03x03 - The Stilwell Brain

[bunniefuu]

♪

"I think, therefore, I am."

But am I?

I think. Ha.

A single microscopic brain cell
cannot think,

is not conscious,

but if you bring in
a few more brain cells,

and a few more,
and connect them all,

at a certain point,
the group itself will

be able to think

and experience emotions

and have opinions
and a personality

and know that it exists.

How can such astonishing things

be made from
such simple ingredients?

Well, answering that question
means learning not only who
we are,

but, more importantly,
how we are.

Today, using what
neuroscientists know so far,

I am going to make my hometown

function like a brain!

( all cheering, applauding )

♪

A single brain cell is tiny,

both in size and abilities.

But when enough are together,
they can do amazing things

like be aware of themselves.

When the collective power
of a group working together

is greater than the sum
of their individual parts,

that is called "emergence."

In a similar fashion,
we as individuals

are connected to
the people around us.

Those connections
form communities that,
when functioning properly,

can work together
to accomplish amazing feats.

A great example is
"wisdom of the crowds."

Even if not a single person
in a crowd

knows the right answer
to a question,

collectively, they could all
somehow know the right answer.

In 1987, economist
Jack Treynor

conducted
the "Bean Jar" experiment.

He asked 56 students
to guess the number
of jellybeans in a jar.

Now, as you can probably guess,

not a single one of them
guessed the right answer.

But amazingly, when he took
the average of their guesses,

what he got was a number within
just 3% of the real answer.

Now, some people
guessed way too high,

but others guessed
way too low,

so all together,
their errors balanced out,

and from a whole bunch
of wrong guesses,

the true answer emerged.

What else can a crowd do?

If I got a bunch
of humans together

and had each one of them
act like a brain cell,

turning on or off in response
to the actions of other people,

could I make a neural network

like the one in our brain?

And if I had enough people,

could intelligence, emotions,

a mind, emerge?

If I recruited
every single person

in the country of China

and arranged them like neurons,

would the result not only be
a simple brain,

but something that can think
and feel

and be aware of
its own existence?

Well, this is the China Brain
thought experiment,

first proposed by Lawrence Davis
and, later, Ned Block.

It's never been done before
and, well, unfortunately,

I don't have access
to everyone in China.

I made some calls,
and like a lot of them are busy.

But the first step is to see
what a crowd in real life

could even do.

This hasn't been done
successfully before,

but I want to blow
a neural network

up to the scale of a crowd.

And what better crowd to use
than one made of the people

whose emergent properties
made me who I am today?

That's right, I am going home
to Stilwell, Kansas.

♪

♪

( birds chirping )

Michael:
For help designing the brain

we would make out of people,

I recruited Chris Eliasmith,

director of the Center
for Theoretical Neuroscience

at the University of Waterloo.

So Chris,
we're headed south,

going down to
the heart of Stilwell,

- where I grew up.
- Nice.

We're going to do something
a little bit weird. Um.

I want to create a brain.

- Right.
- OK? But with
a crowd of people.

It sounds like a challenge,
for sure.

I looked into it,

and I found that the roundworm
has a brain

that's made up of only
300-some-odd neurons.

- That's right.
- We can get 300 people,

and where better to get
these people to make a brain

than my hometown of Stilwell?

This was the community
that, in many ways,
made me who I am.

Michael:
This is all downtown Stilwell.

Some of my earliest memories
are from here.

This used to be,
and maybe still is,
a feed store,

and they would have sno-cones
during the summer.

It was the most awesome,
delicious thing ever.

But as you can see,

a lot of corn
is grown in Kansas,

but around here, the main thing
that I saw being grown

- was just sod.
- Oh, really?

Yeah, There's a famous sod farm
around here

whose slogan was
"High on grass."

- ( Chris laughs )
- It was pretty...pretty edgy
for the time.

OK, so back to the brain
that we're gonna make.

You know, building brains is
in my job description.

I wrote a book called
How to Build a Brain.

Michael:
Chris is known for
is neural network,

the Semantic Pointer
Architecture Unified Network,

or SPAUN, which is one of
the world's most complex

computer simulations
of the brain.

It uses 6.6 million
simulated neurons

to perform functions
like counting, reasoning,

and image recognition.

SPAUN is cutting-edge,

but neural networks
are nothing new.

