Good afternoon.

My name is Marvin Chun,

and I am a researcher,

and I teach psychology at Yale.

I'm very proud of what I do,

and I really believe in the value
of psychological science.

However, to start with a footnote,

I will admit that I feel
self-conscious sometimes

telling people that I'm a psychologist,

especially when you're meeting
a stranger on the airplane

or at a cocktail party.

And the reason for this

is because there's
a very typical question that you get

when you tell people
that you're a psychologist.

What do you think that question might be?

(Audience) What am I thinking?

"What am I thinking?" exactly.

"Can you read my mind?"

is a question that a lot
of psychologists cringe at

whenever they hear it

because it really reflects
a misconception about -

a kind of a Freudian misconception
about what we do for a living.

So you know, I've developed
a lot of answers over the years:

"If I could read your mind,

I'd be a whole lot richer
than I am right now."

The other answer I came up with is,

"Oh, I'm actually a neuroscientist,"

and then that really makes people quiet.

(Laughter)

But the best answer of course is,

"Yeah, of course I can read your mind,

and you really should be
ashamed of yourself."

(Laughter)

But all that is motivation
for what I'll share with you today,

which is I've been a psychologist
for about 20 years now,

and we've actually reached this point
where I can actually answer that question

by saying, "Yes,
if I can put you in a scanner,

then we can read your mind."

And what I'll share with you today
is to what extent, how far are we

at being able to read
other people's minds.

Of course this is the stuff
of science fiction.

Even a movie as recent as "Divergent"

has these many sequences

in which they're reading out
the brain activity

to see what she's dreaming

under the influence of drugs.

And you know, again,
these technologies do exist.

In modern day, in the real word,
they look more like this.

These are just standard
MRI machines and scanners

that are slightly modified

so that they can allow you
to infer and read out brain activity.

These methods have been around
for about 25 years,

and in the first decade -
I'd say the 1990s -

a lot of effort was devoted

to mapping out what the different
parts of the brain do.

One of the most successful forms
of this is exemplified here.

Imagine you're a subject
lying down in that scanner;

there's a computer projector

that allows you
to look at a series of images

while your brain is being scanned.

For some periods of time,

you are going to be seeing
sequences of scene images

alternating with other
sequences of face images,

and they'll just go back and forth
while your brain is being scanned.

And the key motivation here
for the researchers

is to compare what's
happening in the brain

when you're looking at scenes
versus when you're looking at faces

to see if anything different happens
for these two categories of stimuli.

And I'd say even within 10 minutes
of scanning your brain,

you can get a very reliable pattern
of results that looks like this,

where there is a region of the brain
that's specialized for processing scenes,

shown here in the warm colors,

and there is also
a separate region of the brain,

shown here in blue colors,

that is more selectively active
when people are viewing faces.

There's this disassociation here.

And this allows you to see

where different functions
reside in the brain.

And in the case
of scene and face perception,

these two forms of vision
are so important in our everyday lives -

or course scene processing
important for navigation,

face processing important
for social interactions -

that our brain has
specialized areas devoted to them.

And this alone, I think,
is pretty amazing.

A lot of this work
started coming out in the mid-1990s,

and it really complimented and reinforced

a lot of the patient work
that's been around for dozens of years.

But really, the methodologies of fMRI

have really gone far beyond
these very basic mapping studies.

You know, in fact,
when you look at a study like this,

you may say, "Oh, that's really cool,

but what does this have to do
about everyday life?

How is this going to help me understand
why I feel this way or don't feel this way

or why I like a certain person
or don't like certain other people?

Will this tell me what kind
of career I should take?"

You know, you can't really get too far

with this scene and face
area mapping alone.

So it wasn't until this study
came up around 2000 -

And most of these studies,
I should emphasize,

are from other labs around the field.

Any studies from Yale,

I'll have the mark of "Yale"
on the corner.

But this particular study was done at MIT,
by Nancy Kanwisher and colleagues.

