Laura Schulz: Cognitive Development and Commonsense Reasoning, Part 1
June 4, 2014
June 4, 2014
All Captioned Videos Brains, Minds and Machines Summer Course 2014
Topics: Historical perspective on underestimating the challenge of commonsense intelligence in AI; studying children may provide key insights; early representations of objects (e.g. object permanence, Spelke objects, expectations of object behavior), causality, agents and goals (e.g. discriminating objects vs. agents, intentional actions), and learning; drawing rich, abstract inferences from sparse data is critical to rapid learning
LAURA SCHULZ: So today I'm going to begin by introducing what I think is the fundamental problem of cognitive science in the 21st century. And it's the problem of commonsense reasoning. Right now, I'm violating your commonsense reasoning. And I'm violating it in a particularly trivial way. There's nothing physically impossible about setting a lecture from the back of the class. [INAUDIBLE] But by messing with your commonsense reasoning, the kind of knowledge you may not know that you had until it's violated. I've messed with all kinds of things. Their ability to hear me and across and in that direction.
Your ability to categorize me, your ability to make predictions and explanations, even your basic ability to orient your attention. It's just a really, really fundamental set of background knowledge. I did it [INAUDIBLE] and I think it is one of the super hard, complicated central problems. I thought that problem had been routinely underestimated.
Most of you are familiar with this gentleman, Alan Turing. He's [INAUDIBLE] computers, right? He also brought up the very important types of intelligence, that [INAUDIBLE] the idea of the Turing test, the idea that one way to know repeated and really understanding and engineering intelligence would be if you could design a computer that could fool a person into thinking it was a person. So the Turning test, as many of you know, [INAUDIBLE] to have a person back here and a computer back here and someone out here who is typing questions and getting answers back. And if the person, you know they couldn't tell which was which, then you could justifiably say that you'd achieved something like artificial intelligence.
And what Alan Turing believed was that this was a pretty tractable problem. In 1950, he thought that 14 years ago, we would have nailed this problem, that computer-- people would have about 70% of making the right identification. And as we know, we're still facing that problem.
Barbara Minsky had a similar reaction. So within a generation, the problem of creating artificial intelligence wasn't substantially solved and a few years later, [INAUDIBLE] one of the hardest they solved. There's a reason why [INAUDIBLE] As many of you may also know, no computers yet pass the Turing test. None have been reliable [INAUDIBLE]. So what's the big advancement of [INAUDIBLE] which is the computer that did the best job. It's never actually confused with a person, but it does better than all the other computers.
Let me show you what some of the recent winners looked like and why they don't actually confuse interrogators about their humanness. Here's a little transcript from a winner of the past.
Yes, what's your favorite novel?
There's no such thing.
Meaning you don't have a favorite novel or novels don't exist?
What's your favorite book then?
Are you serious.
Et cetera. OK. So there's another reason why not.
Do you have a pet?
No, but I like cats very much.
Why don't you have a cat?
Do I need one?
Only if you think you do.
I think I do.
Well then get one, by God.
Good reasoning! Where can I get one by God?
It's interesting, that they haven't solved the problem. Because [INAUDIBLE] wanted engineer commonsense reasoning. Well, how would you do it and when? [INAUDIBLE] Like we have all this time to live in the world. You have bodies. You have all his information brought in and computers just have what we call [INAUDIBLE]. And maybe they just don't know enough. They don't know about the world. Let's just give it a lot more facts about the world.
And you might be troubled by this, because facts seem somewhat uncomfortable. Suppose I want to know that Barack Obama is in Washington. And now I want to know if his knee is in Washington or his right toe is in Washington or his second toe or his third toe. Or [INAUDIBLE] But the philosopher, Hubert Dreyfus, was on my dissertation committee at Berkley, was deeply skeptical. He said, if you got all that knowledge into a computer, you would not know to retrieve it. It would just be a [INAUDIBLE]
And Hubert Dreyfus was wrong, right? Many of you may have heard of this. Let's see if it plays. This is my [INAUDIBLE]
LAURA SCHULZ: [INAUDIBLE]
LAURA SCHULZ: [INAUDIBLE]
Here we go. Jeopardy Many of you hopefully saw this. Can I just see a show of hands if you watched Watson beat [INAUDIBLE] Jeopardy champion? Great. I'm going to go in a little bit of the way. Actually [INAUDIBLE]
ALEX TREBEK: [INAUDIBLE] In May of 2010, 5 paintings worth $125 million by Robert Matisse and three others left Paris's Museum of this art period.
LAURA SCHULZ: [INAUDIBLE]
What I really want to show you guys 2 seconds later here, because I want to focus on--
ALEX TREBEK: --lost the same amount. But Watson retains control of the board.
LAURA SCHULZ: He's killing it. [INAUDIBLE]
ALEX TREBEK: Answer-- hedgehogs are covered with quills and spines, which are hollow hairs made stiff by this protein. Roger?
ROGER CRAIG: What is keratin?
LAURA SCHULZ: I wanted to see if he gets it right, but I also want you to see this next step gets the guy. This is where I want to stop it. Watson was a remarkable contestant. Watson handled natural language processing. Watson [INAUDIBLE] problem is not necessarily a problem. It turns out that this computer as able to understand natural language, take access to an enormous amount of facts and spit it out in real time faster than [INAUDIBLE] So we can solve that problem. But he [INAUDIBLE]
And the hard problem may be that this kind of knowledge, the ability to access all that real knowledge is neither necessary or sufficient for passing the Turin test. You could actually have all this information and not be recognized as a human. And that is probably because it did recognize as a human to make an error of this kind, right? You may not know that keratin is what hedgehog spines are made of.
But you surely know that porcupine isn't even in the ballpark of the kind of answer that we should know. Right? It's related. It's a relevant fact, but it's not an answer to that question. You structure your knowledge and your representation of something. You may not have as much of it, but it's not organized well.
Now, I want to examine contracts. I want to pick on the kind of organism that not only passes the Turing test, it solves other problems [INAUDIBLE]. It solves problems of better planning, of proper reasoning, of number understanding, of spatial navigation, of theory of mind, of natural language acquisition, of moral reasoning. It does all of this in unsupervised learning, which is about the organisms, not the learning. [INAUDIBLE] is what? And that is young children.
There are a lot of things young children don't do. They're not going to be allowed to drive cars. We'll have self-driving cars. We're not going to be having four-year-olds driving them. Young children are really bad at tests. They're extremely bad at Jeopardy. They can't even remember to ask the question in the form of a question.
But young children can answer this and this and this. Everything that you needed to know to pass the Turning test, you learned before kindergarten. They have structured knowledge. They have structured representations. If you put them behind there, you might not mistake them for a grownup, but no one is going to mistake them for a computer.
Again, the kind of things are effortless for our children are exactly the kind of things we don't have the AI machine learning tools at all, and that's of course, the challenge we want to hand back to you all today today. Why can computers can play Jeopardy but they can't parse natural [INAUDIBLE] and tell you what's going on and [INAUDIBLE] answer infinitely many generated questions about it sensibly, because it's [INAUDIBLE]. The answers may be wrong, but they'll be [INAUDIBLE].
So understanding commonsense, that's the hard problem of cognitive science. To understand that problem, solve that problem, we're going to need to understand two kinds of things. We're going to need to understand the structure of children's early representation [INAUDIBLE] and we're going to need to understand the processes that support inferential learning. And the way I understand we're going to organize this course is that today I'm going to talk a little bit about the structure of the earliest representations and we're going talk more on Friday about [INAUDIBLE].
