20:19
TED2015

Laura Schulz: The surprisingly logical minds of babies

Filmed:

How do babies learn so much from so little so quickly? In a fun, experiment-filled talk, cognitive scientist Laura Schulz shows how our young ones make decisions with a surprisingly strong sense of logic, well before they can talk.

- Cognitive scientist
Developmental behavior studies spearheaded by Laura Schulz are changing our notions of how children learn. Full bio

Mark Twain summed up
what I take to be
00:12
one of the fundamental problems
of cognitive science
00:14
with a single witticism.
00:18
He said, "There's something
fascinating about science.
00:20
One gets such wholesale
returns of conjecture
00:23
out of such a trifling
investment in fact."
00:26
(Laughter)
00:29
Twain meant it as a joke,
of course, but he's right:
00:32
There's something
fascinating about science.
00:34
From a few bones, we infer
the existence of dinosuars.
00:37
From spectral lines,
the composition of nebulae.
00:42
From fruit flies,
00:47
the mechanisms of heredity,
00:50
and from reconstructed images
of blood flowing through the brain,
00:53
or in my case, from the behavior
of very young children,
00:57
we try to say something about
the fundamental mechanisms
01:02
of human cognition.
01:05
In particular, in my lab in the Department
of Brain and Cognitive Sciences at MIT,
01:07
I have spent the past decade
trying to understand the mystery
01:12
of how children learn so much
from so little so quickly.
01:16
Because, it turns out that
the fascinating thing about science
01:20
is also a fascinating
thing about children,
01:23
which, to put a gentler
spin on Mark Twain,
01:27
is precisely their ability
to draw rich, abstract inferences
01:29
rapidly and accurately
from sparse, noisy data.
01:34
I'm going to give you
just two examples today.
01:40
One is about a problem of generalization,
01:42
and the other is about a problem
of causal reasoning.
01:45
And although I'm going to talk
about work in my lab,
01:47
this work is inspired by
and indebted to a field.
01:50
I'm grateful to mentors, colleagues,
and collaborators around the world.
01:53
Let me start with the problem
of generalization.
01:59
Generalizing from small samples of data
is the bread and butter of science.
02:02
We poll a tiny fraction of the electorate
02:06
and we predict the outcome
of national elections.
02:09
We see how a handful of patients
responds to treatment in a clinical trial,
02:12
and we bring drugs to a national market.
02:16
But this only works if our sample
is randomly drawn from the population.
02:19
If our sample is cherry-picked
in some way --
02:23
say, we poll only urban voters,
02:26
or say, in our clinical trials
for treatments for heart disease,
02:28
we include only men --
02:32
the results may not generalize
to the broader population.
02:34
So scientists care whether evidence
is randomly sampled or not,
02:38
but what does that have to do with babies?
02:42
Well, babies have to generalize
from small samples of data all the time.
02:44
They see a few rubber ducks
and learn that they float,
02:49
or a few balls and learn that they bounce.
02:52
And they develop expectations
about ducks and balls
02:55
that they're going to extend
to rubber ducks and balls
02:58
for the rest of their lives.
03:01
And the kinds of generalizations
babies have to make about ducks and balls
03:03
they have to make about almost everything:
03:07
shoes and ships and sealing wax
and cabbages and kings.
03:09
So do babies care whether
the tiny bit of evidence they see
03:14
is plausibly representative
of a larger population?
03:17
Let's find out.
03:21
I'm going to show you two movies,
03:23
one from each of two conditions
of an experiment,
03:25
and because you're going to see
just two movies,
03:27
you're going to see just two babies,
03:30
and any two babies differ from each other
in innumerable ways.
03:32
But these babies, of course,
here stand in for groups of babies,
03:36
and the differences you're going to see
03:39
represent average group differences
in babies' behavior across conditions.
03:41
In each movie, you're going to see
a baby doing maybe
03:47
just exactly what you might
expect a baby to do,
03:49
and we can hardly make babies
more magical than they already are.
03:53
But to my mind the magical thing,
03:58
and what I want you to pay attention to,
04:00
is the contrast between
these two conditions,
04:02
because the only thing
that differs between these two movies
04:05
is the statistical evidence
the babies are going to observe.
04:08
We're going to show babies
a box of blue and yellow balls,
04:13
and my then-graduate student,
now colleague at Stanford, Hyowon Gweon,
04:16
is going to pull three blue balls
in a row out of this box,
04:21
and when she pulls those balls out,
she's going to squeeze them,
04:24
and the balls are going to squeak.
04:27
And if you're a baby,
that's like a TED Talk.
04:29
It doesn't get better than that.
04:32
(Laughter)
04:34
But the important point is it's really
easy to pull three blue balls in a row
04:38
out of a box of mostly blue balls.