They actually had subjects
look at nothing.

They had subjects
close their eyes in the scanner

while they gave them instructions,

and the instructions
were one of two types of tasks.

One was to imagine
a bunch of faces that they knew.

So you might be able to imagine
your friends and family,

cycle through them one at a time

while your brain is being scanned.

That would be the face
imagery instruction.

And then the other instruction
that they gave the subjects

was of course a scene imagery instruction,

so imagine after this talk,
you walk home or take your car home,

get into your home,
navigate throughout your house,

change into something more comfortable,

go to the kitchen, get something to eat.

You have this ability to visualize
places and scenes in your mind,

and the questions is, "What happens
when you're visualizing scenes,

and what happens
when you're visualizing faces?

What are the neural correlates of that?"

And the hypothesis was, well, maybe
imagining faces activates the face area

and maybe imagining scenes
would activate the scene area,

and that's in fact indeed what they found.

First, you bring in a bunch
of subjects into the scanner,

show them faces, show them scenes,

you map out the face area on top,

you map out the place area
in the second row.

And then the same subjects,

at a different time in the scanner,

have them close their eyes,
do the instructions I just had you do,

and you compare.

When you have face imagery,

you can see in the second row
that the face area is active,

even though nothing's on the screen.

And in the fourth row here,
the second row of the bottom pair,

you see that the place area is active
when you're imagining scenes;

nothing's on the screen,

but simply when you're imagining scenes,
it will activate the place area.

In fact, even just
by looking at fMRI data alone

and looking at the relative activity
for the face area and place area,

you can guess over 80% of the time,

even in this first study,

what subjects were imagining.

Okay?

And so, in this sense, in my mind,

I believe that this is the first study
that started using brain imaging

to actually read out
what people were thinking.

You can take a naïve person,
look at this graph, you know,

and ask which bar is higher,
which line is higher,

and they can guess better than chance,
way better than chance,

what people were thinking.

Now, you may say, "Okay,
that's really cool,

but there're a whole lot more
interesting things to do with a scanner

or with what we want to know

than whether people
are thinking of faces or scenes."

And so the rest of my talk

will be devoted to the next 15 years
of work since this study

that have really advanced
our ability to read out minds.

This is one study
that was published in 2010

in the New England Journal of Medicine

by another group.

They were actually studying patients
who were in a persistent vegetative state,

who were locked in,
had no ability to express themselves;

it was not even clear if they were
capable of voluntary thought,

and yet because you can instruct people
to imagine one thing versus another -

in this case, imagine walking
through your house, navigation,

in the other case, imagine playing tennis

because that activates
a whole motor circuitry

that's separate from the spacial
navigation circuitry -

you attach each of those imagery
instruction to "yes" or "no,"

and then you can ask them
factual questions,

such as "Is your father's name Alexander,
or is your father's name Thomas?"

And they were able to demonstrate
that in this patient,

who otherwise had no ability
to communicate,

he was able to respond
to these factual questions accurately,

and even in control subject
who do not have problems.

This method is so reliable

that almost over 95%
of their responses were decodable

just through brain imagining alone -
"yes" versus "no."

So you can imagine,
with a game of 20 questions,

you can really get insight
into what people are thinking

using these methodologies.

But again, it's limited;

we only have two states here -

"yes," "no,"

think tennis or think
walking around your house.

But fortunately, there're tremendously
brilliant scientists all around the world

who are constantly refining the methods
to decode fMRI activity

so that we can understand the mind.

And I'd like to use this slide
from Nature magazine

to motivate the next few studies
I'll share with you.

One huge limitation of brain imaging,
at least in the first 15 years,

is that there are not
many specialized areas in the brain.

The face area is one.
Place area is another.

These are very, very important visual
functions that we carry out every day,

but there is no "shoe" area in the brain,
there is no "cat" area in the brain.