So my job today isn't about [INAUDIBLE]
It doesn't represent the structure of children's early representations [INAUDIBLE] in the next hour or so. This is going to be extremely haphazard. I mean, I apologize for that in advance. What I want to do is give you a feel for the fact that they exist, that there are rich, structured representations very early on. Some of that knowledge is possibly innate. Other of that knowledge is early developing that we learn. People talk about it as being structured like a [INAUDIBLE] knowledge that supports prediction categorization explanation [INAUDIBLE]
And I'm going to try to give you some of the highlights, sort of best of and careful data from developmental technology to show a little bit of each of these areas. And then I'm just going to very briefly, show some of the talk on Friday on complex learning. I'm not going to be getting through all this material. [INAUDIBLE]
Let's talk about the object mystery. [INAUDIBLE] makes clear why commonsense knowledge is a problem. And as I said, you only recognize it when it's violated. So why are objects a problem? So they're a problem because [INAUDIBLE] If you look around this room, you may think that you see objects and you may think that you see angels. But of course, you don't. You see a whole bunch of [INAUDIBLE] We operate on these [INAUDIBLE] We count them. We name them. We manipulate them. We represent spatial relations and causal relations among them. They are fundamental [INAUDIBLE] and you also have a lot of social evaluation of them. Right?
This is your computer, not his computer, your cup of coffee, not his cup of coffee. So how did we get there? And especially, what do you have inside to take you back in time 20, 30 [INAUDIBLE]
[INAUDIBLE] who has actually laid out a lot of this problem, a lot of this space and said, we need to understand, if we're going to understand cognition, we need to understand how children understand objects, agents, [INAUDIBLE], morality, and space.
Babies-- we're going to start with what people assume through generations, which would be the [INAUDIBLE] from the ground up, right? And in particular, he said babies don't have an object concept. They can represent what is immediately accessible to their sensory motor system, but when it goes [INAUDIBLE]. And he had a lot of evidence for this. In particular, this is a--you know, you can try this at home. Find a baby. Hand that baby a really exciting toy rattle. The baby will be enthralled. Remove the rattle from the baby's sight and if the baby is six months or less, it will be untroubled.
When you bring it back, it's just as excited. But when it goes away, the baby will simply [INAUDIBLE] If the baby is old enough to start reaching searching, if you hide the object, the baby may look at it. But if you keep hiding it in that place, and then [INAUDIBLE] move it to another place, the baby keeps going back to the initial place where it was hidden. It's not going to go and search to where you moved it. [INAUDIBLE]
And Piaget said that's because babies have a very unstable ability to represent abstraction at all. They can't have a representation for something that's not immediately accessible [INAUDIBLE]. You may have heard babies lack object permanence. It's still taught in preschools and in early childhood education programs over the country. It's false. It's patently false. Nonetheless, Piaget's experiment [INAUDIBLE]. So I want to talk a little bit about [INAUDIBLE]
Why do we know it's false? Why do we think that babies actually have very early abstract representations of objects? Actually, not just babies, but all kinds of animals? Chicks do it too. We have chicks raised in the dark. You can hatch them and you can show that they [INAUDIBLE]. Why do we know that babies-- [INAUDIBLE]
OK. What you see on the screen here is called a habituation event. You're seeing it at a funny angle here. What you're supposed to be seeing is [INAUDIBLE] a drawbridge going back and forth like this. OK. It goes all the way down and all the way up like that. [INAUDIBLE] And they're watching this. And a habituation event means an event that eventually bores the baby. They see this so often, again and again and again and again, that whatever their looking time in the first viewing trials, was the average dropped in half by the end of their trials. [INAUDIBLE]
Now all you need to do with the baby, is introduce something like an object that is behind this drawbridge right here. And you're going to show them one of two things. Either exactly the same event they just got bored by seeing, they're going to see that bridge collapse on top of that object and push it down. Now that's not possible if there's an object right there. You know that object has got to stop-- one solid object has got to stop another solid object. Babies have just been bored by this event, they've seen it a lot. But this is an impossible [INAUDIBLE].
You wouldn't expect them to recover interest. You would expect them to look longer, even though this is a familiar event, familiar forced event. This is a novel event. They've never seen that drawbridge stop halfway. But if you represent a hidden object, that is the expected event. That is what you should imagine is going to happen to that object. And what [INAUDIBLE] they do expect the drawbridge [INAUDIBLE] And in fact, not only that, they're extremely good at estimating about what that drawbridge should stop [INAUDIBLE] that drawbridge should be able to take. If you hand the baby an object first and let them play with it, they can find out if that object is extremely rigid or it's like a sponge and it's a little bit squishy, and they adjust their expectations accordingly.
So now they use their own firsthand knowledge of the object's properties to represent those properties when that object [INAUDIBLE] representation of that object and their properties in three-month-old babies. And again, as I say, we can now show that even in brand newborn, newly hatched chicks. So there's something quite rich and structured about [INAUDIBLE]
Here we need to pause and say, why do they occasionally get it wrong? [INAUDIBLE] thousands of manuscripts, monographs and books, and like I say, you try to set home with any baby, and you're [INAUDIBLE] The short answer is there are a lot of childhood development and psychology with performance competence that [INAUDIBLE] In this case, there are several fairly simple ones. So recognizing that something can disappear and appear and then figuring out what to do about it, I frequently find myself searching repeatedly in the same place, for instance, for my keys, even when I have just looked there, because I don't really know where else to search. I know it's not there and I know it's not actually there, and yet, again, [INAUDIBLE]
And in particular we know with babies, that the least little things can affect their ability to search for [INAUDIBLE] but if you have babies that have been searching repeatedly in one location, and then you move the object to the new location, if you let the babies-- you shift their posture, and then put them back down, they do a much better job in executing the new reach if their posture has been changed [INAUDIBLE] Some of it is just the motor [? separation ?].
Adele Diamond showed [INAUDIBLE] prefrontal cortical maturation, babies have trouble exhibiting abilities [INAUDIBLE] representation [INAUDIBLE] it's a performance. It's a performance [INAUDIBLE] And you can imagine this yourself, when, although you can presumably represent the position of that tennis ball moving forward toward you, you sometimes miss when you are swinging your racquet at it. It's sometimes hard to adjust your performance to the [INAUDIBLE].
Babies know that [INAUDIBLE] and they know some things about their properties. But what is an object for a baby [INAUDIBLE]. What are its core properties without which it wouldn't be an object at all. This is a question that Spelke [INAUDIBLE] really changed the view of the field and really much of what we know about it [INAUDIBLE]
And in particular, she investigated, quite early in the '80s some [? post ?] constraints on what you might know about objects that you might need to have in your [INAUDIBLE] In particular, she proposed that it's actually core to the fact that something is an object at all that you have continuity, right? If I'm going to toss something across the room, you expect it to go in a complete trajectory all the way to the back of the room. You don't expect it to start here and have an empty space between them and then end up at the blackboard. Things don't jump over [INAUDIBLE] In the physical world, there aren't worm holes in instant intuitive theory [INAUDIBLE]
She also said you should understand the [INAUDIBLE] as follows: that [INAUDIBLE] study. You don't expect two objects to occupy [INAUDIBLE] babies know everything. In particular, she thought babies [INAUDIBLE] gravity. [INAUDIBLE] know about inertia. And the argument which she based this on is an interesting one. She based her argument essentially from error.
She said that adults never make errors about continuity. They never make error or expect objects to pass through one another. But adults cannot predict to balls that you drop from a moving train. Adults, as you may know from some studies of intuitive physics are inclined to think that if a ball is going round and round in a tube like this, they are a little bit uncertain about what physical trajectory that ball is going to take when it pops out the [INAUDIBLE]
Here, [INAUDIBLE] adults can get confused and be kind of [? contagious ?] in a way that they never get confused about these types of
LAURA SCHULZ: Core in their sense, means a [? ungenetically ?] old and [? priogenetically ?] early representation that remains consistent with that development, even to elaborate it even if you get new knowledge, even if it contradicts--
LAURA SCHULZ: [INAUDIBLE] it depends arguably on evolutionary typical experience [INAUDIBLE] it might not develop, but it's triggering it, conceptual triggering. So in environmentally normal conditions, you would expect that knowledge to be present and they would have good evidence [INAUDIBLE] So their signature [INAUDIBLE] possibly a foreknowledge is that you can find it in a species. You can find it extremely early and you can find it signatures of it in adults who know that in some ways who should have richer knowledge or even object contradiction [INAUDIBLE]
So she [? learned how to ?] test these [INAUDIBLE] she could find evidence that even very young infants had this kind of knowledge. So again, you'll [INAUDIBLE] animation here. See here again, the familiarization [INAUDIBLE] Try to show you just one thing at a time, [INAUDIBLE] Again, they see the ball fall all the way to the ground, just like balls do. Right? And [INAUDIBLE] where you either see a novel [? instance ?]. The ball never ended up here.
