

20:07

lobster and also what? How would

20:09

these findings have impacted neuroscience more

20:12

broadly? Yeah, so there's a circuit

20:14

in the lobster gut. that controls

20:16

rhythms of digestion. And you would

20:18

think that this is again simple,

20:20

right? Just like we just want

20:23

something that oscillates, you know, it

20:25

should have a nice rhythm to

20:27

it. But it's been studied particularly

20:29

by Eve Martyr, who does experiments

20:32

and also computational modeling. And there

20:34

are a few kind of. principles

20:36

that probably apply pretty broadly to

20:38

the brain that have been kind

20:40

of highlighted by studying that circuit.

20:43

And so one is this idea

20:45

that this circuit can actually do

20:47

many different things. This one circuit,

20:49

even if you kind of hold

20:51

the connections between the neurons constant

20:54

in some way, you can add

20:56

things like neuromodulators, which change a

20:58

little bit how neurons work or

21:00

how they respond to inputs and

21:02

then you can create, you can

21:05

get this one circuit to create

21:07

completely different rhythms. And so this

21:09

is relevant because, you know, we

21:11

feel like in order to understand

21:14

the brain we need to know

21:16

the connections between neurons, we need

21:18

to know what kind of neurons

21:20

connect to what other neurons and

21:22

what structures they form and we

21:25

definitely do need to know that,

21:27

but it's clear that that's not

21:29

sufficient. to be able to understand

21:31

what a circuit is going to

21:33

do if you just saw a

21:36

diagram of neuron A is connected

21:38

to neuron B and everything else,

21:40

that wouldn't be enough to actually

21:42

know what behavior that circuit would

21:45

be able to produce. So that's

21:47

one thing that this simple lobster

21:49

gut circuit has taught us. Another

21:51

thing is that, so in the

21:53

same way that one circuit can

21:56

produce many behaviors, the same behavior

21:58

can be produced by many different

22:00

circuits. So there's... you know, a

22:02

certain rhythm that it's useful for

22:04

this circuit to produce and different

22:07

individual lobsters can have different neuroconnectivity

22:09

that leads to the same behavior.

22:11

And so this again kind of

22:13

makes this connection between the structure

22:15

of the circuit and the function

22:18

it performs a little bit looser.