04:42
You could do that with your eyes closed.
04:44
It's plausibly a random sample
from this population.
04:46
And if you can reach into a box at random
and pull out things that squeak,
04:49
then maybe everything in the box squeaks.
04:53
So maybe babies should expect
those yellow balls to squeak as well.
04:56
Now, those yellow balls
have funny sticks on the end,
05:00
so babies could do other things
with them if they wanted to.
05:02
They could pound them or whack them.
05:05
But let's see what the baby does.
05:07
(Video) Hyowon Gweon: See this?
(Ball squeaks)
05:12
Did you see that?
(Ball squeaks)
05:16
Cool.
05:20
See this one?
05:24
(Ball squeaks)
05:26
Wow.
05:28
Laura Schulz: Told you. (Laughs)
05:33
(Video) HG: See this one?
(Ball squeaks)
05:35
Hey Clara, this one's for you.
You can go ahead and play.
05:39
(Laughter)
05:51
LS: I don't even have to talk, right?
05:56
All right, it's nice that babies
will generalize properties
05:59
of blue balls to yellow balls,
06:02
and it's impressive that babies
can learn from imitating us,
06:03
but we've known those things about babies
for a very long time.
06:06
The really interesting question
06:10
is what happens when we show babies
exactly the same thing,
06:12
and we can ensure it's exactly the same
because we have a secret compartment
06:15
and we actually pull the balls from there,
06:18
but this time, all we change
is the apparent population
06:20
from which that evidence was drawn.
06:24
This time, we're going to show babies
three blue balls
06:27
pulled out of a box
of mostly yellow balls,
06:30
and guess what?
06:34
You [probably won't] randomly draw
three blue balls in a row
06:35
out of a box of mostly yellow balls.
06:38
That is not plausibly
randomly sampled evidence.
06:40
That evidence suggests that maybe Hyowon
was deliberately sampling the blue balls.
06:44
Maybe there's something special
about the blue balls.
06:49
Maybe only the blue balls squeak.
06:52
Let's see what the baby does.
06:55
(Video) HG: See this?
(Ball squeaks)
06:57
See this toy?
(Ball squeaks)
07:02
Oh, that was cool. See?
(Ball squeaks)
07:05
Now this one's for you to play.
You can go ahead and play.
07:10
(Fussing)
(Laughter)
07:18
LS: So you just saw
two 15-month-old babies
07:26
do entirely different things
07:29
based only on the probability
of the sample they observed.
07:31
Let me show you the experimental results.
07:35
On the vertical axis, you'll see
the percentage of babies
07:37
who squeezed the ball in each condition,
07:40
and as you'll see, babies are much
more likely to generalize the evidence
07:42
when it's plausibly representative
of the population
07:46
than when the evidence
is clearly cherry-picked.
07:49
And this leads to a fun prediction:
07:53
Suppose you pulled just one blue ball
out of the mostly yellow box.
07:55
You [probably won't] pull three blue balls
in a row at random out of a yellow box,
08:00
but you could randomly sample
just one blue ball.
08:04
That's not an improbable sample.
08:07
And if you could reach into
a box at random
08:09
and pull out something that squeaks,
maybe everything in the box squeaks.
08:11
So even though babies are going to see
much less evidence for squeaking,
08:15
and have many fewer actions to imitate
08:20
in this one ball condition than in
the condition you just saw,
08:22
we predicted that babies themselves
would squeeze more,
08:25
and that's exactly what we found.
08:29
So 15-month-old babies,
in this respect, like scientists,
08:32
care whether evidence
is randomly sampled or not,
08:37
and they use this to develop
expectations about the world:
08:40
what squeaks and what doesn't,
08:43
what to explore and what to ignore.
08:45
Let me show you another example now,
08:50
this time about a problem
of causal reasoning.
08:52
And it starts with a problem
of confounded evidence
08:55
that all of us have,
08:57
which is that we are part of the world.
08:59
And this might not seem like a problem
to you, but like most problems,
09:01
it's only a problem when things go wrong.
09:04
Take this baby, for instance.
09:07
Things are going wrong for him.
09:09
He would like to make
this toy go, and he can't.
09:10
I'll show you a few-second clip.
09:13
And there's two possibilities, broadly:
09:21
Maybe he's doing something wrong,
09:23
or maybe there's something
wrong with the toy.
09:25
So in this next experiment,
09:30
we're going to give babies
just a tiny bit of statistical data
09:32
supporting one hypothesis over the other,
09:35
and we're going to see if babies
can use that to make different decisions
09:38
about what to do.
09:41
Here's the setup.
09:43
Hyowon is going to try to make
the toy go and succeed.
09:46
I am then going to try twice
and fail both times,
09:49
and then Hyowon is going
to try again and succeed,
09:52
and this roughly sums up my relationship
to my graduate students
09:55
in technology across the board.
09:58
But the important point here is
it provides a little bit of evidence
10:02
that the problem isn't with the toy,
it's with the person.
10:05
Some people can make this toy go,
