Live 1057

Free Internet Radio Stations
Lecture 3.1: Liz Spelke – Cognition in Infancy (Part 1)

Lecture 3.1: Liz Spelke – Cognition in Infancy (Part 1)

The following content is
provided under a Creative Commons license. Your support will help
MIT OpenCourseWare continue to offer high quality
educational resources for free. To make a donation or
view additional materials from hundreds of MIT courses,
visit MIT OpenCourseWare at ELIZABETH SPELKE: I want to
start with an observation about this summer school. There’s a lot of development
in this summer school. You’ve got two full mornings
devoted to it– today and on Thursday. It also came up pretty majorly
in Josh Tenenbaum’s class last Friday and I learned
early this morning also in Shimon Ullman’s
class that I couldn’t be here for yesterday afternoon. And the issues have come up
in many other classes as well, including Nancy’s, Winrich
Freiwald’s, and so forth. Now, what’s come up is not
only the general questions about development, but specific
questions about human cognitive development. Questions that have
been addressed primarily through behavioral
experiments, not experiments using neural
methods or computational models. And the topic that I’m
going to be trying to– that Allie and I will try
to get you to think about for this morning is
even narrower than that. It’s about the cognitive
capacities of human infants. And I think a fair
initial question would be, why so much focus on
early human development? And that question
will get sharper if you look at where
major organizations are putting their research money. They are not putting it
into the kind of work that I’m going to be
talking about today. There is no– in the
Obama BRAIN Initiative, where they’re looking
for new technologies, there’s no call for
new technologies to figure out what
human knowledge is like at or near
the initial state and how it grows over
the course of infancy. And the European
Human Brain Project doesn’t have development as
a major area in it, either. So I think it’s fair
to ask, why is CBMM taking such a different approach
and putting so much emphasis on trying to get you guys
to think about and learn about human development? And two general
reasons, I think. One is, it’s
intrinsically fascinating. Come on. We are the most cognitively
interesting creatures on the planet. And we’re extremely flexible. At the very least, we know
that a human infant can grow up to be a competent adult in
any human culture of the world today and any human culture
that existed in prehistory. And that means
extremely varied– they’ve had to learn
extremely different things under different
circumstances and have succeeded at doing that. We also know that by the
time they start school, if they go to school at
all, the really hard work of developing a common-sense
understanding of the world is done. That is, it’s not explicitly
taught to children. Most of it isn’t
even very strongly implicitly taught to them in
the form of other people trying to get them to learn things. What you’re trying to do
when you have a young kid, as those of you who have them
know, or have had them know, is you’re trying to get them not
to climb off cliffs or explore the hot pots on the
stove and so forth. You’re really not spending
very much of your time trying to get them to learn new stuff. They’re doing that on their own. So it’s I think a really
interesting question, how do we do that? Intrinsically interesting
in its own right, even if it were of
no other use to us. But historically it’s
also been recognized as being really
important for efforts to understand the human mind,
understand the human brain, and build intelligent machines. So Helmholtz, who came up
in Eero’s talk last night, was not only a brilliant
neurophysiologist and a physicist, he was
extremely interested in perception and cognition. And he wrote about
fundamental questions about human perceptual
knowledge and experience. How is it that we experience
the world as three-dimensional? He concluded that we
didn’t know the answer and never could know
the answer, unless we could find ways to do systematic
experiments on infants of the sort that could already
be done to reveal mechanisms of color vision, for
example, as were described last night on adults– systematic psychophysical
experiments on infants. But he looked at
infants and said, I don’t see any way to do that. We can’t train them to make
psychophysical judgments and so forth. But he was aware of
their centrality. So was Turing, who
in thinking ahead to how one might build
intelligent machines, suggested that one aim
to build a machine that could learn about the
world the way children do. And a side of the
work that’s come up so many times in the whole
Hubel-Wiesel tradition that started in the
late ’50s, I think one of the most exciting
and important developments within that field,
we’re not just focusing on the response
properties of neurons in mature visual
systems, but rather on the development
of those neurons and the effects of
experience on them. When you discover that you
get these gorgeous stripes of monocularly-driven
cells in V1, it then immediately became
really interesting to ask, suppose an animal were
only looking at the world through one eye? Or suppose they could look at
the world through the two eyes, but not at the same time,
or not at the same things at the same time? What would happen
to those cells? And there was gorgeous work
addressing those questions from the beginning. Now, that work has somewhat
receded from attention. I think that’s a mistake. I think that
there’s a great deal to be learned from those
kinds of studies now. And if I get nothing else
across over this time, I hope you’ll at
least get the idea that this is a field
worth following, looking at development in humans,
looking at development of perceptual and
cognitive capacities in animal models of human
intelligence as well. So more specifically,
I think there are three questions
about human cognition for which studies
of early development in general and in human
infants in particular can shed light on. Two of them I’m not going to
really be talking about today, except indirectly. One is the question,
what distinguishes us from other animals? We come into the world with
very similar equipment. But look what we do with it. We create these utterly
different systems of knowledge that no other animal
seems to share. What is it about us that
sets us on a different path from other animals? That’s question one. And the other question
I won’t talk about is– well, I’ll talk about
it a tiny bit, but not directly– is, where do
abstract ideas come from? It seems like we not only
develop systems of knowledge, but those systems
center on concepts that refer to things that
could never in principle be seen or acted on. Like the concept “seven,”
or the concept “triangle,” or the concept “belief,” or
ethical concepts and so forth. Abstract concepts
organize our knowledge. But since they can’t be
seen or touched or produced through our actions, how
do we come to know them? I think studies of
early development can shed light on that as well. But the question I want to focus
on today is the third question, and it’s the one that
Josh raised on Friday. How do we get so much
from so little as adults? As adults, you look at
one of the photographs he showed of just
an ordinary scene and you can immediately
make predictions about, if you were to bang
it, what would happen? What would fall? What would roll? We seem to get this
very, very rich knowledge from this very, very