22:20

It's not just you know the

22:22

structure, you know the function, multiple

22:24

different structures can produce the same

22:27

function, one structure can produce multiple

22:29

different functions. And so yeah, these

22:31

are things that were best identified

22:33

in a simple circuit where we're

22:35

able to have a lot of

22:38

control and we're able to observe

22:40

kind of everything we need to

22:42

know about it. But there are

22:44

principles that... have to hold for

22:46

the brain as a whole. It's

22:49

demonstrating this very nice elegant way

22:51

both experimentally and through computational modeling

22:53

by focusing on this very simple

22:55

circuit, but it definitely represents principles

22:58

that get, you know, referred to

23:00

throughout neuroscience as kind of a

23:02

cautionary tale when people are trying

23:04

to make some sort of connection

23:06

between structure and function. It's like,

23:09

hey, remember the lobster gut, it's

23:11

not that easy. Indeed, and

23:13

it also puts into a bit

23:15

of different light discovery from parallel

23:17

that you described in a chapter

23:19

before that the brain's computations are

23:22

supposed to be efficient because it

23:24

has evolved and we kind of

23:26

assume that it is a... Efficient

23:28

and this lobster story probably points

23:30

to that's the direction that efficient

23:32

doesn't necessarily mean that it is

23:34

optimized for in some algorithmic manner,

23:36

but more like that it's robust

23:38

enough to have some redundancy in

23:40

it. Can you tell some more

23:42

examples where this redionation flexibility can

23:45

be observed in the brain? Yeah,

23:47

so I think that is the,

23:49

so as you mentioned, the kind

23:51

of reason that people come in

23:53

with this belief that. the brain

23:55

is efficient and anything that it

23:57

does is efficient is because the

23:59

brain is an evolved system. We

24:01

assume there's a lot of pressure

24:03

in terms of, you know, maintaining

24:05

low energy use but still being,

24:08

you know, smart enough to actually

24:10

survive in the world. And so

24:12

we assume that the brain has

24:14

found these optimal ways of representing

24:16

information and carrying out computations and

24:18

all of that. And I mean,

24:20

one, I would just say. that

24:22

itself is complicated, that assumption and

24:24

what we do with that assumption

24:26

is a little complicated because the

24:28

brain is an evolved system so

24:31

yes it's we believe evolution can

24:33

find optimal solutions if given enough

24:35

time or at least can find

24:37

good solutions but it also the

24:39

way that evolution explores the space

24:41

is a little weird it's you

24:43

know it's like what happens if

24:45

you change this gene and then

24:47

something changes in the brain and

24:49

that's good or bad and so

24:51

the solutions that the brain finds

24:54

clearly worked well enough and we're

24:56

able to be produced by the

24:58

way that genetics produces brains, but

25:00

that still makes it pretty hard

25:02

for us to understand why they're

25:04

the best solution because we don't

25:06

know the full history and we

25:08

don't know all the constraints and

25:10

we don't know enough about development

25:12

even to know exactly how different

25:14

genes change behaviors. So yeah, but

25:17

that still is obviously a guiding

25:19

principle that informs a lot of

25:21

the ways that people approach the

25:23

study of the study of the

25:25

brain. So yeah, and then so

25:27

ways in which systems can be

25:29

robust. I think one example that

25:31

people think about or point to

25:33

might be places in the prefrontal

25:35

cortex where you get the brain

25:37

that's kind of an area where

25:40

a lot of different information is

25:42

coming into the brain about sensory

25:44

systems and then it guides motor

25:46

control and that kind of thing.

25:48

And so as a result it's

25:50

usually hard to make sense of

25:52

what neurons in the prefrontal cortex

25:54

are doing because they have these

25:56

responses that are a mix of

25:58

a bunch of different stuff. But

26:00

that kind of mixing up of

26:03

stuff that not having just a

26:05

hard coded way of doing something,

26:07

but actually just kind of creating

26:09

this jumble might actually be a

26:11

pretty kind of in some ways

26:13

efficient way to have a system

26:15

that can do a lot of

26:17

different stuff and can be robust

26:19

to different styles of input, but

26:21

still be able to do the

26:23

same type of computations that it

26:26

needs to do to produce the

26:28

right kind of outputs. And people

26:30

also talk about how. the way

26:32

that the brain represents certain information

26:34

can actually shift over time, but

26:36

we can still perform the task

26:38

the same way. So there's something

26:40

called representational drift, where the one

26:42

neuron on one day is representing

26:44

information some way and that it

26:46

changes over time. And so there's

26:49

all of this kind of noise,

26:51

what looks to us like noise

26:53

in the system, in terms of

26:55

how activity changes over time or,

26:57

you know, the same neuron in

26:59

response to the same stimulus, can

27:01

be different. But these things we

27:03

believe could point to underlying mechanisms

27:05

that actually make the system really

27:07

flexible potentially and have other nice

27:09

properties of being able to respond

27:12

quickly to different inputs. So I

27:14

think we're still though very much

27:16

in the phase of trying to

27:18

figure out. Is this thing that

27:20

we're observing optimal or is it

27:22

just kind of a side effect

27:24

of the fact of what the

27:26

brain had to work with to

27:28

make a system that works well?

27:30

That's always, you know, the, it's

27:32

a big question in neuroscience. What

27:35

do you consider as the actual

27:37

thing the brain is doing and

27:39

then what do you call noise?

27:41

What do you say? This is

27:43

just, you know, some messiness on

27:45

top of the thing that it's

27:47

doing. It's not the primary thing

27:49

that it's doing. you know, big

27:51

breakthroughs can happen when you figure

27:53

out that the noise is actually

27:55

a design feature and that bug

27:58

in the system. This sleep about,

28:00

is this a noise or is

28:02

this something? relevant is especially highlighted

28:04

in the chapter about oscillations where

28:06

you say that it's still

28:08

an ongoing debate. So what is

28:10

the status of this debate

28:13

right now? What do oscillations

28:15

do in the brain or how can

28:17

they be observed? Yeah, so

28:20

yeah, in that chapter, I'm talking

28:22

about excitatory and inhibitory forces in

28:24

the brain. So different neurons either

28:26

make other neurons fire more or

28:29

they reduce their firing. And so...