10:08
and some can't.
10:11
Now, when the baby gets the toy,
he's going to have a choice.
10:12
His mom is right there,
10:16
so he can go ahead and hand off the toy
and change the person,
10:18
but there's also going to be
another toy at the end of that cloth,
10:21
and he can pull the cloth towards him
and change the toy.
10:24
So let's see what the baby does.
10:28
(Video) HG: Two, three. Go!
(Music)
10:30
LS: One, two, three, go!
10:34
Arthur, I'm going to try again.
One, two, three, go!
10:37
YG: Arthur, let me try again, okay?
10:45
One, two, three, go!
(Music)
10:48
Look at that. Remember these toys?
10:53
See these toys? Yeah, I'm going
to put this one over here,
10:55
and I'm going to give this one to you.
10:58
You can go ahead and play.
11:00
LS: Okay, Laura, but of course,
babies love their mommies.
11:23
Of course babies give toys
to their mommies
11:27
when they can't make them work.
11:30
So again, the really important question
is what happens when we change
11:32
the statistical data ever so slightly.
11:35
This time, babies are going to see the toy
work and fail in exactly the same order,
11:38
but we're changing
the distribution of evidence.
11:42
This time, Hyowon is going to succeed
once and fail once, and so am I.
11:45
And this suggests it doesn't matter
who tries this toy, the toy is broken.
11:49
It doesn't work all the time.
11:55
Again, the baby's going to have a choice.
11:57
Her mom is right next to her,
so she can change the person,
11:59
and there's going to be another toy
at the end of the cloth.
12:02
Let's watch what she does.
12:05
(Video) HG: Two, three, go!
(Music)
12:07
Let me try one more time.
One, two, three, go!
12:11
Hmm.
12:17
LS: Let me try, Clara.
12:19
One, two, three, go!
12:22
Hmm, let me try again.
12:27
One, two, three, go!
(Music)
12:29
HG: I'm going
to put this one over here,
12:35
and I'm going to give this one to you.
12:37
You can go ahead and play.
12:39
(Applause)
12:58
LS: Let me show you
the experimental results.
13:04
On the vertical axis,
you'll see the distribution
13:07
of children's choices in each condition,
13:09
and you'll see that the distribution
of the choices children make
13:12
depends on the evidence they observe.
13:16
So in the second year of life,
13:19
babies can use a tiny bit
of statistical data
13:21
to decide between two
fundamentally different strategies
13:24
for acting in the world:
13:27
asking for help and exploring.
13:29
I've just shown you
two laboratory experiments
13:33
out of literally hundreds in the field
that make similar points,
13:37
because the really critical point
13:40
is that children's ability
to make rich inferences from sparse data
13:43
underlies all the species-specific
cultural learning that we do.
13:48
Children learn about new tools
from just a few examples.
13:53
They learn new causal relationships
from just a few examples.
13:58
They even learn new words,
in this case in American Sign Language.
14:03
I want to close with just two points.
14:08
If you've been following my world,
the field of brain and cognitive sciences,
14:12
for the past few years,
14:15
three big ideas will have come
to your attention.
14:17
The first is that this is
the era of the brain.
14:20
And indeed, there have been
staggering discoveries in neuroscience:
14:23
localizing functionally specialized
regions of cortex,
14:27
turning mouse brains transparent,
14:30
activating neurons with light.
14:33
A second big idea
14:36
is that this is the era of big data
and machine learning,
14:38
and machine learning promises
to revolutionize our understanding
14:43
of everything from social networks
to epidemiology.
14:46
And maybe, as it tackles problems
of scene understanding
14:50
and natural language processing,
14:53
to tell us something
about human cognition.
14:55
And the final big idea you'll have heard
14:59
is that maybe it's a good idea we're going
to know so much about brains
15:01
and have so much access to big data,
15:05
because left to our own devices,
15:06
humans are fallible, we take shortcuts,
15:09
we err, we make mistakes,
15:13
we're biased, and in innumerable ways,
15:16
we get the world wrong.
15:20
I think these are all important stories,
15:24
and they have a lot to tell us
about what it means to be human,
15:27
but I want you to note that today
I told you a very different story.
15:31
It's a story about minds and not brains,
15:35
and in particular, it's a story
about the kinds of computations
15:39
that uniquely human minds can perform,
15:42
which involve rich, structured knowledge
and the ability to learn
15:45
from small amounts of data,
the evidence of just a few examples.
15:49
And fundamentally, it's a story
about how starting as very small children
15:56
and continuing out all the way
to the greatest accomplishments
16:00
of our culture,
16:04
we get the world right.
16:08
Folks, human minds do not only learn
from small amounts of data.
16:12
Human minds think
of altogether new ideas.
16:18
Human minds generate
research and discovery,
16:20
and human minds generate
art and literature and poetry and theater,