The first was made
by Dr. Frank Rosenblatt

of Cornell University in 1957.

His network, called
the Perceptron,

was designed for
image recognition,

and he hoped it would become
capable of learning,

just like a brain.

But the project was only
partially successful,

and after some controversy,
fell by the wayside.

It was only when researchers
in the 1980s

came back upon
Dr. Rosenblatt's work,

and as computing power
increased,

that the field of artificial
neural networks

came back to the mainstream.

Today, it is alive and well.

SPAUN, and even neural networks
used in self-driving cars,

are expanding the possibilities
of computer learning.

If I want to make
a brain out of people,
where do I start?

That's a good question.

I think the first thing
we want to do is figure out

what we want our brain to do.

I would recommend
something like vision.

Vision. Let's make
this brain see.

Michael:
Before we can design
the intricacies

of the brain we're making,

let's look at how
visual processing works.

Let's say we look at a cat.

Light information
from every point on the cat

lands on the retina.

This information gets sent
to our visual cortex.

The visual cortex
is structured in layers--

V1 through V6.

Each of these layers
are made up of neurons

activated by specific features,

like lines, angles, and shapes.

The features that are detected

are sent to
the infratemporal cortex

which puts all the pieces
of the image together,

and we get our Eureka! moment

where we recognize
the object we're looking at,

what it means what feelings
we have towards it.

I love cats.

But what should we have
our brain recognize?

We don't want
really high resolution images

or images that depend
on too much detail,

- Ok.
- so things like letters
and digits.

Let's say we use digits. Ok?

- Ok.
- I want to be the one
who draws a digit,

and then you will be
on the output side.

You should be able
to determine what I've drawn;

not because I showed it
to someone and they
telephoned it back to you,

but because
they processed it
intelligently.

That's what we need
to figure out,

how we're gonna
show an input to our people.

So we should take
some small number of them

and put them at the front,
as the retina,

and really just show them each
a little bit of the image.

So if, for instance,
we're able to put like
25 people in that kind
of front row,

the "input" layer,
then whatever image we show

should be made up
of 25 pixels.

- Exactly. Right.
- Twenty-five pieces.

I'm gonna draw
25 people.

1, 2, 3, 4, 5,

6, 7, 8, 9, 10,

11, 12, 13, 14, 15,

16, 17, 18, 19, 20,

21, 22, 23, 24, 25.

See? I can count.

These are our retina cells,

and each one is
an individual person

that's literally standing,
like, in a field.

What do they then do next?

They merely need to indicate
whether or not

their cell is on or off.

- All right.
- So, they should start f*ring.

They should start spiking
like a neuron.

What could they do
to indicate

that they're f*ring or not?

They could jump up and down,

they could wave a flag...

OK, I like that.

When Chris and I use words
like "f*ring" and "spiking,"

we're talking about
how brain cells, neurons,

talk to one another

by sending an electric message
from one cell...

to another.

It's called
an "action potential,"

and it travels down the axon
of the cell.

When the ionic flow
into a brain cell

reaches a certain threshold,

the cell will fire an electronic
message down its axon.

So a neuron
can either be on or off.

It's either f*ring
or it's not f*ring.

What we need to find
is a way for a person

to be either on or off--

raising a flag or their hand
should do the trick.

To illustrate our visual input,

I will be drawing a number
from 0 to 9

onto a grid divided
into 25 squares, or pixels.

Now each person, or neuron,
will receive one pixel.

If a neuron receives
a pixel with writing on it,

it will fire.

The V1 layer identifies
pixels in the retinal layer

that form particular lines
in the number,

and the V2 layer identifies
particular combinations

of lines from V1
that form angles.

- What does V3 do?
- V3 is more sensitive to color.

We're only working with
black and white in this case.

So you're saying we won't even
need to have a V3 in our brain?

We're skipping V3 altogether.

All right, sorry, V3.

- So we're gonna go
straight to V4?
- Yeah.

Michael:
V4 neurons will fire

when their assigned
combinations of angles

have been detected.

At this point
the basic shape of a number

is beginning to take form.

And so actually the next one
is called IT.

- Ooh!
- And that stands for
infratemporal cortex.

Michael:
Now don't worry.

We haven't forgotten V5 and V6.

They exist, they're responsible

for higher-level image
processing in our brain.

But for our demonstration,
we don't need them.