limited body of information at any given time. And what that
suggests is that we are able to bring to
bear in interpreting that scene a whole
body of knowledge that we already have about
the world and how it behaves. But that raises the question,
what is it that we know and how is our
knowledge organized? What aspects of the world do we
represent most fundamentally? Which of our concepts
are most important to us and generate the other
concepts and so forth? How can we carve human
knowledge at its joints? And now this can be
studied in adults and you’ve seen a number
of examples of this. You saw it in Nancy’s
talk last Tuesday, right? Anyway, last week sometime. Studies using
functional brain imaging to get at our representations
of human faces. You saw it in Josh’s talk. He was mostly using
data from adults to be probing the knowledge
of intuitive physics that he was focused on and that
his computational models are trying to capture. You’re going to see
it on Thursday in– no, tomorrow in
Rebecca Saxe’s talk, where she’ll talk about human
adults’ attributions of beliefs and desires and other
mental states to people. It’s certainly
studyable in adults, but it’s difficult to
answer these questions. It’s difficult to answer these
questions in any creature, but I think it’s
especially difficult to answer these questions in
adults for a couple of reasons. One is that our knowledge
is simply too rich. By the time we get to be
adults, we know so much and we have so many
alternative ways of solving any
particular problem, that it’s a real
challenge to try to sift through
all our abilities and figure out what the really
fundamental, most fundamental concepts that we have are. And the second
problem with adults is we not only know too
much, we’re too flexible. We can essentially relate
anything to anything. We can use information
from the face to answer all sorts of
questions about the world. And here, I think, infants are
useful for a maybe seemingly paradoxical reason. They’re much less
cognitively capable. They know much less about
the world and they’re far less flexible– I’ll show you examples of this– far less flexible in
the kinds of things that they can do with the
knowledge that they do have. Nevertheless, they seem to
come into the world equipped with knowledge that
supports later learning. And because it’s
supporting later learning, it’s being preserved
over that learning. It’s being incorporated
in all of the later things that we learn. So it remains fundamental
to us as adults. And I think this can
help us, to think about how our own knowledge
of the world is organized. OK. So that’s a general overview. How do we study infants? Now here’s where the
tables turn radically. We have way better methods for
studying cognition in adults than we do in infants,
just as Helmholtz thought. They can’t talk to us. They don’t understand
us when we talk to them, so we can’t give them structure. Oh, and unlike willing
trained animals, you can’t train
them to do things, at least not in
any extended sense. They can’t do much. I’m most interested in
infants in the first four months of life before they
even start reaching for things, much less sitting up by
themselves or moving around. The interesting thing is,
from day one, from the moment that they’re born, they’re
observing the world. They’re looking at things and
they’re getting information from what they see. Now, their observations– we’ve
learned over the last half century or so that their
observations are systematic and they’re reflected in very
simple exploratory behaviors, like when a sound happens
somewhere in the visual field, turning the head and
orienting it toward the sound. Even newborn infants
will do that. Or if something new or
interesting is presented, infants will tend to look at it. And these behaviors
I think can tell us something about what
infants perceive and know. And before getting to the
real substance of what I want to focus on
today, let me give you a few examples of this. What kinds of things
do infants look at? Well, if you present
even a newborn infant– infants at any age, really– with two displays
side by side, and vary properties of those displays
and the relation between them, you’ll see that they look at
some things more than others. So they’ll look at
black-and-white stripes more than they’ll look at
a homogeneous gray field. That’s useful. It allowed people to
get initial measures of the development of visual
acuity which infants– it actually overturned
a somewhat popular view that at birth, infants
couldn’t see at all. We know from these simple
studies that they can. And we also know that their
acuity starts out very low but gets pretty good
by the time they’re four to six months of age. It doesn’t reach full adult
levels until about two years. We also know that they
look at moving arrays more than stationary
arrays, and they look at
three-dimensional objects more than
two-dimensional objects. In addition to having
intrinsic preferences between different
things, they also have a preference for looking
at displays that change or displays that
present something new. So jumping from the ’50s
when those first studies were done up to the ’80s,
there was a whole flurry of studies showing babies pairs
of cats on a series of trials and then switching
to a cat and a dog. And the babies
would look longer– these are
three-month-olds– would look longer at the dog than
at a new example of a cat. So they’re able to
orient to novelty. And they also look
longer at a visual array that connects in some way
to something they can hear. Now, one of the things I spend
a lot of my time studying is foundations of mathematics–
numerical and spatial cognition in infants. I’m not going to talk
about it at all today. But I kind of couldn’t resist
giving just one example of looking at what you hear that
connects to infant sensitivity to a number. This is a study that
was conducted in France by Veronique Izard
and her colleagues with newborn infants in
a maternity hospital. She played infants
sequences of sounds, and each sequence involved
repetitions of a syllable. For half the infants, each
syllable appeared four times. For the others, it
appeared 12 times. And for the ones for which
it appeared four times, each syllable was
three times as long. So the total duration
of a sequence was the same for the
two groups, but one involved four syllables
and one involved 12. And after they heard
that for a minute, the sound continued to
play and now she showed, side by side, an
array of four objects versus an array of 12 objects. And the babies tended to
look at the array that corresponded in number to
what they were hearing. Now, all of this gives us
something to work with, but it raises a nasty problem. And the problem
is, what are babies perceiving or understanding? Today we’re not
going to be asking, how can babies classify things? What do they respond
to similarly? What do they respond
to differently? We’re going to be asking, what
sense do they make of them? What are they representing? What the content of
the representations that they’re forming
in each of these cases? And these studies as I’ve just
described them don’t tell us. Let’s take the case of the
sphere versus the disk. When this study was
first conducted, the author concluded that
babies have depth perception, that they perceive
three-dimensional solid objects. Is that a justifiable
conclusion? AUDIENCE: Not necessarily. ELIZABETH SPELKE: Why not? AUDIENCE: Because they
are not [INAUDIBLE].. ELIZABETH SPELKE: Yeah. OK. So when you present things