And yet, you would, of course, expect it to. If objects are solid, they don't pass through things. But the ball has to end up here. Or you could [? use your ?] familiar instinct, where the ball goes all the way to the ground, which is just what they've been [INAUDIBLE]. But in this context, they [INAUDIBLE] of what [INAUDIBLE]
You see the baby's initially interested in the familiarization event, they get [? forged ?] to that event. And then you show them the consistent event or the inconsistent event [INAUDIBLE] with respect to that little bench being there, but when it's although superficially more familiar to the familiarization, it's [? in fact, ?] a violation if [INAUDIBLE]
So, what she also found is that in very similar events, superficially [INAUDIBLE] babies didn't show this kind [? of battering ?] behavior. So it seems equally obvious to us that it is a violation of how the ball stopped in midair. Like that does not happen unless you happen to be holding the ball. An unsupported object doesn't drop [INAUDIBLE] It's a violation of gravity. It's a violation of inertia. Once that object is in motion, if nothing stops it moving, here in our intuitive [INAUDIBLE] we expect that thing not to just pause in midair. We don't expect objects to totally change trajectory.
Babies [INAUDIBLE] Similarly, with violations of gravity, if you have that ball roll across a plane and you put a gap in the plane, this is certainly possible. It can roll under the gap. Or you can put it by this [INAUDIBLE] it's totally impossible that it should do that. But the babies didn't care that there was difference here.
Rene [INAUDIBLE] followed up on these kinds of studies and went on to develop some ideas about how babies would gradually begin to develop an understanding of what the constraints are on [? support motions ?] on objects. They know they're falling and they and they know [INAUDIBLE] but it takes time for babies to appreciate not just that objects need to be supported, but the kind of supports they specifically need.
So at six and one half months, babies are perfectly happy-- I'm sorry, five months-- are perfectly happy to accept that objects need support but they are indifferent to where on the object that support comes from. To understand why that's not crazy, look around this room and you'll see any number of objects that are actually affixed by a vertical support, like that clock over there. So they have lots of evidence that make it possible, that objects can be supported in other ways.
Then they recognize that the support should be underneath but they don't really care about much of a supporting surface there is and it relatively late in development that they really begin to understand the relationship between the mass of the object and the distribution of that to the support. And all of this, as Rene points out, goes along with babies active experience pf the world, the fact they are, at six months, starting to go to meet and manipulate objects in a competent way. They can sit up now. They can move things around themselves. They can experiment with stacking and with sorting. But they don't need that information to know that the object is a solid. They don't need to know [INAUDIBLE]
AUDIENCE: What do you see with animals [INAUDIBLE]
LAURA SCHULZ: Strikingly. Again, the--
AUDIENCE: What kinds of animals?
LAURA SCHULZ: The ones that I think make like Spelke the happiest are [? control ?] reared chicks because you can actually take a chicken and [? chip it down ?] to any visual experience. Right, they can hatch right out of the egg and then you can [? project ?] them into situation experiments. You can actually get chickens to collectively-- I want to show you one other thing about Spelke objects. And then I want to talk [? about pathology. ?] And you can get newly hatched chicks to collectively imprint on things they might classify as things that move independently as opposed to being caused to move by other objects. So I think that it's the newborn chicks [INAUDIBLE]
But if you think about it, they have lots of special animals. There are birds that hatch and fly. There are gazelles that are born and [? die ?] and mountain goats that go up mountains. It would [? be interesting ?] if you didn't have some sense of how you might build in innate [? sense ?] of what the physical world was really like. [INAUDIBLE]
AUDIENCE: I have a-- it's probably really dumb question here, but these experiments, are they done on a computer or do you have a [INAUDIBLE] that's mocked up to video them like--
LAURA SCHULZ: They're usually actually [INAUDIBLE]. So most of these experiments have been done live with real objects in front of real babies.
LAURA SCHULZ: I'm sorry?
AUDIENCE: If you do everything on a computer screen, would that change the results?
LAURA SCHULZ: I believe the experiment that has been done on chicks had been done on a computer screen. Many of these things have been switched to [INAUDIBLE] The main thing is you tend to [? depress ?] it, in terms of looking overall, if they're not looking at real world objects. They're actually more interested in real world objects than the screen, depending on what stimulates them. Since you want to get as much out of your audience as you can, That may be changing as our computer animation [INAUDIBLE]
Let's talk about causal reasoning. If you've spent any time around a baby or shopping for baby gifts, you can find all kinds of toys like this. And they're going to tell you, to make claims like these, this toy is going to increase your baby's cause and effect understanding. This toy here is going to also help your baby learn about cause and effect.
All these marvelous toys-- what has always struck me about these toys is that their the computer products of our age. We can-- in 1970, no toys would have been marketed like this. You would have had the same kind of baby toys. You would actually call them basically rattles. Because they do something that makes a noise. And they were widely available. But you wouldn't talked about cause and effect understanding because that wasn't the state of developmental theory at that time.
Piaget has been very clear that infants will [? precausal. ?] That they couldn't understand cause and effect. There was no point in giving a baby a toy cause and effect reasoning because babies have an undifferentiated sense of efficacy that can distinguish cause and effect. [? Causality ?] is a later version, [? hashtag ?] [INAUDIBLE]. It's not as though babies are much smarter.
Why were the young [INAUDIBLE] in the '60s and '70s? In part because Piaget asked babies, children, even much older children to explain how [INAUDIBLE] events, how vegetable grows, why a river flows. Many kids gave pretty nonsensical explanations or they attributed them to psychological principles. The river flows because it wants to.
He thought they couldn't distinguish between physical and psychological reality. By 1990, children were much smarter. Why? Because there's been a lot of good knowledge that children do have, as [INAUDIBLE] documentation of the physical world that makes distinct provisions for [INAUDIBLE]
So that's one view. But they also have particular pieces of [INAUDIBLE] which is an interesting example of So let me talk about it a little bit. One contribution to the idea that babies really do represent [INAUDIBLE] I told you about a couple of [INAUDIBLE] continuity and solidity. But another really important principle about object is contact causality.
So if I want to pick up that coffee cup, I could plead with it, talk to it, do all kinds of things, and that coffee cup would stay right where it is. But if I would like Josh to move, I could say, Josh, could you please get up and [INAUDIBLE] I can act on Josh, but I can't act on that coffee cup. [INAUDIBLE] We take that, again, as commonsense. [INAUDIBLE]
But it's just that babies have their expectations too. If you give them that object that is ambiguous. Maybe it is an object, maybe it is an agent. If you give it other cues, and it kind of cues agency and we're going to talk a lot about agency, but they know that this thing moves spontaneously by itself, or if you give it morphological features, some eyes, and things, then they expect that object to both resist being moved and move itself and move other things.
But if you don't give them those cues, the babies expect that objects obligatorily move when hit, when struck [INAUDIBLE] and they are surprised if it moves by itself when there was no contact [INAUDIBLE] So, these things together, by the way, [INAUDIBLE] but what an object really is to babies, [INAUDIBLE] the object is to say that it respects these three [? components. ?] And that is essentially a possibly innate, possibly biogenetically broad representation, abstract representation. It goes for all of them. And it falls [INAUDIBLE] predictions about [INAUDIBLE]
Let's talk a little bit about [? disperceptual ?] causality. [INAUDIBLE] all their children maybe need some help with all these toys to really master it, but [INAUDIBLE] And one of the reasons they thought it was a hard problem is that you can see one thing after another, after another, after another. You don't actually get to see causality [? as well ?]. [INAUDIBLE] the realist view that you just sort of watch the billiard balls and see one cause another. [INAUDIBLE] you just see one of the billiard balls moving, and the other. You don't see them necessary [INAUDIBLE]
Interestingly, work by a [INAUDIBLE] friend of mine, Michottean, [INAUDIBLE] said that [? Kim ?] may be right in principle, that you should be able to see it. But in fact, the perceptual properties of certain displays seem to give rise to a representation of causality that is extremely [INAUDIBLE] parameters. And if you manipulate the [INAUDIBLE] slightly, the perception of causality goes away. And if you have them treated stably, the perception of causality is robust and that [INAUDIBLE]
And interestingly they reported even actually when they know causality isn't present. For instance, when Michottean was doing it, he didn't have animation software, he were just able to use you know, the flip books that you make out of paper and stuff. And he has what looks like black balls hitting one another.