You don't see separate blobs
or patches of activity

for these two different categories,

and indeed, for most categories
that we encounter

and are able to discriminate,

they don't have separate brain regions.

And because there are
no separate brain regions,

the question then is,

"How do we study them,
and how do we discriminate them,

how do we get at these more fine
detailed differences that matter so much

if we really want to understand

how the mind works
and how we can read it out

using fMRI?

Fortunately, some neuroscientists
collaborated with computer vision people

and with electrical engineers
and computer scientists and statisticians

to use very refined mathematical methods

for pulling out and decoding
more subtle patterns of activity

that are unique to shoe processing

and that are unique
to the processing of cat stimuli.

So, for instance,
even though the shoes and cats,

you show a whole bunch
of them to subjects,

it'll activate same part of cortex,
the same part of the brain,

and so that if you average
activity in that part,

you won't see any differences
between the shoes and the cats.

As you can see here on the right,

the shoes are still associated
with different fine patterns,

or micropatterns of activity

that are different from the patterns
of activity elicited by cats.

These little boxes over here,
these cells are referred to voxels;

the way fMRI collects data from the brain

is it kind of chops up your brain
into tiny little cubes,

and each cube is a unit of measurement -

it gives you numerical data
that you can use to analyze.

And if you imagine
that the brightness of these squares

corresponds to a different
level of activity,

quantitatively measurable
different level of activity,

and you can imagine over a spacial region

that the pattern of these activities
may be different for shoes versus cats,

then all you need to do
is to work with a computer algorithm

to learn these subtle
differences in patterns,

so we have what we might call a "decoder."

And you train up a computer decoder

to learn the differences
between shoe patterns and cat patterns,

and then what you have
is here in the testing phase:

after you train up
your computer algorithm,

you can show a new shoe,

it will elicit a certain pattern
of fMRI activity,

as you can see over here,

and the computer
will tell you its best guess

as to whether this pattern over here

is closer to a shoe
or closer to that for a cat.

And it will allow you

to guess something
that's a lot more refined

and that is not possible

when you don't have areas
like the face area or the place area.

And now people are really taking off
with these types of methodologies.

Those started coming out in 2005,
so a little than a decade ago.

This is a study published
in the New England Journal of Medicine,

year 2013, literally last year,

a very important clinical
application of studying

or trying to find a way
to objectively measure pain,

which is a very subjective state, we know.

You can use fMRI,

and these researchers have developed ways

of mapping and predicting
the amount of heat

and the amount of pain people experience

as you increase the temperature
of a noxious stimulus on their skin,

and the fMRI decoder
will kind of give you an estimate

of how much heat
and maybe even how much pain

someone is experiencing,

to the extent that they can predict,
just based on fMRI activity alone,

how much pain a person is experiencing
over 93% of the time,

or with 93% accuracy.

This is an astonishing version
that I'm sure many of you have seen,

published in 2011,

from Jack Gallant's group
over in UC, Berkley.

They showed people a bunch
of these videos from YouTube,

here shown here on the left,

and then they trained up
a classifier, a decoder,

to learn the mapping
between the videos and fMRI activity;

and then when they showed
people novel videos

that elicited certain patterns
of fMRI activity,

they were able to use
that fMRI activity alone

to guess what videos people were seeing.

And they pulled the guesses
off of YouTube as well

and averaged them,

and that's why it's a little grainy,

but what you'll see here -
and this is all over YouTube -

is what people saw

and, on the right, what the computer
is guessing people saw

based on fMRI activity alone.

So it's pretty astonishing;
this literally is reading out the mind.

I mean, you have complicated models,
you have to train it and such,

but still, this is based
on fMRI activity alone

that they're able to make
these guesses over here.

I had a student at Yale University.
His name is Alan Cowen.

He had a very strong
mathematics background.

He came to my lab,
and he said, "I really love this stuff."

He loved this stuff so much
that he wanted to do it himself.

I actually did not have the ability
to do this in the lab,

but he developed it
together with my post doc Brice Kuhl,

who's now an assistant professor
at New York University.