that differ in depth, you’re presenting a host of
different visual features that for us as adults
are cues to depth. The question is, are they
cues to depth for the infant? And the fact that the
infant is looking longer at something we would call a
sphere than at something we would call a disk, doesn’t
tell us whether they’re looking longer because they’re
thinking, “sphere,” or “3D,” or “solid,”
or something like that, or whether they’re looking
longer because they’re seeing a more interesting
pattern of motion as they move their head around,
or because as they converge on one part of the array they’re
getting interesting differences in how in-focus different
parts of it are, and so forth. All of the different
cues to depth could– what we want
to know is, what’s the basis of this preference? And the existence
of the preference doesn’t tell us that. Similarly for the
cats, and similarly for this single
isolated experiment that I gave you
on number, right? Does this say anything
whatsoever about number, or could there be
some sensory variable where there’s just more going
on in a stream of 12 sounds and there’s more going on
in an array of 12 objects, and babies are matching more
with more, independently of number? These studies in
themselves don’t tell us. In order to find out,
what we need to do is take these methods and
do systematic experiments. And these experiments work best
under the following conditions. When you’re studying a
function that exists in adults and whose properties have been
explored in adults in detail systematically, when you have
a body of psychophysical data that you can rest on in
your understanding of what’s happening in adults, and you
can then apply that to infants. So one example of that
took as its point of– this is work by Richard Held,
a wonderful perception psychologist who worked at MIT. Still is active, actually. He’s retired but still active. And he did these
beautiful experiments that started with the
sphere-versus-disk phenomenon. And first of all, he
tried to take it apart and say, let’s just focus
on one cue today, OK? Binocular disparity at the
basis of stereo vision. So he put stereo
goggles on babies. These were babies ranging in
age up to about from birth to about four months, I think. He put stereo goggles on
them and showed them, side by side, two arrays of stripes. In one of the arrays, the
same image went to both eyes. In the other arrays,
the edges of the stripes were offset in a way
that leads an adult to see them as organized in
depth– some stripes in front of others. And he showed that infants
looked longer at the array with the disparity-specified
differences in depth than the array where it didn’t. He did not conclude
from that that they have depth perception,
but it gave him a basis for doing a whole series
of experiments that asked, in effect, do you
see this effect under all and only the
conditions in which adults have functional stereopsis? So he showed, for example,
that if you rotate the array sideways 45 degrees
so that you still have double images
on the stereo side, but we wouldn’t see depth
because our eyes are side by side, not one above the
other, the effect goes away. He varied the
degree of disparity and showed that you
only get this preference within this narrow range where
we have functional stereopsis. And he was able to show the
striking continuity between all of the properties
of stereo vision in adults and in these infants. So that study and a bunch of
others using other methods, I think have resolved this
question of whether depth– when depth perception begins. Its beginning very early. Stereopsis comes in around
two to three months of age. Other depth cues
come in at birth. It’s beginning very, very early. But it didn’t come from
single experiments. It came from systematic
patterns of experiments. In the case of cats versus
dogs, we don’t really have a psychophysics
of cat perception, but steps have been taken to
try to get to what the basis is of infants’ distinction
between dogs and cats in those experiments. And interestingly, what’s
popped out are faces. Turns out, you can occlude the
cat and the dog’s whole bodies, and if you leave their
faces, you get these effects. If you occlude their faces
and leave their bodies, you mostly do not,
unless you cheat and give other obvious features, like
all the dogs are standing and all the cats are sitting,
or something like that. But in the normal case,
faces are coming out as an important ingredient
of that distinction. In the case of abstract
number, there’s also a lot of work in
adults on our ability to apprehend at a glance
approximate numerical value of sounds in a sequence
or visual arrays. We’ve learned a lot about the
conditions under which we can do that and the conditions
under which we can’t. That’s not my topic for today,
but Izard and her collaborators have been testing for
all of those conditions in newborn infants. And so far, so good. It looks like there
is a similar alignment between the patterns of– the factors that influence
infants’ responses in those studies
where they hear sounds and see arrays of
objects and the factors that influence our abilities to
apprehend approximate number. OK. So this gives us some good
news and some bad news. The good news is that
I think questions about the content of infants’
perception and understanding of the world can be addressed. The bad news is that we
can’t do it very fast. You can’t do it with a single
silver-bullet experiment. You have to do it with a
long and extensive pattern of research. In the past, research on infants
has gone extremely slowly. Basically, the
methods that we have allow you to ask each baby who
comes into the lab maybe one, or if you’re lucky, a
couple of questions, but not more than that. So it takes a long time
to do a single experiment. I do think, though,
that this work is poised to accelerate
dramatically and that we’re poised to– this is a good time to be
thinking about infant cognition because I think we’re soon going
to be in a different world, where we can start asking
these questions at a much more rapid pace. That’s for at least two reasons,
both of which, by the way, are being fostered by the Center
for Brains, Minds and Machines and undertaken by people
who are part of that center. One is, there are now efforts
underway to be able to test infants on the web. These basic simple
behavioral studies, you can assess looking
time using the webcam in an iPad or a laptop, and
you can test babies that way. And there’s attempts
to do that, which would make it possible
to collect data doing the same kinds of
experiments that have been done in the past,
but much more quickly. Two, as Nancy already
mentioned and Rebecca may talk about
tomorrow, there are efforts underway to use
functional brain imaging to get at not only
what infants look at, but what regions of the brain
are activated when they look at those things,
which will give us a more specific signal of
what infants are attending to and processing, someday,
hopefully, in the near future. And we just had a
retreat of CBMM, where there was a
lot of brainstorming about new technologies
to try to get more than just simple looking
time out of young babies. So maybe some of that
will work as well. But what I want
to focus on today is that even this
slow, plodding research has gone on for long
enough at this point that I think we’ve learned
something about what infants perceive and what they know. And I tried to put what I think
we learned into two slides. Here’s the first one. I think that very