And it's the ball. And one ball came in and directly hit and moved the other ball. Adults reported it as causal. So this is a crude illustration. I apologize but something like that, right? And it looks like the blue ball is moving the green ball. So I've got to watch that. To change that ever so slightly, by putting for instance a spatial gap between the two balls, adults no longer report causality. They report it as an independent event. [INAUDIBLE]
It looks like, actually, under some conditions, we are motivated to see causality in the world. What about babies, right? Do they just see these as one event next to the other? Do we learn that these balls cause [INAUDIBLE] That's a hard question to ask. How do you ask the question if they see a direct [INAUDIBLE] a spacial gap event or a temporal gap event as non-causal? Leslie came up with what was a rather ingenious solution.
He said, well, let's show babies [? each of these events ?] forward, right? Blue ball hits the green ball, either directly or with a spatial gap or a temporal gap. And then [INAUDIBLE]
And what he said is, in all of these cases, a lot of those things change, right? You're reversing the trajectory. You're reversing the time of these events. But only in the causal events are you reversing the agent [INAUDIBLE] information. Only in the causal events does the thing that was bonked because of the bonker. Right? It's basic that you understand the event as causal. It's not just a change in space. It's not just a change in timing or trajectory. It's a deep change in the structure of that event.
First the blue ball may seem [INAUDIBLE] now the green ball [INAUDIBLE] equally interpreted that the reversal of each of these events should be equally interesting. But they care about the causal [? scripture ?], a much deeper violation happening here versus Hawthorne [INAUDIBLE] And that is exactly what he had very many of [INAUDIBLE] good evidence that newborn kicks did the same thing. OK. [INAUDIBLE]
AUDIENCE: It might be an inane question, but why [INAUDIBLE] at an earlier age, [INAUDIBLE] very early, but if they're hungry and they were given food, they will not be hungry afterwards. [INAUDIBLE] It takes time for them to make this connection, but [INAUDIBLE]
LAURA SCHULZ: One of the things we know as far as genetically broad, the ability to form associations between events, and actually there are two sets of abilities that we know happen and in any organism we might be interested in testing. All animals can draw predictions from repeated pairings of events. So if A happens, then B happens. If A happens, they will learn to expect B.
Depending on the intelligence of the organism and similarly you may have to vary the number of trials, but if A repeatedly predicts B, most animals can learn that A predicts B. So that's [? causal ?] conditioning, right? And all animals [INAUDIBLE] learning. If the [INAUDIBLE] do something and something happens, if they push this button, they get apple juice and they get fed, they push this button, and they get shocked, they quickly learn the association between their own action and the world and what happens to them.
What we strikingly don't have a lot of evidence in the animal literature is the ability to bridge this gap. So animals can repeatedly for instance, hear a tone and receive food, right? And they will even make different predictions about the events they see in the world and events they [? entertain on. ?] But what they did do, is having heard a tone and gotten food, and then say, learning how to activate the tone, they don't actually design an intervention. They say, if I do this, I'm going to be able to get the food.
They don't go from observing this food to [INAUDIBLE] We don't have good evidence for their ability to bridge that kind of gap. What you're seeing here in the babies is the ability to represent what might just be an association in all these cases, be it [INAUDIBLE] and represent something richer about it, represent a causal connection. Even though they're not directly involved in that, even though there's no instrumental [INAUDIBLE]
Let's talk a little out of the physical world, and move [INAUDIBLE] and ask another question. You've just traveled to another planet. Scientists tell us that there are all kind of planets out there that might be supporting intelligent life so we shouldn't limit our discussion of intelligence parochially to us Earthlings. How are you going to recognize agents on that planet? What are you going to do? Anything here?
AUDIENCE: Looking at things which appear to be acting on other things.
LAURA SCHULZ: OK. So you might look at things that are acting on other things. What else are you going to do? First, the wind acts on the leaves. You could be mistaken. [INAUDIBLE] But maybe it's an agent. What else?
AUDIENCE: You'll interact with it.
LAURA SCHULZ: What? Oh, interact with you. So [INAUDIBLE] right away. Or does it attack you? That would be a good cue.
AUDIENCE: Change information [INAUDIBLE]
LAURA SCHULZ: OK. So it's adapting to its environment. Is it moving in ways that we want? We're going to move on to talk about what we mean by intelligence is that some of the questions we want to look at is what kind of behavior will recognizably cue us to an unfamiliar organism that is, is it behaving in any kind of [INAUDIBLE] way? And although this is sort of a silly hypothetical question when you think about going to another planet, [? it's argued ?] that the kind of question that you are all presented with when you came into this course.
Here you are. First time, how are you going to recognize these? Who out there is an agent? How do you distinguish them from the objects? [INAUDIBLE] So we've considered the [INAUDIBLE] And there's a flip side of this question. Suppose you showed the babies were sensitive to objects that acted on other objects. They were sensitive to objects that responded differently to themselves.
They were sensitive to how [INAUDIBLE] adapted in a very [? spacious ?] environment. Would you think they had a [? competent ?] agent? Would you think that they had an intuitive technology? Would that be sufficient to credit an intent with that kind of representation [INAUDIBLE]
To give you a sense of why this problem again- and part of my goal here is to just break apart commonsense a little bit, it seems so trivial. Of course, you can [INAUDIBLE] Look around. [INAUDIBLE] But you could see the world otherwise. Right?
Here's a little thought experiment. What you see here are bags of skin stuffed into pieces of cloth and draped over chairs. There are small, restless black plastic spots that move at the top of the bags of skin, a hole underneath that irregularly makes noises. The bags move in unpredictable ways, et cetera, et cetera. This is a description of Gopnik can't imagine what the world might look like if you were genuinely [? mind blind. ?]
And people have sometimes proposed [INAUDIBLE] This is what I see, right? [INAUDIBLE] But it's nearly impossible. It's not a given [INAUDIBLE] So how do you get it? Well, one thing [INAUDIBLE] little bit of work on all types of cases, [INAUDIBLE] long after a lot of basic categories seem to drop out [INAUDIBLE] There's still a distinction between agents and objects.
How do you do it? Well, one of the things that makes it challenging is, of course, [INAUDIBLE] objects. [INAUDIBLE] pop in and out of existence. You can't go through walls and you have a lot of different superficial properties of objects. And so you have to rely on these other kinds of cues for [INAUDIBLE] morphological cues, [? key ?] to a lot of things [INAUDIBLE] I mentioned [INAUDIBLE] faces. I think you'll probably get a lot about faces in the course of becoming [INAUDIBLE] in various ways. And we'll talk more about this dynamic [INAUDIBLE]
Premack and Premack wrote this [? narrative ?] [? intelligence ?] which suggested that one of the [? clear tooth ?] agencies [INAUDIBLE] even snakes are sensitive to this kind of thing. [INAUDIBLE] in the environment and that may just be a perceptual trigger. And [INAUDIBLE] for all kinds of animals to recommend. [INAUDIBLE] And we know that perception is movement is important [INAUDIBLE] contact
And I think some lovely work demonstrated this [? construct ?] [INAUDIBLE] on Friday. She sent a bean bag flying over our wall. And it landed right there. And all you can see here actually is an object. But the object is behaving in a not very object way. It appears to be moving spontaneously. But if you resist that [INAUDIBLE] to our object, then you have to believe that something moved that object. And that something is probably an agent.