We do, however, need
the infratemporal cortex,

which is the final layer needed
for visual processing

in the brain we're designing.

Our IT will consist
of ten neurons

representing the numbers
0 through 9.

They will be looking at neurons
in V4,

and will only fire when their
corresponding neurons fire.

For example, if one
or multiple V4 neurons

representing the shapes
of an 8 fire,

the IT neuron
representing the 8
will also fire.

Voila! We just recognized
a number.

So what happens after
the infratemporal layer?

So after that,
I think that's where I'll be.

It's gonna be me making a
decision about what digit
I think was actually shown at
the end.

So I think we're gonna need
a couple hundred people,

so one question is,
where do you put
that many people?

I would love to use
my high school football field.

The question is,
is it gonna work?

- I am hopeful right now.
- We got our work cut out for
us.

Michael: Knowing how
we want to structure

our neural network,
it was now time for us

to head to my high school
football field

where our "brain"
will take form.

I spent a lot of time here

impressing the world
with my body's prowess--

at the clarinet.

OK, so Chris,
I brought you here

because we need to talk about

the actual logistics of getting
all these people together.

So here's the plan
as I see it.

Everyone is going to be
wearing a shirt

that is a different color,
based on what layer they're in.

We're also gonna give
everyone one of those, um,

like, "I'm running
in a marathon" kind of...

- The big bibs?
- Bibs. Thank you. Yes.

I knew that too. You see,

I'm always checking, because
Chris is one of those nerds

who doesn't know
what us sportos
talk about.

Anyway, the bibs
will give every person
an individual number,

so if something goes wrong,

we can target
that one brain cell

and say, "Are you damaged?
What do you need?"

Take out the problem.

OK, we're here
on this 40 yard line.

I will be here.
This is gonna be
the input layer.

The retinal cells will be
all in front of me,
all 25 of them.

You are gonna be
way down in the end zone
on the output side.

And I'm gonna be
using the scoreboard.

-Ok.
- So when you
make your prediction,

based on what you think
the brain has figured out,

we'll put that on
the Visitor's side

and I'll reveal the Home number
as what I really wrote.

- Sounds good.
- So literally from here
to that end zone

is the amount of space
we're going to need

for these hundreds of people

to also have
the right eye lines.

Just have to make sure that
communication lines are open,

meaning that it's easy to see
whoever you have to
pay attention to.

Michael:
Our human brain will have
a couple hundred people

spread out in five layers

across half a football field.

Every single participant
will be assigned

to only react
to certain neurons

in the layer ahead of them.

And it's complicated,
so their positions on the field

had to be carefully chosen
so that every neuron

has a clear line of sight

to the neurons
they are connected to.

In a way...

something will be born
on this field tomorrow.

( laughs )

- We shall find out.
- We'll find out!

All right.

( crowd chattering )

Michael:
So what does it take
to turn Stilwell into a brain?

Well, seven tents, 550 chairs,

twenty gallons of coffee,
three hundred flags,

t-shirt and hats,

our drone operator Jeff,

this cute little Gator,

two hundred cinnamon rolls,

and of course,
our medic, Brian.

Now all that's left
is to pull this all off.

A community is something
that is bigger

that the sum
of all of its parts,

and so is a brain.

Now, today, I'm feeling
pretty excited about

the neural network
we're gonna build out of people,

because there's a zero percent
chance of rain,

but a 100% chance of brain.

OK, we better get started.

The gates are open,
and our neurons are filing in.

First, they're all given
color-coded t-shirts

associated with the layer of
the brain they will represent.

Just want go in the center?

Michael:
And then they will
take to the field

to get in position.

Michael:
You all in the orange shirts
are the retina.

Your job is to say,
"Is there writing on my square?"

or "Not." If your square
has any writing

or black marks on it,
raise your flag--
oh, and stand up.

Now I'm sure all you
mega-brainiacs out there

remember every detail about
how this brain

is going to recognize numbers.

But just in case you don't,
here's a refresher.

I will draw a number
on a 25-pixel grid,

break the squares up,
and hand them out

to every person
in the retina layer.

The retinal neurons
will only fire if they
have writing on their pixel.

The people in the yellow shirts,

you guys are V1.

Each V1 neuron will be watching

three retinal neurons
in front of them

and fire only if all three

of their assigned retinal
neurons fire,

revealing lines
that make up the number.

You guys are V2.