early in development, baby in the newborn
period, but anyway, before babies are starting
to reach for things and move around on
their own, they already have a set of functioning
cognitive systems, each specific to a
different domain. One is a system for representing
objects and their motions, collisions, and
other interactions. Another is a system
for representing people as agents who act on
objects, and in doing so, pursue goals and cause
changes in the world. A third is a system
for perceiving people as social beings who can
communicate with, engage with other social beings
and share mental states. And then three other systems
that I won’t talk about today. One system of
number, which I think is being tapped in that
first Izard experiment. And two systems capturing
aspects of geometry, one supporting navigation of the
sort that Matt Wilson studies, the other supporting visual form
perception of the sort that IT and occipital cortex represent. I think each of these
systems operates as a whole. In Josh’s terms
from last Friday, it’s internally compositional. Infants don’t just come
equipped with a set of local facts about how objects
behave, they come equipped with a set of more general rules
or principles that allow them to deal with objects
in novel situations and make productive inferences
about their interactions and behavior. Each of these
systems is partially distinct from the other systems. It’s distinct in three ways. First, each of them operates
on different information. It’s elicited under
different conditions. Second, it gives rise to
different representations with different content. And third, most deeply, it
answers different questions. So for example, we
have two– infants have two systems for
reasoning about people, but each system is answering
a different question. The agent system is
answering the question, what is this guy’s goal? What is he trying to accomplish? What changes is he
affecting in the world? The social system is asking,
who is this guy related to? Who is he connected to? Who is he communicating with? Each of the systems are limited,
extremely limited relative to what we find in adults. Each captures only
a tiny part of what we as adults know
about objects or agents or social interactions. Each of them, I
think, interestingly, is shared by other animals. I didn’t expect
that to be true when we started doing this research. But as far as we can see so far,
it’s hard to find anything that a young human infant can do
that a non-human animal can’t. And I’ll give you
examples of that, too. And finally– and I won’t talk
about this much, unfortunately. I think each of these
systems continues to function throughout life
and supports the development of new systems of knowledge. So when we think thoughts
that only humans think, we engage these
fundamental systems that we’ve had since infancy
and other animals share. I also think this
research tells us something about how we do that. I think that in
addition to having these basic early
developing systems, we have a uniquely
human capacity to productively
combine information across these systems, and
through those combinations, to construct new concepts. I think these new
concepts underlie, or they tend to be
abstract, and they underlie a set of very important
later-developing systems of knowledge,
including knowledge that allow us to form
taxonomies of objects, of tools, of natural kinds like
animals and plants, and to reason about
their behavior, such that when we
encounter some new thing, we already know a lot
about the kind of thing that it is and can
use that to infer many of its specific
properties, and also to direct our learning
very explicitly to fill in the gaps in our knowledge. Another is the systems
of natural number in Euclidean geometry. Natural number, children seem to
construct over the first three to five years of life. Euclidean geometry seems to take
much longer, much, much later. Molly Dillon, who’s also
here, has been trying to work on understanding– and
so has Veronique Izard– how children go from
six years of age, where they seem absolutely
clueless about the simplest properties of Euclidean
geometry, to 12-year-olds who, whether they’re in the Amazon
and have never been to school, or studying geometry
in school, seem to have a basic rudimentary
understanding of points and lines and figures
on the Euclidean plane. A third is a system of
persons and mental states. And I won’t talk
about it, but I’m only talking for the first
half or so of this time, then Alia Martin’s
going to take over. And you’ll touch on– you’ll get to some
of those issues. Now, as Nancy said last week, I
have this out-there hypothesis that I don’t think anybody
else in the world believes, but I still believe it. That this productive
combinatorial capacity either is or is
intimately tied to what’s the most obvious cognitive
difference between us and other animals, namely our
faculty of natural language. In particular, I
think that there are two general properties
of natural language that make it an ideal medium
for forming combinations of new concepts. One is that the words
and the rules of– well, three
properties, actually. One is that the syntactic
and semantic rules of natural languages
are combinatorial and compositional. That is, if you learn
the meanings of words and you learn how
to combine them, you get the meanings of
the expressions for free. You don’t need to
go out and learn what a brown cow is if
you know what brown is and you know what a cow is. Second, the words and the
rules of natural language apply across all domains. They’re not restricted to
one domain or another the way infants’ other cognitive
capacities seem to be. So if you learn
how “cow” behaves in the expression “brown cow,”
and then you hear “brown ball,” or something that a different
domain of core knowledge would be capturing,
you can immediately interpret that
combination as well. And then the last thing about
natural language that I think makes it so useful for cognitive
development is that it’s learned from other people. And other people
talk about the things that they find useful
to think about, right? Word frequency is
a really good proxy for what the useful
concepts out there are. So a child who has a very
powerful combinatorial system that can create a
huge set of concepts is going to have a search
problem when they try to apply those concepts to the world. Something will
happen in the world. And if they now have
a million concepts that they could bring to
bear, which one are they going to use? Are they to test them all out? Having too many concepts,
too many innate concepts, would not necessarily
be a blessing. But if you use language to guide
you to the useful concepts, I think you’ll do better. The ones people are going
to talk about around you most frequently are going
to be the ones that it’s going to be most useful for you
to be learning at that point. So let’s go back to that first
set of questions, which is what I want to be focusing on today. And as I said, I’ll
talk particularly about three domains
where infants seem to develop
knowledge quite rapidly over the course of infancy. And I’ll spend most of my time
on the first one, objects. So object cognition
is really interesting and it seems to span
this really big range. It seems to involve many
different kinds of processes. If you’re going to figure
out what the objects are, what the bodies are
in a scene, then you need segmentation abilities. You need to be able to
take an array like this and break it down into
units, figuring out what different parts of that