And so she shows babies [INAUDIBLE] a hand. And the hand either moves to the origin of the movement as it appears on this side. Or the hand [INAUDIBLE] And the argument is, you might expect an agent on this side, because don't they have to be responsible for the spontaneous movement of the object, the it should be an agent. But you have no reason to support the hand over here. This should be [INAUDIBLE]. And indeed, babies looked longer at this hand than at this hand.
And they only looked longer if it was a hand. If it was a train, another object that moved that there, the babies showed no [INAUDIBLE] Moreover, you might say, oh, look, hands are very familiar but babies actually see hands throwing objects. And also, hands actually can throw objects. [INAUDIBLE] But if you really care about the basic ontological distinction, the fact there's a kind of thing that can initiate movement in the world, then you might have a much more abstract representation than [INAUDIBLE]
What if I don't give you something familiar? What if I don't even give you something with hands [INAUDIBLE] I could give you something with other cues agency, like those morphological cues, like for instance, [? a snake. ?] Well, it so happens that this little guy here that they've never seen before and who doesn't have hands, but who can move spontaneously by itself, and who does have a face, babies treat it like a hand and not as a train. They expect it over here and they look longer when it appears over here than when it's here.
AUDIENCE: How long?
LAURA SCHULZ: The original results were in 10 to 12-month-old babies. And I'm never very good [? at agents. ?] I think she would have gotten a little younger, if anyone knows [INAUDIBLE]
AUDIENCE: They might have gone down to seven month old.
LAURA SCHULZ: I think you are right. [INAUDIBLE] If that [INAUDIBLE] that jumped over the wall, the babies no longer expected that. So what you see is this creature with eyes. And this creature with eyes, instead of the bean bag over here, well, that creature with eyes that can move spontaneously can do whatever it wants. And they don't expect that a hand initiated it.
AUDIENCE: So the train's been moved out?
LAURA SCHULZ: The train moved out in exactly the same trajectory as the hand. So the argument that it controlled here, the low level control for one thing emerged on this side, the bean bag. The babies [? didn't expect ?] other things to emerge from that side. And to show that wasn't true, that it was specific to the hand, if a train emerges from this side, it's just as surprising as [INAUDIBLE]
AUDIENCE: What are the differences between [INAUDIBLE] expecting, not expecting?
LAURA SCHULZ: Statistically significant but I don't have those actual drafts up here. They look longer. You usually get two measures from instant looking time, so how long on average they look at one as opposed to the other, and how many individual babies show the looking time concept.
Objects don't move by themselves, agents do. But is that sufficient for a baby to-- again, if they had that cue, would you be satisfied that they attributed it to [INAUDIBLE] that they really understand everything about agents. Is there some reaons to think not? I mean I have to show you parts of this study in pieces. It's a really nice study from [? Susie Johson's ?] lab, where again, we're going to use something, bean bag like here. This is a fuzzy green blob, basically. And the blob has a little light up there that's attached and it beeps in some conditions. It has the capacity to move [INAUDIBLE]
And what [INAUDIBLE] So what would babies use-- do they have a whole [INAUDIBLE] If you show them one thing, do they expect other things like that? And so what she was interested in [INAUDIBLE] We know by 10 months of age, babies do something quite remarkable, which is, they go like this. I don't know if I managed to get a list of [INAUDIBLE] but I'm that some of you are following my gaze outside the window. Because I turned and some of you looked in that direction.
And if I pointed, I think you guys would follow my attention. You didn't do that [INAUDIBLE] Which is to say, you didn't look at me, you didn't look at my finger, you looked where I was looking, used my attention to gauge your attention. That's development called joint attention that emerges toward the end of that first year.
Now, what she wanted to know was when would babies treat an entity as an appropriate subject for their attention. When would they follow a [? gauge? ?] So she had a fuzzy green blob, who moved spontaneously. And the fuzzy green blob turned and oriented a particular direction. And she wanted to know, would the babies turn and follow that entity's [? gauge. ?] And spontaneous movement was [? nominal. ?]
Anyway, to follow the [? gauge yet. ?] But as I will show you soon, you don't need a face to get babies to follow the gaze. But you do need [INAUDIBLE] So we're going to come back to that. So if it did move with the face, babies would follow [INAUDIBLE] babies didn't treat it as an agent [INAUDIBLE]
Well, maybe you need a little bit more than spontaneous movement. As noted, she went to another plan and looked at something that spontaneously moved, you could be very distracted by whatever the other kind of physical force, perhaps unfamiliar ones were causing things to move [INAUDIBLE] not because of their psychological ones. So somebody mentioned, well, you really want to see if they [? adapt ?] to their environment. Most of our objects do. [INAUDIBLE]
And so, infants might use all of these things, once they understand that objects do exhibit properties of inertia, then it might be a violation to say that that has self-generated movement. That is something like an agent. And the nicest work on this comes from [INAUDIBLE] lab where you can see that subtle cues to changes in the motion trajectory and in particular, heading, makes a big difference to the [INAUDIBLE]
So I can move around this room in a lot of ways and you can takei t to be an agent. But one think I don't usually do is move like this. I move in the direction of my head. I don't just move about in all kinds of other directions. And it turns out that that heading and the ability to follow and change [INAUDIBLE]
LAURA SCHULZ: Excellent.
So what do you guys see here? Anyone.
AUDIENCE: A bunch of airplanes [? looking ?] at a ball.
LAURA SCHULZ: Airplanes [? looking ?] at a ball.
AUDIENCE: The green is running away.
LAURA SCHULZ: Anyone want to be that green ball?
[LAUGHING] The green ball looks like it is being pursued. [INAUDIBLE]
But I want to show you how quickly you can make that go away. So everything is the same here except for the [INAUDIBLE] And absent that, [INAUDIBLE] So a lot of our perception of agency, like our perception of [INAUDIBLE] subtle spacial parameters. [INAUDIBLE] But babies are also sensitive to cues like changes in stop and start and trajectory and heading.
I think that a more interesting finite [? key ?] agency is this [INAUDIBLE] So someone mentioned that one way you could go to a planet and find out what the agents were is the way [INAUDIBLE] respond to cues. If they ran away from you, they'd have to, if they like tried to make contact, and trade things with you, you might be [INAUDIBLE] to think they were agents.
[? So Susie Duncan ?] tried this as well. She had that fuzzy green blob there and then, babies tend to do things spontaneously, They go duh. And the blob goes beep, beep, beep. and the baby would go duh. And the blob goes beep, beep, beep. And then the baby would move its head. And the blob would go beep, beep, beep.
It was the responses that were huge to the babies' reaction in one condition. In the other condition, the baby has got the behavior, the green blob behavior as the previous baby. [INAUDIBLE] So in one case, the [INAUDIBLE]
So they seem to treat that as again, a cue to agency, again absent [INAUDIBLE]
AUDIENCE: Are there any experiments that show [INAUDIBLE]
LAURA SCHULZ: Well, she did an interesting followup. You might wonder how this thing-- we kind of think this is the head. But it's possible this is the head. How do you know? It doesn't really have a face. It could have a really big head and a very small body. Or it could have a small head and a great big body. And it could have a little pebble on the thing that is diagonally, right? [INAUDIBLE] And the head is that way. [INAUDIBLE] And she talks to the green blob in one position next to the baby. Or it talks to the baby.