You're a bit more advanced.

You're looking for combinations
of features

that make, for instance, angles.

The V2 neurons will be watching
the V1 layer,

and fire only if
their assigned V1 neurons fire,

revealing angles.

V4 neurons will be
looking at V2 neurons.

Their f*ring reveals
combinations of angles

that begin to form the number.

Finally, the purple shirts.

You all are
infratemporal cortex.

Extremely important role,

not more important
than the others, though.

Part of the IT's function
is to inhibit

incorrect results it receives
from the V4 layer.

For example, if V4 neurons
are indicating

both a 6 and an 8,

an 8 will outrank a 6
because an 8 has more features.

Chris will determine the number
by interpreting the results
from the IT layer.

Got it? Good. Because
it's happening now.

Michael (over loudspeaker):
All right!

It is time for me
to draw my first numeral.

Stand by.

There it is.

Now it's time to distribute
these pixels

to the photoreceptors
in the retina.

25, 24...

19...

13...

All right.

I have distributed

the input to the retina layer.

Is everyone ready?

( all cheering, applauding )

Three, two,
one, think!

And they're off.
The retina layer has fired,

passing off signals to V1.

V2 sees V1 f*ring,
and also fires,

cuing V4 and IT.

There's a lot of flags
on the play, folks.

Look at all this processing.

Good work, stay up,

I'm now walking back to Chris

where he will tell me
what you guys have processed.

All right, Chris,
it actually happened

way faster than I thought.

It took me forever
to get over here.

They were already done
"thinking."

It was really interesting
to watch.

We had a little bit of noise
in the system, for sure,

because we actually kind of
have two answers at the end.

So I'm going to be doing
something that brains do,

which is kind of
make a guess sometimes
based on the best evidence.

Michael:
All right, Chris.

What numeral
do you think I drew?

Chris:
I think you drew a 3.

Let's get that up
on the Visitor's scoreboard.

The numeral I truly drew...

was a 3!

( all cheering, applauding )

Chris:
Nice work.

There was some noise
in the system.

I think we can perfect this
a little bit,

because it wasn't
a confident 3,

but Chris did still
get it right.

And by "Chris,"
I mean all of you.

Michael:
We gave each of our IT neurons
a clear tube and plastic balls

to make their job easier.

Every time they see one of the
neurons they're watching fire,

they put a ball in their tube.

If a neuron
they're inhibited by

has more balls than they do,

they stop f*ring.

Michael:
Our model had a mistake.

You should have been
inhibited by 254,

meaning that if 254 is f*ring,

and puts a ball in the tube,

- you just sit down.
- OK.

But we didn't have 254
written down into your code.

I'm gonna do that right now.

With that kink worked out,

it was time to try again.

Take that one.
Thank you.
24...

Michael:
Three, two, one...

go!

Michael:
Ooh. There's not much happening

in V2 and V4.

Or IT, for that matter.

Looks like our brain d*ed.

Michael:
Something definitely
went wrong.

The processing stopped in V2.

So, I think our brain broke.

Not a single person in V4

or the infratemporal cortex

has been activated.

Chris: Seems like something
very strange happened

when the light
went through the lens
and got to the retina.

OK, but what you really mean to
say is that I handed the pixels
out to the wrong neurons.

- It's lookin' that way.
- Wow, who would have thought

that the worst-working
part of the machine

would be the actual people
who get to be people

- and not brain cells?
- I know, right?

OK, I think
I'm gonna do an 8 this time.

And I'm gonna put it
kind of up in this corner,

or along that side.

This is pretty weird,
it's not centered,

it's not filling up
the whole space.

Let's see if our brain
can recognize it.

Each pixel of our image
needs to go

to a particular retinal neuron.

To make sure I didn't
mess it up again,

we put numbers
on the back of each one.

Is everyone ready?

- ( all cheer )
- OK.

Three, two, one, go!

Michael:
Oh, yeah.

Whoa!
That was fast.

Michael:
Blink, and you'll miss it,
so let's take a break,

because this is the Mind Field
Play of the Game.

The 8 I drew contained
13 pixels,

and bam! the 13 retinal cells
connected to those locations
are f*ring.

Now, that's what I call
a sensation.

V1 reads the formation
perfectly.

They don't even know it,
but each one f*ring

means a horizontal, vertical,

or diagonal line
has been caught.