array lie on the same object and what parts lie
on different ones. So early mechanisms
for doing that can participate in
object representation. But also to perceive
objects, arrays are cluttered and objects tend to be opaque. And when they are,
it’s never the case that all of the
surfaces of one object are in view at the same time. And it’s often the
case that you’re only seeing a little bit of any
given object at a time. Yet somehow we’re
able to see this as a continuous table that’s
extending behind everything that’s sitting on it, and even
sort of as a continuous plate, a single plate that’s partly– that’s on the table behind
the vase, and so forth. So to represent
objects, we’ve got to be able to take
these visual fragments and put them together in
the right sorts of ways. Something that’s harder
to show in a static image, but that of course is
radically true about the world is that our perceptual
encounters with objects are intermittent. We can look away
and then look back, or an object can
move out of view and then come back
into view, yet what we experience is a world
of persisting objects that are existing and moving
on connected paths, whether we’re looking
at them or not. And finally, objects
interact with other objects and we need to work
out those interactions. And the working out
that I’m interested in is not what this
little boy is doing, but what his younger
sister is doing as she’s sitting
in her infant seat and observing him
acting on these towers and wondering what’s
going to happen next. OK? At least that’s the problem
on the table for today. OK, so a standard view
for a very long time has been that different
mechanisms solve these different aspects of
the problem of representing objects. That segmentation
depends on relatively low-level mechanisms. Completion and
identity through time, it’s going to depend on how
much time we’re talking about and how complicated the
transformations are. They’re sort of in the middle. And this is all about
reasoning, about concepts that go beyond
perception altogether, like the mass of an
object, which we can’t see directly, and so forth. And I kind of believed
that that was true when we started doing this work. And because I did and wanted to
know where the boundaries were of what infants
could do, I started by working on these
problems here. And that’s what I’m going
to talk about today. But let me flag at the
outset that I no longer believe that the real
representations of objects that organize infants’ learning
about the physical world, I no longer believe
that they’re embodied in a set of diverse systems. I think there’s a single system
that’s ultimately at work here. Of course it has
multiple levels to it, including low-level of edge
detection, and so forth. But that there’s a single
system at work that both– that tells us what’s
connected to what and where the
boundaries of things are in arrays like this,
how things continue where and when they’re
hidden, and how they interact with other things. That’s one unitary
system, and I’ll try to show you what
evidence supports that view, though,
of course, jump in with questions or criticisms
or alternative accounts. OK, so here’s an intermediate
case to start with. You present a– it was studied
a lot by Belgian psychologist Elvin Meshot back in the 1950s,
I think– ’50s or early ’60s. Take a triangle, present
it behind an occluder, and ask babies, in effect, what
do you see in that triangle? Do you see a connected
object or do you see two separate visible fragments? We did these studies
with four-month-olds because they’re not
yet reaching for things and manipulating objects. We used the fact that
they tend to like to look at things that are new. So we presented this display
repeatedly– we, by the way, is Phil Kellman, now
at UCLA and studying all this stuff in
adults primarily, also studying mathematics now. Anyhow, so we presented
displays like this repeatedly to babies until they
got bored with them. And then we took the occluder
away and in alternation, presented them with
a complete triangle and with a triangle that
had a gap in the center. And we reasoned that there
were two possible outcomes of the study. Possibility one is that as
empiricists and the then-very influential child psychologist– developmental
psychologist Jean Piaget argued, for a four-month-old
infant who isn’t yet reaching for
things, the world is an array of visible fragments. So they will see
this thing as ending at this edge where the occluder
begins, and this display will look more similar to
them than this display, so they’ll be more
interested in that one. There was also the theory
from Gestalt psychologists and others that predicted the
opposite, that there would be automatic completion
processes that would lead any creature, whether
they were experienced or not, to perceive the simpler
arrangement, which is this one. Those, it seemed to us,
were the only two options. Baby research is
really fun because it can surprise you
even when you think you’ve covered all the bases. Neither of those
turned out to be true. What happened instead was that
when we took the occluder away, you still saw an
increase in looking both to the connected object
and to the separate object, and those two
increases were equal. Now, this could have been for
an extremely boring reason. Maybe babies were
only paying attention to the thing that was
closest to them in the array. So we very quickly tested for
that in the following way. Instead of contrasting an array
with a small gap to an array that had it filled in,
we contrasted an array with a small gap to an
array with a larger gap, too large to have fit
behind the occluder. And there, babies looked
longer at the array with the larger gap. So we know it’s not
that they’re not seeing this back form
and its visible surfaces, but they seem to be uncommitted
as to whether those surfaces are connected behind
the occluder or not. They don’t see them as ending
where the occluder begins, but they don’t clearly see
them as connected, either. And we showed that this
was quite generally true, both for simpler arrays and
for more complicated– well, for richer ones, like a sphere. We did this with a bunch
of different arrays. And under these
conditions, where the arrays are stationary,
that’s what we found. But there was one
condition where we got a different
finding, and that’s when we took one of
these arrays and moved it behind the occluder, never
moving it enough to bring the center into view,
but moving it enough such that the top and bottom
were moving together. And when we did that,
now babies looked longer at the display
that had the gap. That raised the question, why
is motion having this effect? And the immediate
possibility, we thought, is motion is calling their
attention to the rod, so they’re tending to it
more than they otherwise would, and it’s leading them
to see its other properties, like the alignment of its edges. So to test that, we gave them
misaligned objects differing in color, differing in texture. All of the edges–
none of the edges were aligned with each other. If motion was just calling
attention to alignment, it shouldn’t do
that in this case. But in fact, we found that
after getting bored with that, infants expected something like
this, not something like that. They looked longer at
the display with the gap. So it looks like the
motion is actually providing the information
for the connectedness, and the alignment is not
playing much of a role at all. Now, what could
be going on here? This is the kind
of thing I think that Josh likes to call a