28:31

this creates this like push-pull effect

28:33

which has some benefits that lets

28:36

the system respond potentially very quickly

28:38

to an incoming input because it

28:40

can just kind of push the

28:43

excitatory side of the network and

28:45

it can respond very quickly. And

28:47

then another effect of having this

28:50

kind of push-pull relationship is that

28:52

you get these oscillations where you

28:54

get kind of a wave of activity

28:56

but then that activity recruits its own

28:58

inhibition and so it shuts down but

29:01

then another wave starts and so you

29:03

get oscillations that can be observed in

29:05

many different forms across different brain regions.

29:08

Sometimes people note that the frequency or

29:10

other properties of these oscillations will change

29:12

with different tasks or different brain regions

29:15

will have different kind of characteristic

29:17

frequencies and these kinds of oscillations. And

29:19

then, yeah, the question is... Are these doing

29:21

anything or is it that the

29:23

brain wanted excitatory and inhibitory neurons

29:25

for other reasons and a consequence

29:28

of that is that there are

29:30

oscillations? A lot of people for a

29:32

long time have put forth theories

29:34

about how these oscillations could be

29:36

useful, how they could be helpful

29:38

for encoding information. For example, you

29:40

can imagine like different bits of

29:42

information are encoded at different parts

29:45

of this kind of wave that's

29:47

happening in the neural activity. That's

29:49

one method. There's ways in which

29:51

oscillations could be helpful in learning

29:54

if you can compare like two

29:56

time points. So people, especially because

29:58

oscillations to some extent... have been

30:00

a little bit easier to observe with

30:02

certain methodology. So for example, like EEG,

30:04

which people use to do human neural

30:07

recordings without having to actually open the

30:09

skull, one of the primary things that

30:11

you can get out of that is

30:13

you can measure oscillations and different frequencies

30:15

and stuff. You can't measure the activity

30:18

of individual neurons, but you can get

30:20

a sense of general oscillations. And so...

30:22

because this is something that people can

30:24

measure and they see correlations between things.

30:27

They see signatures in these oscillations that

30:29

relate to behavior and to what a

30:31

person is doing during the experiment. Yeah,

30:33

there's been this big hope of, you

30:35

know, truly identifying the role of these

30:38

oscillations. I think it's complicated for a

30:40

few reasons. As I said, there's... different

30:42

styles of oscillations in different brain regions,

30:44

so they're probably doing different things and

30:46

maybe they're playing more of an important

30:49

role in some brain regions or under

30:51

certain circumstances than in others. And then

30:53

the other kind of complication is that

30:55

whenever you want to test is this

30:57

oscillation important, really ideally you'd be able

31:00

to modulate the oscillation and change something

31:02

about the oscillation and show that it

31:04

changes behavior in some way. And we

31:06

have some ways of trying to do

31:08

that, but usually any time you change

31:11

a nature of the oscillation, you're changing

31:13

something about the neural firing itself. And

31:15

then so people have questions of like,

31:17

oh, was it really the oscillation or

31:19

was the neural firing? And so you

31:22

get into these debates of like people

31:24

saying that these oscillations are just byproducts

31:26

of the actual firing that matters. So

31:28

I don't know that it's necessarily resolved

31:30

or anything. I think the debate is

31:33

ongoing. From my personal perspective, I guess

31:35

I do feel like I've seen more

31:37

people. take some of these theories more

31:39

seriously lately, that might just be like

31:41

kind of crowd of people that I

31:44

know who their research interests have got

31:46

in such a way that they're more

31:48

seriously exploring what these oscillations could do,

31:50

whereas maybe previously that same kind of

31:52

crowd of researchers may have thought that

31:55

these were the important thing to study.