They wanted to ask the question

of "Can we limit this type
of analysis to faces?"

As you saw before,

you can kind of see what categories
people were looking at,

but certainly the algorithm
was not able to give you guesses

as to which person you were looking at

or what kind of animal
you were looking at.

You're really getting
more categorical guesses

in the prior example.

And so Alan and Brice,

they thought that, well, if we focused
our algorithms on faces alone -

you know, the faces
being so important for everyday life,

and given the fact

that there are specialized mechanisms
in the brain for face processing -

maybe if we do something
that's that specialized,

we might be able to individually
guess individual faces.

As a summary, that's what they did.

They showed people a bunch of faces,

they trained up these algorithms,

and then using fMRI activity alone,
they were able to generate good guesses,

above-chance guesses,

as to which faces
an individual is looking at,

again, based on fMRI activity alone,

and that's what's shown here on the right.

Just to give you a little bit more detail

for those who kind of
like that kind of stuff.

It's really relying
on a lot of computer vision.

This is not voodoo magic.

There's a lot of computer vision
that allows you

to mathematically summarize
features of a whole array of faces.

People call them "eigenfaces."

It's based on principle
component analysis.

So you can decompose these faces
and describe a whole array of faces

with these mathematical algorithms.

And basically what
Alan and Brice did in my lab

was to map those mathematical
components to brain activity,

and then once you train up a database,
a statistic library, for doing so,

then you can take a novel face
that's shown up here on the right,

record the fMRI activity
to those novel faces,

and then make your best guess
as to which face people were looking at.

Right now, we are about 65, 70% correct,

and we and others are working very hard
to improve that performance.

And this is just another example:

a whole bunch of originals
on the left here,

and then reconstructions
and guesses on the right.

This study actually just came out,
just got published this past week,

so I'm pretty excited

that you might even see
some media coverage of the work,

and again, I really want to credit
Alan Cowen and Brice Kuhl

for making this possible.

Now, you may wonder,

going back, kind of, to the more
categorical reading out of brains -

That work came out in 2011.

I've been talking about it a lot
in classes and stuff,

and a really typical
question that you get -

and this question I actually like -

is, "Well, if you can read out
what people are seeing,

can you read out what they're dreaming?"

because that's the natural next step.

And this is work done
by Horikawa et al. in Japan,

published in Science last year.

They actually took a lot of the work
that Jack Gallant did,

modified it so they can decode
what people were dreaming

while they were sleeping
inside the scanner.

And what we have here on the -

Ignore the words,

these are just ways to try to categorize
what people were dreaming about.

Just focus on the top left.

That's the computer
algorithm's guess, or decoder's guess

as to what people were dreaming about.

You can imagine
it's going to be a lot more messier

than what you saw before for perception,

but still, it's pretty amazing

that you can decode dreams now
while people are sleeping,

you know, while they are
in rapid eye movement sleep.

So what we have here -
I'll turn it on, it's a video.

I will mention that,
a little stereotypically,

this is a male subject,

and male subjects tend to think
about one thing when they're dreaming,

(Laughter)

but it is rated PG,
if you'll excuse me for that.

And again, you know, it's not easy.

It's, again, remarkable
that they can do this,

and it gets better
as it goes towards the end,

right before they awakened the subject
to ask them what they were dreaming,

but you can start seeing
buildings and places,

and here comes the dream,
the real concrete stuff,

right there.

Okay.

(Laughter)

Then they wake up the subject and say,
"What were you dreaming about?"

and they will report something
that's consistent with the decoder,

as a way to access its validity.

So, we can decode and read out the mind;

the field, the kind
of neuroscience can do that.

And the question is, "What are we
going to use this capability for?"

Some of you may have seen this movie, 
the Minority Report.

It's a fascinating, amazing movie.
I encourage you to see it if you can.