suspicious coincidence, right? That an infant is
looking at this array, and isn’t it odd that
we’re seeing this– I’m seeing the same
pattern of motion below the occluder as
I’m seeing above it? Now that could be two
separate objects that just happen to be
moving together, but that would be
rather unlikely. You’re much more likely
to see a pattern like that if in fact there’s a between it
and it’s just one object that’s in motion. I think that’s
probably the right way to think about what’s going
on in these experiments. But if it is, notice
that not all coincidences that are suspicious for us
are suspicious for infants. For us, it’s a
suspicious coincidence that this edge is
aligned with that edge. For infants, it’s not. I think this is a
case where we can see infants can be
useful for thinking about our own
cognitive abilities because they seem to share some
of our picture of the world, but not all of our
picture of the world. And that can be a hint as to how
that picture gets put together and how it’s organized. So what kind of motion? We’ve tried a bunch
of different ones. One of them is vertical motion. That’s interesting because
it’s also a rigid displacement or motion in depth. They’re both rigid displacements
in three-dimensional space. Actually, all of
these three are. But in this case, you don’t
get any side-to-side changes in the visual field. I think I animated this. Yeah. So this is kind of what
the baby is seeing. By the way, all of these studies
were done with real 3D objects and they had textures
on them, and so forth. They’ve also all
since been replicated in other labs using computer
animated displays, which we didn’t have– which weren’t available
back in the day. And you get the same
result. So I’m just doing cartoon
versions of them here, but actually babies
showed these effects across a range of
different displays. So there’s vertical motion. Here is motion in depth. Oh, and by the way, we’re not
restraining babies’ heads, so it’s not going to be
anything near as, like, simple uniform as
what’s at their eye, is what I’m showing here. And then rotational motion,
like that, around the midpoint. And what we found is that
babies used both vertical motion and motion in depth
about as well as they used horizontal motion to
perceive the connectedness of the object. They did not use rotary motion. So I know there’s a lot of
interest and projects focused on perceptual invariance. And I think there’s an
interesting puzzle here, and it’s one that Molly is
very interested in, in the work that she’s doing on geometry. These are all rigid motions. But somehow, rotation seems
to be a whole lot harder for young intelligent beings
to wrap their heads around than translation is– including translation in
depth or vertical translation. There’s something hard
about orientation changes. And in fact, I think they
remain hard for us as adults. If you think of things like
how the shape of a square seems to change if you rotate
it 45 degrees so it’s a diamond. It’s no longer obvious that
it’s got four right angles. There’s something about
orientation that’s harder than these other things. And I think we were
seeing that here. When an object– when a baby
is sitting still and a rod is moving behind
an occluder, it’s moving both relative to
the baby and relative to the surroundings,
which of those things matters to the baby? So Phil Kellman did the
ambitious experiment of putting a baby
in a movable chair and moving the baby
back and forth. In one condition, the baby is
looking at a stationary rod, but his own motion is such
that if you put a camera where the baby’s head is, you’ll see
the image of that rod moving back and forth behind the block. In the other condition,
the motion of the rod is tied to the
motion of the baby, so it’s always staying in
the middle of the baby’s visual field, but it’s actually
moving through the array. And it turned out that it’s– OK, so whether the baby
was still or moving didn’t matter at all. So if the object is–
sorry, I did these wrong. This should be still,
that should be moving. If the object is still, and
whether the baby is still or moving, it doesn’t work. If the object is moving– the diagram is right. It was just my
label that’s wrong. If the object is
moving, it doesn’t matter whether it’s being
displaced in the infant’s visual field or not. It’s seen as moving. Now, this isn’t magic. The studies are not
being done in a dark room with a single luminous object
where the baby wouldn’t be able to tell. There’s lots of surround–
it’s in a puppet stage and that is stationary. So there’s lots of information
for the object moving relative to its
surroundings in all of the conditions of this study,
and I’m sure that’s critical. But for the point of view
of the infant’s connecting of the visible
ends of the object, the question he’s
trying to answer is, is that thing moving? Not, am I experiencing movement
in this changing scene? Retinal movement. So if it’s the case that– what those last
findings suggest is that the input representations
to the system that’s forming objects out of
arrays of visual surfaces already capture a lot of the 3D
spatial structure of the world. This is a relatively
late process. And it allows us to ask, is
it even specific to vision? Would we see the
same process at work if we presented babies with the
task of asking, am I feeling– are two things that are
moving in the world connected? Or are they not, in areas
that I’m not perceiving? We can ask that in
other modalities. So we did a series of studies–
this is with Arlette Streri. We did a series of studies
looking at perception of objects by active touch. By taking four-month-old babies
and putting a bib over them. Now I said they can’t
reach for objects, but if you put a ring in a
baby’s hand, even a newborn’s hand, they’ll grasp it. So we put rings in
their two hands. And in one condition, the
rings were rigidly attached, although the array was set up
so that they couldn’t actually feel that attachment and
they couldn’t see anything, about the object,
anyway, because they had the screen blocking them. But as they moved one, the other
would move rigidly with it. In the other condition,
the two were unconnected, so they would move
independently. And after babies explored that
for– over a series of trials, and as in the other
studies, we then presented visual
arrays in alternation where the two rings
were connected or not. And found that in the condition
where they had moved rigidly together, infants
extrapolated a connection and looked longer at the
arrays that were not connected. In the case where they
moved independently, they did the opposite. Now, that doesn’t
tell us that there is a single system at work here. It could be that there
are, as Shimon, I believe, was saying yesterday
afternoon, there are redundancies in the system. You have different
systems that are capturing the same property. That’s still true. But here’s a reason to– we went on to ask not
only what infants can do, but what they can’t do. And I think it gives us
reason to take seriously the possibility that
there’s actually a single system at work here. What we did– I haven’t
pictured it here– is, instead of varying the
motion of the things, we did vary their motion, but we also
varied their other properties. So their rigidity. We contrasted a ring
that was made out of wood with a ring
that was made out of some kind of spongy,
foam-rubbery material– their shape, their
surface texture. Asking, do infants take
account of those properties in extrapolating a connection? Are they more likely