31:57

But that could be. just a personal

31:59

bias as to what I'm seeing. So

32:01

yeah, I think it remains a thing

32:03

that people, you know, some people are

32:06

going to pursue these things and they're

32:08

going to try to figure out what

32:10

these are for and other people are

32:12

going to think, meh, I don't have

32:15

to care about oscillations, I'll care about

32:17

other signals in the neural code and

32:19

it'll be fine. And then, you know,

32:21

hopefully people come together to create the

32:23

full picture. But yeah, usually these things

32:26

are studied a little bit a little

32:28

bit separately. You have

32:30

also mentioned that oscillations might

32:32

play very different roles in

32:35

different parts of the brain

32:37

and in chapter 8 you

32:39

point out that one brain

32:41

region which turned out to

32:44

be particularly tricky to study

32:46

is the motor cortex. Do

32:48

you have an idea by

32:50

this is more complex than

32:53

other systems? Yeah, it's

32:55

a good question. So yeah, the motor

32:57

cortex has proven a bit challenging. And

32:59

I think maybe to understand the nature

33:01

of the challenge, it's important to think

33:04

about actually how we study other systems

33:06

and especially in the history of neuroscience,

33:08

where a lot of people were studying

33:10

vision and the kind of. original gold

33:12

standard in vision research was that you're

33:15

able to show a visual stimulus, like

33:17

you put something on a screen and

33:19

you record from a neuron and you

33:21

can clearly say this neuron responds to

33:23

this thing. If I, you know, put

33:26

it on the screen, the neuron is

33:28

firing a lot, if I take it

33:30

off, the neuron is not firing, if

33:32

I change it slightly, the neuron fires

33:34

less and less, and so you're trying

33:36

to find this thing that the neuron

33:39

responds to that the neuron responds to.

33:41

And so basically people try to do

33:43

that same thing in motor cortex by

33:45

having Subjects in these experiments make movements

33:47

and then try to find okay when

33:50

The arm moves in this direction then

33:52

this neuron fires a lot and when

33:54

it moves slightly differently from that direction

33:56

the direction the neuron fires less and

33:58

trying to find these things that, you

34:00

know, these neurons are tuned for or

34:03

these neurons are responsible for encoding, whatever,

34:05

the direction of movement. That didn't pan

34:07

out in the exact same way that

34:09

it did in the visual system where

34:11

you could kind of identify at least

34:14

for parts of the visual system you

34:16

can cleanly identify things that these neurons

34:18

respond to. And I think there could

34:20

be a variety of reasons for that.

34:22

And one is just that the... relationship

34:25

between the motor cortex and movement is

34:27

not like the it's not this direct

34:29

line where the motor cortex is the

34:31

only thing that's driving movement. There's all

34:33

of these nested loops and hierarchies that

34:35

exist in the motor system where you

34:38

know you can think of something as

34:40

simple as a reflex which doesn't even

34:42

involve the brain you know you hit

34:44

your me and your leg moves that's

34:46

something those there are reflexes like that

34:49

are just the spinal cord controlling movement

34:51

you know you get some sensory input

34:53

and the spinal cord itself says okay

34:55

now we're going to move the muscle

34:57

and then you have more elaborate versions

35:00

of those that go through kind of

35:02

subcortical regions or different regions of the

35:04

brain as well. And so you get

35:06

this stack of little loops where all

35:08

of these different influences are controlling movement.

35:10

And so to try to directly say,

35:13

what is the relationship between this one

35:15

neuron and motor cortex and this movement

35:17

when the movement is the result of

35:19

many influences and also the movement requires

35:21

many muscles, you might just not be

35:24

able to get that kind of relationship,

35:26

that interpretable relationship. And so. It's a

35:28

little bit like the study of motor

35:30

cortex was led astray by the study

35:32

of sensory systems where you could get

35:34

a little bit more of a one-to-one

35:37

correspondence like that. It might just be

35:39

that that's not the way to think

35:41

about the motor cortex, but then the

35:43

question becomes what is the most productive

35:45

way to think about the role of

35:48

the motor cortex? It might be that

35:50

you have to understand all of these

35:52

other lower level loops first before you

35:54

can really appreciate what the motor cortex

35:56

is doing on top of those things.

36:00

How much does this difference come from

36:02

the fact that in one case we

36:04

are trying to understand the information processing

36:06

activity of the plane with the sensory

36:09

systems and in the motor of cortex

36:11

we are more focused on the output?