So one of the other was more plausibly the head of the object. And after the third party interaction, and the baby selectively oriented to the direction to which she spoke to the head of the green blob. So they determined the heading by the direction of [INAUDIBLE] So these are fairly [INAUDIBLE]
By quite early on, babies discriminate many characteristic features of agents and objects, based on these dynamic cues and also based on [INAUDIBLE] function. But again, we might want to know when babies bind all of these features together. And in particular, not just when they predict one cue off of another cue off of another cue, but when we can really decide that these features cue the things we can't see and we can't discern. The fact that you are hanging out there, not just with changes in trajectory and changes in responding and behavior, but with an inner [INAUDIBLE] with goals, intentions [INAUDIBLE]
So goals are precursor [INAUDIBLE] I can give you a lot more about the theory of mind reasoning from Rebecca [INAUDIBLE] But today, I'm just going to walk you through what you would need even to get there, right? So there are set behaviors. [INAUDIBLE] A rock rolling downhill is not about me rolling downhill [INAUDIBLE] I'm an agent doing it. But I'm acting [INAUDIBLE]
Intentional action is about [INAUDIBLE] It's either goal-directed or referential. And [INAUDIBLE] But if I wave my hands around vigorously, and flap them, my actions are either functional, I'm trying to make the lights turn on. Or I'm trying to do something communicative, referential. These are the stupid lights that tend to go out when you're always having a quiet moment in a lecture.
So when do babies appreciate these aspects of [INAUDIBLE] intentional action. I could use one of these. [INAUDIBLE] [INAUDIBLE] I could infer the action into this entity [INAUDIBLE] I could decide it was referring to something without really [INAUDIBLE] But if you're going to get [INAUDIBLE] So this is a prerequisite to I'm going to start with a classic video that a student-- have you had [INAUDIBLE]? Well, if you haven't, I'd like to introduce you to [INAUDIBLE] circa 1944.
All right. Can you describe to me what it is that you just saw there?
AUDIENCE: Kids interacting with adults.
LAURA SCHULZ: Kids [INAUDIBLE] What else?
AUDIENCE: Big triangle defending his castle.
LAURA SCHULZ: Big triangle defending his castle. Let's get a couple more. Very classic.
AUDIENCE: Two friends were playing and they were assaulted by something.
LAURA SCHULZ: Two friends did what?
AUDIENCE: Were playing and they were assaulted by someone.
LAURA SCHULZ: Two friends were playing and were assaulted. Ah. And many of you are privately imagining much richer narratives here. But you're embarrassed to say because part of you is thinking, no, no, no. [INAUDIBLE] And that's remarkable and strange and you should be troubled by it. And it is a problem for intelligence and we think commonsense and reasoning is a hard problem. I think more than anything-- this is a very old demonstration, right? From [INAUDIBLE] data, from motion, trajectories, and [INAUDIBLE] a lot of rich, structured [INAUDIBLE] you're telling a story that goes far beyond that, far beyond the data. And yet, all of you thought, [INAUDIBLE]
And if I showed it to children, they would tell you a story, like there's bullying and [INAUDIBLE]. They would feel fear. They would giggle. They would be delighted and happy and think. And one question you might want to ask, is what about [INAUDIBLE] You can might say, OK, well, look, you can do that to me because I grew [INAUDIBLE]
What about babies? Some lovely work by [INAUDIBLE] in the '90s. They were asking questions about what babies could see in all directions. They habituated babies to events like this. This looks like a very neat sketch, but it was a very dramatic moment, especially if you're a baby. But you have over here, this small ball. What you have over here is a big ball. But what you see is the small ball backing up and back and forth [INAUDIBLE] going beep, beep and flashing. And the big ball over here saying bleep, bleep, bleep. And the small goes beep, beep. And the big ball bleep, bleep, bleep. And irresistibly you have a baby ball that's separated from a mama ball. [INAUDIBLE]
What do the babies really think? Well, what you see is you see that the baby ball back up jump over the obstacle. [? Continue ?] with each of the terms here but I think it's justified here as you will see. They say that over and over again. But of course, really you see is the balls and beeps and the trajectories again and again and again. [INAUDIBLE]
And the you show them either the old action, which they have just been [INAUDIBLE] Because the agents have a goal and it's a rational actor, which is to say, it's acting effectively. It's trying to take the shortest path before [INAUDIBLE] Then the ball will go like this. This is a new action. The babies have never seen the ball go like that.
But that is a rational action if you [INAUDIBLE] goal-directed action. If you just [INAUDIBLE] this is a familiar action. This is what [INAUDIBLE] This ought to be boring. That [INAUDIBLE]
And what you're seeing here is that infants do [INAUDIBLE] great repetition of six, nine, and 12 age. I think the six-month-olds were the longest. The nine-month-olds were the [INAUDIBLE] Not only do babies infer the action the relevant action [INAUDIBLE] But you have quite selectively, because you just put the ball behind here where it's no longer a ball in the first place [INAUDIBLE] an inefficient action, the babies [INAUDIBLE]
And [INAUDIBLE] the babies the action itself, [INAUDIBLE] so they can't see what's behind here by [INAUDIBLE] the babies will posit an obstacle they've never seen. So they are more surprised that no obstacle is present than if an obstacle is covered. Although they never saw the obstacle. They have just invented it to make sense of the action. So again, [INAUDIBLE]
So we know babies are tentative to other agents actions. [INAUDIBLE] And as you'll see in a minute, [INAUDIBLE] But let's talk about some of the earliest [INAUDIBLE] research suggesting that babies see actions, agent's actions as goal-directed. Amanda Woodward showed babies a reach. and there were two objects and the hand reached for one of the two, let's say the ball, repeatedly. Ball, ball, ball, ball, ball, ball, ball, ball.
And then the hand [INAUDIBLE] just switched the position of the object. And then the hand is doing exactly the same thing. It's going to reach, continues to reach in exactly the same place. But now there's a new goal. The hand [INAUDIBLE] an entirely new reach but it's for the same goal. Now the baby doesn't care primarily that that low level [INAUDIBLE] primarily sensitive to just the change in trajectory, then [INAUDIBLE] But if they make observations about goals and what they care about with agent action is goals, then you obviously see the reverse. This action, although superficially familiar is much more interesting. This action, although superficially [? nominal ?] [INAUDIBLE]
And in fact, six-month-olds should [INAUDIBLE] when it was a hand but not when it was attached [INAUDIBLE] But more interestingly, three-month-olds didn't care a lot. But three-month-olds did not seem to treat the goal as an important agent action. And one hypothesis is that is because three-month-olds themselves [INAUDIBLE] at all. They have a three-month-old [? spacing ?] around. They flail. Occasionally they happen to whack things. But they're incapable of executing it.
So Jessica Sommerville took two-month-old babies and she put Velcro mittens on them. And now the babies were given a few minutes for the first time in their little lives and the first time in the history of the world [INAUDIBLE] Three-month-olds were probably a tad bit more able to be trained to pick up an object and now this is a brand new event for babies and a small step for humanity. And they did that for a few minutes of play time.
And then they saw these same videos and the babies see themselves had executed a reach, now made the distinction like the six-month-old baby [INAUDIBLE] goals and in the change of the trajectory. The three-month-olds who did not have that change, failed [INAUDIBLE] The babies' ability to recognize the agent goals are related to their own inability.
AUDIENCE: And that training was only for a few minutes?
LAURA SCHULZ: Only for a few minutes. Because you try keeping sticky mittens on three-month-old babies for any longer than that.
AUDIENCE: They still showed no interest in the [? cloth? ?]
LAURA SCHULZ: And [INAUDIBLE] Actually that is an excellent question. Do I know that they [INAUDIBLE]. I do not.
Let's talk now about some recent research that I think has once again expanded our understanding of education [INAUDIBLE] This is by [INAUDIBLE] and I'm going to show you the first groundbreaking new paper here, but she now has literally some 20 studies [INAUDIBLE] experiments. It is one thing to understand the goal to reach for an object to go over a wall. But one of the really interesting things about human action is that my goals can be your goals, which is to say, I can help you.
If I know what your goals are, my goal can be the same as your goal. Or I can be [INAUDIBLE] my goals can support your goals. So if I can represent your goals and if babies understand one agent's ability to take on another agent's goals, for better or for worse, then they should know [? the represent ?] action matches this goal directive, but is potentially as helpful [? or as useful. ?] Because I want to show you a few clips suggesting that infants [? can do this ?] [INAUDIBLE]
What you see is a very small red ball trying to get up a hill. And now there's a little yellow triangle. What you all see is that the yellow triangle [INAUDIBLE] So that's a little [? helping ?] video. Here is an example of a not so helpful agent. Try that again.