to think two things are connected to each
other if they’re both made of foam rubber
than if one of them is made of foam rubber and
the other is made of wood? We never found any effect
of those properties, just as we didn’t
in the visual case. So we see not only the same
abilities, but the same limits. And while that’s
not conclusive, I think it adds weight
to the idea that what we could be studying here– we
started in the visual modality. But what we could
be studying here is something that’s more
general and more abstract. Basic notions about how objects
behave that apply not only when you’re looking at things,
but when you’re actively– when you’re feeling them,
actively manipulating them, exploring them in
other modalities. So I put a question mark because
it’s not absolutely conclusive, but I think we should take
seriously that possibility. OK. Only motion. Is motion the only
thing that works? Or will other changes work, so
if an object changes in color? We created a particularly
exciting color change by embedding colored
lights within a glass rod so it’s flashing on and off. Succeeded in eliciting very
high interest in that array. Babies looked at
it for a long time, but only the motion array
was seen as connected behind the occluder. So it looks like not all
changes elicit this perception. It’s an open question what the
class of effective changes is. Maybe it’s broader
than just motions, but it doesn’t seem
like all changes work. Finally, is motion
the only variable that influences
infants’ perception of– the only property of
surfaces that influences infants’ perception of objects? The answer to that
seems to be no. So we studied that in
a different situation for which this is just a
very impoverished cartoon. We took two block-like objects– of different colors and
textures in some studies, same color and
texture in others. It didn’t matter. And put one on top of the
other and either presented them moving together
or moving separately. And then tested whether babies
represented them as connected in either of two ways. Some of the studies
were done with babies who were old enough to reach. And then we could ask,
are they reaching for it as if it were a single
body or as if there were two distinct bodies there? I could give you more
information about that if you’re interested. The other was with
looking time, where we had a hand come out and
grasp the top of the top object and lift it. And the question is,
what should come with it? Will the bottom object
come with it as well or will the top
object on its own? When the things had
previously moved together, they expected it all
to move together. When they’d moved
separately, they expected only the top
object would move by itself. And when there was
no motion at all, findings vary somewhat
from one lab to another, but mostly they tend to be
ambiguous in the case where there’s no motion. So there it looks like
motion is doing all the work. But if you make one simple
change to this array that you can’t do in
the occlusion studies, you simply change the
size of this object and present it such that there’s
a gap between the two objects. And you can either do it with
this guy floating magically in midair, or you can do
it with two objects side by side, both stably
supported by a surface. If there’s a visible
gap between them, the motion no longer matters. They will be treated as
two distinct objects, no matter what. So what I think is
going on here is that babies have a
system that’s seeking to find the connected, the
solid connected bodies. The bodies that are
internally connected and will remain so over motion. And that’s what’s
leading them to see these patterns of relative
motion or these visible gaps as indicating a place where one
object ends and the next object begins. I did want to get on to the
problem of tracking objects over time, perceiving not what’s
connected to what over space, but what’s connected
to what over time. Under what conditions are
the thing that I’m seeing now the same thing that I
was seeing at some place or time in the past? So conceptually, it feels like
continuity of motion over time is related to connectedness
of motion over space. And it’s been tested for
in a variety of ways. Here’s one set of
studies that we did, where we have an
object that moves behind a single screen, and then
either is– and it starts here, ends up here. And either is seen to move
between the two screens or is not. And we ask babies
in effect, how many objects do they think
are in this display, by boring half the
babies with this, half the babies with that,
and then presenting them in alternation
with arrays of one versus two objects,
neither of which ever passes through the
center, but the arrays differ in number. In the one case, it’s
either moving over here or it’s moving over there
on different trials. And what we find is
that in this case, they expect to see one object
and look longer at two. In this case, they
expect to see two objects and look somewhat longer at one. There’s actually an overall
preference for looking at two, but you get that
interaction and there’s a slight preference for looking
at one in that condition. Providing evidence,
I think, that babies are tracking objects over
time by analyzing information for the continuity of– or discontinuity of
their object motion. Now, Lisa Feigenson
has conducted stronger tests of this, I think,
with somewhat older babies. When babies get older
and they do more, you can do stronger tests. So these are babies
who are old enough to crawl, old enough
to eat, and old enough to like graham crackers. So she puts the baby back here,
and in one set of studies, she takes a single graham
cracker, puts it in one box, and then takes two graham
crackers, one at a time, and puts them in the other box. And then the baby, who’s
being restrained by a parent, is let loose. And the question is,
which box will they go to? And they go to the box with
the two graham crackers. My favorite study, though,
in this whole series was one that she and Susan
Carey ran as a boring control condition. I think it’s the most
interesting of the findings, though. In the boring control
condition, they were worried about the
fact that maybe babies are going to the box
with two because they see a hand around that box
for a longer period of time, doing more interesting stuff. So they did the
following boring control. The two condition was
the same as before. So a hand comes out with
a single graham cracker, puts it in the box,
comes out empty, takes a second graham cracker,
returns with a second graham cracker, puts it in the box. In the other condition,
the hand comes out with one graham cracker,
puts it in the box, comes out again with
the graham cracker, and then goes back into the
box with that graham cracker. So you’ve got more graham
cracker sightings on the left. You’ve got a same amount of
hand activity on the two sides, but the babies go
to the box with two. They’re tracking the graham
crackers, not the graham cracker visual encounters. They’re tracking a
continuous object over time. Finally, objects. Scenes don’t
usually just contain a single object that’s either
connected, continuously visible or not, or connected or not. They contain multiple objects
and those objects interact with each other. Shimon talked yesterday
afternoon about the evidence that babies are sensitive
to these interactions, at least down to about six
months of age in the conditions he was talking about. In slightly
different conditions, the sensitivity has