16:23
and human minds take care of other humans:
16:29
our old, our young, our sick.
16:32
We even heal them.
16:36
In the years to come, we're going
to see technological innovations
16:39
beyond anything I can even envision,
16:42
but we are very unlikely
16:46
to see anything even approximating
the computational power of a human child
16:48
in my lifetime or in yours.
16:54
If we invest in these most powerful
learners and their development,
16:58
in babies and children
17:03
and mothers and fathers
17:06
and caregivers and teachers
17:08
the ways we invest in our other
most powerful and elegant forms
17:11
of technology, engineering and design,
17:15
we will not just be dreaming
of a better future,
17:18
we will be planning for one.
17:21
Thank you very much.
17:23
(Applause)
17:25
Chris Anderson: Laura, thank you.
I do actually have a question for you.
17:29
First of all, the research is insane.
17:34
I mean, who would design
an experiment like that? (Laughter)
17:36
I've seen that a couple of times,
17:41
and I still don't honestly believe
that that can truly be happening,
17:42
but other people have done
similar experiments; it checks out.
17:46
The babies really are that genius.
17:49
LS: You know, they look really impressive
in our experiments,
17:50
but think about what they
look like in real life, right?
17:53
It starts out as a baby.
17:56
Eighteen months later,
it's talking to you,
17:57
and babies' first words aren't just
things like balls and ducks,
17:59
they're things like "all gone,"
which refer to disappearance,
18:02
or "uh-oh," which refer
to unintentional actions.
18:05
It has to be that powerful.
18:07
It has to be much more powerful
than anything I showed you.
18:09
They're figuring out the entire world.
18:12
A four-year-old can talk to you
about almost anything.
18:14
(Applause)
18:17
CA: And if I understand you right,
the other key point you're making is,
18:19
we've been through these years
where there's all this talk
18:22
of how quirky and buggy our minds are,
18:25
that behavioral economics
and the whole theories behind that
18:27
that we're not rational agents.
18:29
You're really saying that the bigger
story is how extraordinary,
18:31
and there really is genius there
that is underappreciated.
18:35
LS: One of my favorite
quotes in psychology
18:40
comes from the social
psychologist Solomon Asch,
18:42
and he said the fundamental task
of psychology is to remove
18:45
the veil of self-evidence from things.
18:47
There are orders of magnitude
more decisions you make every day
18:50
that get the world right.
18:55
You know about objects
and their properties.
18:56
You know them when they're occluded.
You know them in the dark.
18:58
You can walk through rooms.
19:01
You can figure out what other people
are thinking. You can talk to them.
19:02
You can navigate space.
You know about numbers.
19:06
You know causal relationships.
You know about moral reasoning.
19:08
You do this effortlessly,
so we don't see it,
19:11
but that is how we get the world right,
and it's a remarkable
19:14
and very difficult-to-understand
accomplishment.
19:16
CA: I suspect there are people
in the audience who have
19:19
this view of accelerating
technological power
19:21
who might dispute your statement
that never in our lifetimes
19:24
will a computer do what
a three-year-old child can do,
19:27
but what's clear is that in any scenario,
19:29
our machines have so much to learn
from our toddlers.
19:32
LS: I think so. You'll have some
machine learning folks up here.
19:38
I mean, you should never bet
against babies or chimpanzees
19:41
or technology as a matter of practice,
19:45
but it's not just
a difference in quantity,
19:49
it's a difference in kind.
19:53
We have incredibly powerful computers,
19:55
and they do do amazingly
sophisticated things,
19:57
often with very big amounts of data.
20:00
Human minds do, I think,
something quite different,
20:03
and I think it's the structured,
hierarchical nature of human knowledge
20:05
that remains a real challenge.
20:09
CA: Laura Schulz, wonderful
food for thought. Thank you so much.
20:11
LS: Thank you.
(Applause)
20:14

▲Back to top

About the Speaker:

Laura Schulz - Cognitive scientist
Developmental behavior studies spearheaded by Laura Schulz are changing our notions of how children learn.

Why you should listen

MIT Early Childhood Cognition Lab lead investigator Laura Schulz studies learning in early childhood. Her research bridges computational models of cognitive development and behavioral studies in order to understand the origins of inquiry and discovery.

Working in play labs, children’s museums, and a recently-launched citizen science website, Schultz is reshaping how we view young children’s perceptions of the world around them. Some of the surprising results of her research: before the age of four, children expect hidden causes when events happen probabilistically, use simple experiments to distinguish causal hypotheses, and trade off learning from instruction and exploration.

More profile about the speaker
Laura Schulz | Speaker | TED.com