This movie is all
about predicting behavior,

you know, reading out brain activity

to try to predict what's going
to happen in the future.

And neuroscientists are working
on this aspect of cognition as well.

As one example, colleagues
over at the University of New Mexico,

Kent Kiehl's group,

scanned a whole bunch of prisoners
who committed felony crimes.

They were all released back into society,

and the depended measure was:

"How likely was it
for someone to be rearrested?

What was the likelihood of recidivism?"

And they found that based on brain scans

while people were
doing tasks in the prison -

in the scanner that was
located in the prison,

they could predict reliably

who was more likely
to come back to prison,

be rearrested, re-commit a crime

versus who was less likely to do so.

In my own laboratory here at Yale,
I worked with a undergraduate student.

This is, again,
his own idea, Harrison Korn,

together with Michael Johnson.

They wanted to see if they could measure
implicit or unconscious prejudice

while people were looking
at these vignettes

which involve employment
discrimination cases.

So you can read through that:

"Rodney, a 19 year-old African American,

applied for a job as a clerk
at a teen-apparel store.

Had experience as a cashier,

glowing recommendation
from a previous employer

but was denied the job because he did not
match the company's look."

Okay?

And et cetera, et cetera, et cetera.

And the question is,

"In this hypothetical damage awards case,

how much would you
as a subject award Rodney?"

And you can imagine a range of responses,

where you would award Rodney a lot
or award Rodney a little,

and we had many other cases like this.

And the question is,

"Can we predict which subjects
are going to award a lot of damages

and which subjects
are going to award few damages?"

And Harrison found that if you scan people

while they were looking
at white faces versus black faces,

based on the brain response
to those faces,

he could predict

the amount of awards that were given
in this hypothetical case later on.

And that's shown here
by these correlations on the left.

As a final thing I want to share
of my lab right now,

we're really interested in this issue
of "what does it mean to be in the zone?

what does it mean
to be able to stay focused

for a sustained amount of time?"

which, you might imagine,

is critical for almost anything
that's important to us,

whether that's athletic performance,
musical performance,

theatrical performance,
giving a lecture, taking an exam.

Almost all these domains
require sustained focus and attention.

You all know that there are times

when you're really in the zone
and can do well

and there are other times
where even if you're fully prepared,

you don't do as well,

because you're not in the zone
or you're kind of distracted.

So my lab is interested
in how we can characterize this

and how we can predict that.

And as work that's not even published yet,

Emily Finn, a graduate student
in neuroscience program,

and Monica Rosenberg,
a graduate student in psychology,

they're both working with me

and Todd Constable over at the Magnetic
Resonance Research Center at Yale.

They find that it's not easy to predict
what it means to be in the zone

and what it means to be out of the zone,

but if you start using measures

that also look at how different brain
areas are connected with each other -

something that we call
"functional connectivity" -

you can start getting at
what it means to be in the zone

versus what it means
to be out of the zone.

And what we have here

are two types of networks 
that Emily and Monica have identified.

There is a network of brain regions
and the connectivity between them

that predicts good performance,

and there is another complementary network

that has different brain areas
and different types of connectivity

that seems to be correlated
with bad performance.

So if you had an activity
in the blue network over here,

people tend to do worse.

If you have activity in the red network,
people tend to do better.

These networks are so robust

that you can measure these networks
even when subjects are doing nothing.

We put them in the scanner
before anything starts,

have them close their eyes
and rest for 10 minutes -

you call that a "resting state scan" -

and based on their resting
state activity alone,

we were able to predict how well
they were going to do over the next hour.

Again, red network corresponds
with better performance;

blue network corresponds
with worse performance.

And just on a variety of tasks,

you have very high
predictive power in these studies.

The implications for studying ADHD
and many other types of domains,

I think, are very large,

so hopefully, this will get
published soon.

So in closing, I hope I've convinced you

that I am no longer embarrassed to say
that we can read your mind,

if anyone asks me.

And I thank you for your attention.

(Applause)