been shown as young as three months of age. Basically, here’s a paradigm
that will show that, if you have a single object
that’s moving toward a screen. Another object is stationary
behind the screen. But at the right time, the
time at which this object, if it continued moving at
the same rate, at the point where it would
contact that object, this object starts to move
in the same direction. And now, after seeing
that repeatedly, the screen is taken
away and babies either see the first object contacting
the second and the second one immediately starting
to move, or they see the first object
stopping short of the second an
appropriate gap in time, and then the second
object starts to move. And they look longer
at this display, providing evidence
that they inferred that the first object contacted
the second at the point at which it started to move. Interestingly, as in the case
of the occluded object studies, if instead of having
the second object move, you have it change
color and make a sound, so it undergoes a change
in state, but no motion, the babies no longer infer
contact in this condition. They are attentive
to those events. They watch them a
lot, but they’re uncommitted as to
whether that first object– this is work of Paul
Muentener and Susan Carey relatively recently. It wasn’t done
with cylinders, it was done with a toy
car that hits a block, I think, or doesn’t
hit the block. They’re uncommitted as to
whether the car contacted the second object or
not, if the second object changes state but doesn’t move. Returning to the case
where they succeed– namely, this thing went behind a
screen, the other thing started to move, infants inferred
that they came into contact– that begins to suggest that
maybe babies have some notion that objects are solid,
that two things can’t be in the same place
at the same time, that when one moving thing
hits another thing, one or the other of them or both,
their motion has to change, because they’re not going
to simply interpenetrate each other. And Josh already
very briefly pointed to some very old studies
suggesting that babies have– make some assumption that
objects are solid as early as– I think in the
earliest studies done with babies it’s about two
and a half months of age. These are these studies
that Renee Baillargeon did that start with simply
a screen, a flat screen, rotating on a table, rotating
180 degrees back and forth on a table. Then she places an
object behind this wall. The screen is lying on the
table with its back edge right here at the middle. She places an object behind
it, and then the screen starts to rotate up
around the back edge and the question to the
infants in effect is, what should happen
to that screen? And the two options
she presents to them is it either gets to the
point where it would contact this object which is
now fully out of view, and stops, and then returns
to its first position, which is a novel motion,
but consistent with the existence,
location, and solidity of that hidden object. Or it continues merrily on
its way and the same pattern of rotation as before. When it does that,
of course, it’s going to come back
flat on the screen and there’s not going
to be any object there. If there had been
an object, it would have had to be compressed. Or what I think actually
went on in those studies, it was quickly and
surreptitiously knocked out of the way. And infants looked less at
this event than at this one– this one, sorry–
providing some evidence that they were representing
these objects, both as existing when they were
out of sight, and as solid. So this is just a summary,
not a claim about knowledge development, about– I’m attempting to
characterize here with motion over just one
dimension of space and time, how infants seem– what infants seem to represent
about the behavior of objects. Namely that each object
moves on a continuous path through space and over time. That it moves cohesively. It doesn’t split into
pieces as it’s moving. So if you’ve seen
something move like this, then you find it unlikely
that if this were lifted, it would go on its own, and
you look longer at that. There is no merging,
where two things that previously moved independently
now move together. So after looking at
this, it would also be unlikely, if you lifted
this, for the whole thing to jump up at once. They move without gaps. They move without
intersecting other objects other objects on their paths
of motion, such that two things are in the same place
at the same time. And they move on contact
with other objects and not at a distance from them. So that’s just a
summary of what I think these studies show about
four-month-old infants, not newborns. They also show that infants’
perception of objects is really limited. There’s all these situations
under which we see unitary, connected, bounded
objects when they don’t. And interestingly, research
by Fei Xu and Susan Carey shows that even when you present
really quite surprisingly old infants, 10-month-olds, with
objects that should be really familiar to them, like
toy ducks and trucks, they don’t assume that these
two objects will be distinct if they undergo
no common motion. If they’re simply
presented stationary, the babies seem uncommitted
as to whether there’s a boundary between them or not. So they’re using very
limited information to be making these basic– building these basic
representations of what’s connected to
what, where one thing ends and the next begins. Now, this changes very
abruptly between about 10 and 12 months of age. They start treating those
as two separate objects, whether they’re moving
together or stationary or not. Now, infants’
tracking of objects shows very similar limits. So I told you they
succeed in perceiving– representing two
distinct objects in a situation like this. But up until and including
10 months of age, they fail in this situation. If a truck comes out on one
side of a single large screen, so you’re not
getting information for the motion
behind that screen, and a duck comes out
on the other side, and you ask babies, in effect,
how many things are there? One or two? By removing the screen
and alternately presenting those two possibilities,
they are uncommitted between those two alternatives. In this situation as
in the previous one, there’s this very abrupt
change between about 10 and 12 months of age. And I can’t resist saying,
even though I’m way over time, that Fei Xu has shown that that
change is interestingly related to the child’s developing
mastery of expressions that name kinds of objects. So she’s been able to show,
for example, that if you simply ask for individual infants, when
did they start succeeding here, their success is predicted
by their vocabulary as reported by parents. She’s also shown that
if you take a younger infant who would be slated– destined to fail this study,
but as you bring objects out on the two sides, either
familiar ones or novel ones, starting at about
nine months of age, if you name them and you give
them distinct object names, they now infer two objects. And in fact, they’ll
even do it if the two things you bring out from
behind a single wide screen look the same. If you bring one thing out
and say, look, a blicket, and put it back in, and then
bring something out and say, look, a toma, even
if it looks the same, they’ll infer two objects. So there seems to be
this change that’s occurring at the end
of the first year quite dramatically that’s
overcoming this basically meant that we’re seeing earlier on.

Leave a Reply

Your email address will not be published. Required fields are marked *