Now look at neuron 40's speed.

It's sensitive only to
a horizontal line

low and to the right,
which my 8 had.

If retinal cells 23, 24, and 25
all fire together,

such a line has been sensed,
and watch this--boom!

Champion reflexes there,
folks.

If the V1 neurons
a V2 player is watching, fire,

they stand, and that means
that the lines V1 caught

made some corner angle.

Standing V4 neurons are shapes
made by those corners,

but it all comes down to IT.

A bunch of them fire.

Lots of numbers contain
the shapes I drew,

but there can only be
one MVP,

and, dang,
look at this teamwork!

3 is inhibited by 8,

0 is inhibited by 8.

If 8 is getting
this much activity,

sit down
and let that neuron score!

This, my friends, is what
we in the cognition sports biz

call...a gr8 play.

Chris (over loudspeaker:
The purple people neurons

make me think that
you wrote an 8.

An 8? Let's put an 8
up on the scoreboard.

And as it turns out,
the numeral that I did write...

was an 8!

( all cheering, applauding )

- Nice job!
- That was good.

Michael:
This was our first
definitive success.

Now it's time to really
put the system to the test.

So I'm gonna draw a 1,

but then I'm gonna add
a line down here

and I'm gonna do
a dot right there.

And we're gonna see if
that noise trips up our brain.

Three, two, one, go!

My guess is a 1.

The number that I drew was,
in fact, a 1.

( cheers, applause )

7...and I'm gonna put
a line through it.

Go!

Oh, yeah.

Man, these people are good.

Chris:
The brain thinks
it's just saw a 7.

I wrote a 7.

- ( cheers, applause )
- Nice work.

Nice work.

Michael:
With two successful
results in a row,

for the last test
I want to see what will happen

if I really
mess with our brain.

I'm not even going to draw
a number;

instead, I'm going to
fill in every single cell.

This means that
every single neuron
in the retina will fire,

will stand up and wave.
Let's see what happens.

My prediction, of course,
it's gonna look like an 8,

because an 8 is a numeral that
fills in a lot of the cells.

Are you all ready?

- ( all cheer )
- Good.

- Chris, are you ready?
- Ready!

Michael:
All right, I'm ready.

Three, two, one, go!

Chris:
What?!

( laughing )

- ( all cheering )
- Michael: Look at that.

- Michael: Holy cow!
- ( Chris laughs )

It looks to me like you
essentially opened up the eye

and shone a laser
right into it.

Right, yeah.
So what I did is,
I didn't even draw a numeral.

I just scribbled
all over the whole thing,
I filled in every single cell.

- So what does the brain
think that I drew?
- The 8.

I had no idea the inhibition
would work that well.

We were able to single
all of that mess

down into just one guess,

and it really was
the smartest guess.

Right, it was the one with
the most features in it.

Michael:
Congratulations to
the entire infratemporal cortex,

V4, V2, V1, retina.

You guys have been amazing.
Great work.

( cheers, applause )

Today, I was a neuron.

My favorite part
about the brain

was that it actually worked.

My favorite part
was just the whole experience,

doing science,
meeting cool people.

It's just a really good
simulation of how
the brain works,

and it was just really cool
to take some information
from that.

Michael:
And as always,

thanks for...being a brain.

( all cheer )

Michael:
Our small-town brain

did work as predicted.

Made up of only
a couple hundred neurons,

each one with no idea
what number I was drawing,

it was nonetheless able
to process the image

and determine
the correct answer.

We created a living,
breathing model

of a part of the human brain.

And our demonstration
was a new way

to illustrate and share

how the human brain
processes visual information.

We were able to watch it
process...think...

in real time,
and that's amazing.

Its success shows what we can
achieve by working together,

and we only used
a small fraction

of the number of neurons found
in an actual human brain.

So imagine how powerful

the connections of
not a few hundred,

but a hundred billion
human neurons could be.

Now, interestingly,

a hundred billion people
is about how many humans

have ever existed
in the history of Earth.

There are only a few billion
alive right now,

so I guess that means
get procreating...please.

I want to make a bigger
superhuman mind.

No. I'd like to thank
every neuron

from my hometown of Stilwell,

and the entire community
that supported us.

Because without them
all working together,

none of this
would have been possible.

And as always...

thanks for watching.

♪

- ( no audible dialogue )
- ♪