36:13

So is it fair to say that

36:15

understanding the output is trickier generally? Yeah,

36:17

I mean, I think especially... if you're

36:20

thinking of kind of unconstrained movement like

36:22

just people deciding how to move, there

36:24

is this question of, you know, where

36:26

does that signal even come from? In

36:29

these experimental settings, people really did try

36:31

to get it as close to the

36:33

kind of sensory experiment as they could,

36:35

you know, the sensory experiment and you

36:38

show something and the subject has to

36:40

look at it and that's it. And

36:42

then these, so they would tell people

36:44

how you're supposed to move. and then

36:46

they would do the movement. So they

36:49

were trying to make it as analogous

36:51

as possible. But yeah, of course, this

36:53

is, it's not the same as receding

36:55

sensory information if you're producing a motor

36:58

output. And so yeah, I think that

37:00

probably, you know, it complicates the picture

37:02

of how you study this system when

37:04

it is this internally generating the thing

37:06

that you're studying. And so you don't

37:09

even really get to see the. like

37:11

the initial stages of the generation. You're

37:13

just picking one point in the system

37:15

to record neurons from and hope that

37:18

that part is responsible enough for the

37:20

movement that you get an interesting signal.

37:22

Whereas with sensory processing, we could trace

37:24

kind of exactly the steps that go

37:26

from, you know, the stimulus on the

37:29

screen into the retina and some cortical

37:31

region and cortex and you could say,

37:33

okay, this is why I'm studying this

37:35

part of the brain because I know

37:38

that it gets visual information via these

37:40

steps. Whereas with the motor cortex, you're

37:42

telling the animal to move and they

37:44

move and motor cortex does something as

37:47

a result of that. And but you

37:49

still, I think, as I said, with

37:51

all of these nested loops, the exact

37:53

place in the system that, you know,

37:55

motor cortex is located is maybe not.

37:58

So. clean and easy

38:00

to describe. Interesting.

38:02

And this book was

38:04

published in 2021. And

38:07

what other topics would you include

38:10

if you were writing a similar

38:12

book now? Yeah, so I try

38:14

to be pretty broad. And these

38:16

are topics that have already been

38:18

studied for decades. And so, you

38:20

know, they kind of cover the

38:22

main things. I think it's more

38:24

that there has been obviously progress

38:26

in some of these areas or

38:28

some changes to how people think

38:30

about things in some of these

38:32

areas. So for example, there's a

38:34

chapter on reinforcement learning and how

38:36

the signal that represents rewards and

38:39

reward prediction and errors of reward

38:41

prediction, how that can be encoded

38:43

in the brain. And there's been

38:45

a lot of recent papers that

38:47

have kind of complicated that picture

38:49

a little bit lately and pointed

38:51

to a system that has more

38:53

complex representations of reward and, you

38:55

know, it made people rethink how

38:57

to interpret different neural responses and

38:59

all of that. So, you know,

39:01

in a lot of these areas,

39:03

there's just been a bit more

39:05

progress that has happened since I

39:08

wrote the book. And so I

39:10

don't know if there's whole topic

39:12

areas that I would add because

39:14

the science doesn't move that quickly.

39:16

So not that much has changed.

39:18

But there are definitely places where

39:20

there's been more progress that it

39:22

would be interesting to cover and

39:24

there are some topics that I

39:26

even knew at the time that

39:28

I just didn't have enough space

39:30

to fit them in and it

39:32

would have been nice. So one

39:34

example is I do talk about

39:37

grid cells and hippocampus and place

39:39

cells and this kind of thing,

39:41

but it's a relatively short section

39:43

and there's actually a lot of

39:45

mathematical theory that surrounds the study

39:47

of how the brain represents space

39:49

and navigation and space that I

39:51

just didn't get the time or

39:53

the space to cover as the

39:55

book was written. So yeah, I

39:57

think in terms of sense, it

40:00

was written, it's mostly just been

40:02

advances or some a little bit

40:04

of shake up in some of

40:07

our understanding of these topics.

40:09

And do I understand correctly

40:11

that there there has been

40:13

some significant progress like with

40:15

the report and revert dictional

40:17

for a big part of

40:20

the conclusion is that it's

40:22

even more complicated than it

40:24

so. Yeah, basically. But you know, that

40:26

kind of thing is, it's welcome when

40:28

if you think that you're not fully

40:31

understanding the system, if, you know, the

40:33

model that you have doesn't fully explain

40:35

behavior or doesn't fully explain neural responses,

40:37

then you want someone to come along

40:40

and be like, actually it's more complex

40:42

in this specific way. And then now,

40:44

you know, we understand it, and it

40:46

becomes part of, you know, the established.