Well, you can imagine the scenario if I can't get it to play. I'll just play this one again and [INAUDIBLE] Red guy tends to go up and this time, instead of coming behind and helping the red guy go up, what you see is a blue square on the other side and [INAUDIBLE] Having seen both the helping and the hindering events, you can then show the babies.
And if I were this red guy, I'd have to make a choice. And in one condition, he does what you expect. He approaches the helpful agent. And in the other condition, he does what you might not expect. He goes ahead and he approaches the agent. Everyone clear on that? Now, I'm just giving you a nice narrative explaining it. Of course, the babies won't get an explanation. They just watch the video.
The question is do they form any expectation based on the behavior that they've seen here about what that agent [INAUDIBLE] And of course, [INAUDIBLE] expect the agent [INAUDIBLE] It matters that that thing is an agent. It doesn't have eyes [INAUDIBLE]
And you might say, you could get that effect if the agents are only sensitive to helping and thinking that agents are helpers or if they're only sensitive to hindering. And you'd say, don't you think you should disprefer agents who block you.
She also did an experiment where she compared helping and hindering respectively to a neutral agent. [INAUDIBLE] neutral agent, just a bystander, just watched what was happening. And the prediction in this case is, of course, that the babies should expect you to approach the helper over the neutral agent, and the neutral agent over the hinderer. [INAUDIBLE] That all clear? So this is just actually a summary of what agents should do there. [INAUDIBLE] She shows this effective in both in six-month-olds and ten-month-olds. And later results in infants are [INAUDIBLE] as young as three-month-olds.
And as Josh will show you shortly, these are not just the independent preferences, but these preferences are predicated on actually knowing what the other agents' goals are. You have to understand what that agent wants to do. I think it is remarkable that no early representations were [INAUDIBLE] but clearly, rich sophisticated kind of structured representation about the nature of goals. But again, [INAUDIBLE] The ones I've shown you of a full theory of mind. [INAUDIBLE] a lot of the [INAUDIBLE]
AUDIENCE: [INAUDIBLE] question. How do you tell if a baby says, oh, yes, I like this or not? Is there a certain sound he makes or--
LAURA SCHULZ: [INAUDIBLE] She's done it in a variety of ways. She's done it with looking times. But she's also done it with children-- and so you bring out the two objects and the babies, the six and ten-months [INAUDIBLE] So you just let the experimenter who comes into [INAUDIBLE] let the babies just see both objects are there. The baby gets to make a choice. And the question is, do they selectively pick up the helper over the hinderer and the neutral agents and they selectively avoid the hinderer over the helper and the neutral agent. And you can see babies doing all this. Yes.
AUDIENCE: [INAUDIBLE] Agents don't have any sort of eyes?
LAURA SCHULZ: They do. In fact, it's important that they have eyes and it's important that their eyes exhibit a heading that's actually [? grooving ?] with the trajectory motion. She does them without eyes and you don't get these effects. So it's important that they're being coded as agents for the kinds of things for whom goals might be relevant. Yes.
AUDIENCE: You said goals are [INAUDIBLE] in the theory of mind.
LAURA SCHULZ: These figures have to understand [INAUDIBLE]. Certainly if they're going to understand goals that [INAUDIBLE]
LAURA SCHULZ: I think that, again, and I think you'll see some of this even later today, that what you really want to know the origins of the theory of mind has come down to [INAUDIBLE] as representational. It doesn't just copy the world as it really is. It is a representation of the world [INAUDIBLE] world. That you might lack knowledge, you might [INAUDIBLE] knowledge, [INAUDIBLE] You really want to say that you understand [INAUDIBLE] in terms of what kinds of things [INAUDIBLE] and understand that agent action is attention [INAUDIBLE] understand about the nature of knowledge and about the nature of [INAUDIBLE] Yes.
AUDIENCE: Do you know if adults seeing similar kinds of things show [INAUDIBLE] preferences that are similar to kind of the way babies look? I mean, was originally like a [INAUDIBLE] that kind of showed that responses between both are similar and that adults--
LAURA SCHULZ: It's a complicated question. In general, it's a complicated question in development about how you show again, a common dependent measure that [INAUDIBLE] So if I want to know what you see, I can just ask you. I don't have to bother [INAUDIBLE] differential [INAUDIBLE] that babies might [INAUDIBLE] 12-month-olds [INAUDIBLE] to events. But already they're looking at just like a third of what it was at six months. By the time you get to be adults you get all of this in a glance. So you end up with different measures when you're trying to ask adults about what happened. But it is important in the infant literature that people work hard to figure out what can you do besides looking [INAUDIBLE] Yes.
AUDIENCE: This is a slightly weird question about agency. Do you have any idea about when children come to know that or come to realize that, adults, up until then have really been screwing with them and assigning agency to objects and really never had it, like [? Mr. Foam? ?] And like stuffed animals?
LAURA SCHULZ: [INAUDIBLE] computer.
AUDIENCE: They're never expected to respond though.
LAURA SCHULZ: It's an interesting question. People have long asked similar questions. We think we're screwing with them because we represent things fancifully or [INAUDIBLE] and it turns out kids are exquisitely good at distinguishing pretend scenarios from real scenarios.
AUDIENCE: Even when they're doing these tasks?
LAURA SCHULZ: Very early.
AUDIENCE: So like they don't really think that the agents are agents. They just think that the adults are calling them agents because they're in this game?
LAURA SCHULZ: Again, you're in trouble when you start trying for a experiment because you don't have the data bearing on it, say what's going on in the minds of the babies. A lot of people have thought about this problem of [INAUDIBLE] representations. It's tough. I'm sure adults know less than children [INAUDIBLE]
You go to the movies, right? You have rich emotional reactions to those. You can get scared. You can get moved by things you see the screen. But you don't call the cops. You don't actually confuse it. And yet you have a lot reaction that are quite specific to the representations that you see, including real emotional aftereffects and real discussions about what happened to these totally fictional characters.
As far as we can tell, babies are perfectly good at working with symbolic representations, at working with cartoons, depictions of working with [INAUDIBLE], making attributions about them, but without confusing them. Same thing goes for when they pick up a block and go vroom, vroom, vroom, vroom, vroom. So they can represent that block as standing in for a car. You get these stand ins or representations very early and they don't seem to be confused.
But I think that those are some of the hardest. Like I'm showing you stuff we do understand. If you want to start asking about imagination of representation and the status of these things, we're in a world of hurt because it's a really, really complicated problem. It is what you spend a lot of your early childhood doing. [INAUDIBLE] as far as we know. We don't understand [INAUDIBLE]
So, with that let me turn a little bit to learning, which I'm just going to [? presage ?] here. And to do this, I'm going to talk a little bit about what you might think [INAUDIBLE] recently because there have been a couple of major traditions that have been [INAUDIBLE] One is the Piagetian tradition. I alluded to Piaget. I haven't talking very much about him. But if you knew a tiny bit about [INAUDIBLE] you would know Piagetian theory. And Piagetia, in theory, was fundamentally stage theory.
Babies start as [? sensory ?] motor learners, gradually their abilities [INAUDIBLE] the world got more and more and more abstract. [INAUDIBLE] So, and a lot of early abilities were not [INAUDIBLE] because they would be discrete stages. All of the research I've shown you has basically marked the demise of Piagetian theory. We now think there is no point [INAUDIBLE] We think there's no point when abstract representation [INAUDIBLE] There is no evidence for huge stage shifts in cognitive representation for development.
That whole theory is [? lost. ?] That doesn't say that Piaget is [? lost. ?] He carved out his [? remains ?] like I said, from morality to probabilities that we should be [? concerned ?] representations over. He posited a lot of ideas about how you would [INAUDIBLE] that talk about between experience and structured representation. But stage theory in particular, [INAUDIBLE].
Another of course, dominant theory of learning until the cognitive revolution, is [INAUDIBLE] continues to be in neuroscience today is just associative learning theory. How you learn. [INAUDIBLE] and I guarantee I can make them anything you want just by manipulating their environment. What governs learning? Environmental [INAUDIBLE] What you can associate in the environment. I can make them a doctor, a lawyer, even a beggerman.