40:49

story that we tell about that system

40:51

and how those things work. And once

40:53

it becomes an established part of it,

40:55

then it doesn't feel that complicated anymore.

40:57

And then you could hope for another

40:59

round of that where even more refinements

41:01

happen and we can incorporate that into

41:03

our understanding. So yeah, so some of

41:05

it is the, yeah, just the things

41:07

that things are more complex and we're

41:09

figuring out the ways in which they're

41:12

more complex, which should allow us to

41:14

explain more data ideally. And then

41:16

I guess there is, you

41:18

know, since this book came

41:20

out, there is the whole

41:22

topic of large language models

41:25

and general kind of

41:27

AI, those models are less related to

41:29

the brain, less related to how the

41:32

brain is built and how the brain

41:34

learns and how the brain functions. But

41:36

they're obviously something that people have a

41:39

lot of curiosity about because when you

41:41

interact with them, they seem kind of

41:43

like a person and they can do

41:45

tasks at human or above levels on

41:48

a lot of things. And so yeah, I

41:50

don't, in the context of this book, which

41:52

is about the study of the brain, I'm

41:54

not sure how they would have fit in,

41:56

but clearly it's kind of a related topic

41:58

area of these are artificial. neural network

42:00

models. They can do a lot

42:02

of interesting things. People are trying

42:04

to understand how they work. So

42:06

that is maybe a topic area

42:08

that I would have tried to

42:10

speak to at least a bit

42:12

in the book if that had

42:14

come out at the time that

42:16

I was writing it. And another

42:18

big event has been since you

42:20

published a book that approximately one

42:22

or half a year later you

42:24

started a lab. And given that

42:26

the book that writes a lot

42:29

about how different fields of science

42:31

interact with each other and learn

42:33

from each other, it's not a

42:35

big surprise that you were left

42:37

started out with. multiple different topics.

42:39

Can you tell about that a

42:41

bit? Yeah, so I have my

42:43

lab at NYU now and we're

42:45

trying to focus on kind of

42:47

three different lines of research. One

42:49

is a continuation of things that

42:51

I did during my PhD in

42:53

postdoc, which is modeling sensory processing

42:55

using artificial neural networks. So we

42:57

focus mostly on the visual system,

42:59

but the main questions are kind

43:01

of how is sensory processing influenced

43:03

by kind of non-sensory things like

43:05

you know what you see is

43:07

not just a result of the

43:09

light that hits your retina there's

43:11

a bunch of other top-down signals

43:13

in the brain things that people

43:15

frequently call attention that can determine

43:17

what you actually perceive and what

43:19

you act on. And so we're

43:21

kind of looking into questions of

43:23

how do we model that, how

43:25

are we able to explain certain

43:27

patterns in behavior by understanding the

43:29

neural mechanisms of those top-down signals

43:31

or those contextual signals that actually

43:33

change how visual information is processed

43:35

in the visual system. So that's

43:37

one topic area. Another area that

43:39

I've got more interested in is

43:41

what would be called interpretability in

43:43

the field of AI, which is

43:45

how do we study artificial neural

43:47

networks? If we build these networks

43:49

and we train them and they

43:52

can do things and then we

43:54

don't really know how they're doing

43:56

what they're doing and we can't

43:58

really predict the way that they're

44:00

going to make mistakes or we

44:02

can't. easily fix the kind of

44:04

mistakes that they make. And my

44:06

interest in that is because neuroscientists

44:08

face a pretty similar question. We

44:10

have this giant neural network. It

44:12

happens to be a real one,

44:14

not an artificial one, but we

44:16

don't really know how it works.

44:18

We would still like to be

44:20

able to understand it and fix

44:22

it. And so can there be

44:24

more of an interaction between these

44:26

two fields as kind of what

44:28

I'm trying to advocate for and

44:30

trying to push research that uses

44:32

artificial neural networks not just as

44:34

a model of the brain, but

44:36

as a testing ground for the

44:38

tools that we use to analyze.

44:40

neural data and just answer questions

44:42

about what does it mean to

44:44

understand the neural network and what

44:46

tools can we find and use

44:48

to make that understanding easier. And

44:50

then there's a third line of

44:52

research in my lab which is

44:54

really separate which is that I

44:56

have some projects on applying machine

44:58

learning. techniques to climate change problems.

45:00

And I'm particularly focused on the

45:02

analysis of satellite imagery because those

45:04

methods are computer vision methods which

45:06

overlap with the methods that I

45:08

used to study the visual system.

45:10

So the motivation for that work

45:12

was more out of acknowledgement of

45:15

the seriousness of the problem and

45:17

the sense that kind of everyone

45:19

needs to pivot what they're doing

45:21

to try to help with this

45:23

huge societal and technological problem of

45:25

climate change. And then, you know,

45:27

the niche area that makes sense

45:29

for me is this analysis of

45:31

satellite imagery. And I've been able

45:33

to collaborate with people and find

45:35

a lot of different applications for

45:37

how, you know, just taking images

45:39

of the earth at different points

45:41

in time can do a lot

45:43

to help us understand things that

45:45

are going on and to help

45:47

plan for adaptation and understand existing

45:49

adaptation efforts and all that kind

45:51

of thing. So it seems a

45:53

little bit far afield, but there

45:55

is an overlap in the methodology

45:57

because I'm studying the visual system

45:59

and using computational models that can

46:01

also be applied to setting satellite

46:03

imagery. And it actually makes a

46:05

lot of sense in the light

46:07

of your book. where you described

46:09

how neuroscience has profited from findings

46:11

in many areas and climate change

46:13

is an even younger field and

46:15

newer field. So what do you

46:17

think what would help it to

46:19

use information from other fields efficiently?

46:21

Yeah, I mean it is, you

46:23

know, the problem of climate change

46:25

is it spans every topic area.

46:27

There's nothing that you can bring

46:29

up that I wouldn't be able

46:31

to tell you how it's connected

46:33

to climate change either because it's

46:35

part of, you know, creating greenhouse

46:38

gases or because it's something that

46:40

will be impacted by a changing

46:42

climate or you know everything is

46:44

involved in some way and so

46:46

yeah you really need all of

46:48

the disciplines to tackle all of

46:50

these many individual problems and so

46:52

yeah I do think coming from

46:54

scientific training going into this more

46:56

kind of applied, applying these techniques

46:58

to problems. I think that that

47:00

is interesting and then it brings

47:02

a different perspective in terms of

47:04

what's useful for being able to

47:06

build systems that will actually help

47:08

people. And I think obviously, you

47:10

know, there's interdisciplinary needs in terms

47:12

of people who actually work on

47:14

policy and like getting things going

47:16

on the ground and how we

47:18

can integrate technology into their decision-making

47:20

process. Also, you know, I know

47:22

people in the field of psychology

47:24

who are trying to understand human

47:26

decision-making and how do people decide

47:28

if they care about climate change,

47:30

what they're going to do, why

47:32

do some people not care about

47:34

it, why do people care about

47:36

it, but not act on it,

47:38

how do people, you know, form

47:40

collective groups that can act on

47:42

it? And so these are just

47:44

a couple of many, many different

47:46

types of interdisciplinary interactions that are

47:48

happening and need to happen so

47:50

that the progress can be made

47:52

faster and more effective. And so

47:54

I'm particularly interested in this intersection

47:56

of bringing machine learning and technology-based

47:58

approaches to problems for people who.

48:00

actually trying to do things on

48:03

the ground and what tools would

48:05

help them and how can we

48:07

better communicate with them about what

48:09

tools are available and how to

48:11

use them. Well, thanks a lot

48:13

and thanks for this conversation

48:15

and if you are the best with these

48:18

fascinating projects and I would like

48:20

to encourage the listeners that if

48:22

you are interested in how the

48:24

brain works and what we know

48:26

about it and what we don't

48:29

know and what the open questions

48:31

and debates are, I definitely recommend

48:33

to check out models of the

48:35

mind, which is a very sort

48:37

of introduction into a complex topic.

48:40

Great, thank you very having me.