And a lot of things contributed to the [? cause ?] of behaviorism. But it was a very [INAUDIBLE] theory and remains so [INAUDIBLE] because it's incredibly measurable. You really manipulate different kinds of kids and their environment and direct behavior of rats today and mice today and neurons to a great extent, just by controlling the [? carriage ?] and the environment.
But Piaget and many others said this is a problem. To explain these successive generalizations by this simple associations explains nothing at all because the problem is precisely to know why these associations are formed and not others among the infinity of combinations possible. So along with things like Tolman's experiments showing that rats learn things which they were never reinforced. They run the entire maze, [INAUDIBLE] Along with Chomsky's argument about the poverty of stimulus and that there was not empirical evidence nor any complicational way where you could get language from simple associations between words and [? conventions. ?]
And along with Turing, who showed actually you can make rigorous, measurable, quantifiable things that have nothing to do with associative learning. And yet, [INAUDIBLE] That would be a program. All of these things together, along with Piagetian theory, suggested that behaviorism wasn't the way to go. you were going to need to consider it [INAUDIBLE] You were going to need to consider [INAUDIBLE] But just in case you aren't quite persuaded of this. And you aren't persuaded that commonsense really is a part of [INAUDIBLE], there are an infinite number of possible foundations.
Let me give you a little multiple choice test. A speaker of Quinian is going to point to this and say, 'Gavagi.' So [INAUDIBLE] What a Gavagi? What does it mean?
LAURA SCHULZ: Don't overthink.
AUDIENCE: A rabbit.
LAURA SCHULZ: A rabbit. OK. Excellent. [INAUDIBLE] a rabbit. And all of you are very likely to agree that it means a rabbit. But as the philosopher Quine pointed out, it could mean a rabbit plus grass. It could mean lunch. It could mean bugs.
It could mean rabbits in the year 2014, but not thereafter. There are infinite things that it could mean animal. It could mean fuzzy thing. It could mean cute. It could mean bugs. How do you know-- have you converged so rapidly, given all of these associations that [INAUDIBLE]
Here's another example. Complete this sentence: The giraffe has a very long--
LAURA SCHULZ: Neck. Or temper Of flight to Kenya. Totally trivial. And yet, there are infinite combinations. You just read through one and the problem of commonsense [INAUDIBLE]
I'll give you one. Well, what's behind the rectangle?
AUDIENCE: [INAUDIBLE] Well, I point out, it commonly occurs in everything from [INAUDIBLE] structures to [INAUDIBLE] What do you think you're going to see if I remove that yellow rectangle?
AUDIENCE: The rest of the bar.
LAURA SCHULZ: A single [? red bar. ?] Exactly. Just like that. But it could be this. It could be this. It could be [INAUDIBLE]
So you are [INAUDIBLE] categories, the shape, the texture, the color, the location, the type, any of those things, their properties they share. How so you either have or learn that that is the right step [? to find ?] by which to constrain these infinite possibilities? What is the structure that lets you learn rapidly with small amounts of data? That is the hard question that we're trying to get at.
And this is a problem, like I say from [? the Modernom ?] system to language to [INAUDIBLE] data. If that is true, you need to change [INAUDIBLE]. I'll give you a last [INAUDIBLE] Suppose I tell you there's research that scientists came up with a technique for changing the gender of mice. They injected 100 mice with glucosin before impregnation. 70% of the newborns were male.
OK. If I asked you, [INAUDIBLE] as evidence for whether they actually succeeded in manipulating the gender of mice? You don't know. Can I see a show of hands if you might read this study. You all are a cagey bunch.
Now I'm going to tell you very similar data. ESP experts from the Institute of Noetic Sciences in Eugene, Oregon. My home town is Oregon. At least one I lived in a long time. --came up with a technique for changing coin flips. They flipped a coin 100 times while chanting heads. 70% of the coin flips came up heads. How many are going to read this study? OK. It's the same statistical data. Why are you incredibly biased and shortsighted? Elitist prejudice on your part? Right?
[INAUDIBLE] Another problem [INAUDIBLE] There must be constraints on learning and a lot of what is going on in psychology is [INAUDIBLE] the nature of those constraints. How many of them are hardwired in [INAUDIBLE] What is the structure of those [INAUDIBLE] How many of them are learned? What are they? [INAUDIBLE] OK?
So, this posed a huge problem, a huge debate, a huge problem. This problem is probably directed right at the school, actually, which is that it is a chicken and egg problem. If you think about Piaget, but really if you just kind of think about your intuitive theory of learning, like well, how am I going to have a kind of abstract structured representation that's going to contain the correct interpretation of the data, the only way I can get it I though, an abstract representation is by getting a lot of specific examples. So I get this example, this example, this example, this example. Then I generalize and I have an abstract theory. And I can use that then to make substitutes and predictions and constrain the information.
But suppose I [INAUDIBLE] but how, if I need to interpret this concrete example, [INAUDIBLE] representation in the first place, how do I ever solve this problem? Why aren't I in a constant chicken and egg problem where I need abstract representation, constrained representation and otherwise, determined by the data. But I need the data to get that abstract representation in the first place.
Some constraints, [INAUDIBLE] plausibly innate. Maybe they [INAUDIBLE] Maybe they re early developing. But surely not all of the things I use to interpret the world are we born with that we are just handed. So how do we solve that problem? And I think a really lovely answer to that is the kind of answer that I think
Sorry [INAUDIBLE] The kind of answer that I think Josh talked about earlier this week, which is the [INAUDIBLE] So I'm sure he mentioned it. I'll remention it. One, which is this idea that you could [INAUDIBLE] Imagine the problem of walking into a room where there are hundreds and hundreds of bags of marbles. Brown bags, so you don't know what's in them. hundreds of them.
And you go in and you open one bag and you pull out a blue marble. And then you pull out the second blue marble from that bag. And you move on to the next bag and pull out a red marble and a second red marble. And third bag, and you pull out a green marble, and a second green marble. Now I have six marbles. And there are hundreds of thousands of marbles in this room. But from six marbles, you can already make a pretty powerful guess that these bags are homogeneous. And that means that a single marble from each of the next bags will tell you what's in that bag.
So you already know something about the abstract structure and you know [INAUDIBLE] So you can use small amount of data to get abstract representations, which then can spin your interpretation [INAUDIBLE] evidence. That could be wrong. It's small amounts of data. Maybe you got some weird sample in your six marbles. But probably you didn't. So if it's a fallible inference, it can certainly be wrong. All of probablistic inferences can be wrong. But it's a good one and it it's right, it will support rapid, rich, powerful learning. And that's the kind of mechanism that we talk about when we talk about the ability to form [INAUDIBLE]
So one thing you might want to know, it's really direct, but if you do something like that marble experiment, in babies, can you show that [INAUDIBLE] work What babies see is they see a huge [INAUDIBLE] a bunch of [INAUDIBLE] from box one, a bunch of squares from box two, a bunch of [INAUDIBLE] from box three, and [INAUDIBLE] You don't know what is in this, right? But you expect that whatever the first shape is, all the other objects will be [INAUDIBLE] You can also put the color and the different [INAUDIBLE]
What she did was show an expected outcome, one that [INAUDIBLE] or an unexpected outcome, one that [INAUDIBLE] And if you've developed an-- you don't know. Right? You've only seen one thing, pulled it out of that box. But if you've already developed an expectation of a kind of thing it should be, then this should be a violation and that should be expected. But it should be a violation of what comes out of this box. It shouldn't be something about just taking two shapes [INAUDIBLE] two of different shapes into the same shape.
And if you pull that [INAUDIBLE] Because it's an [INAUDIBLE] about the population and the distribution of samples of [INAUDIBLE] population. And that's [INAUDIBLE]
AUDIENCE: Should probably
LAURA SCHULZ: All right. So that is, I think, actually the end of the talk. That's good. Thank you very much.
AUDIENCE: When you talked about babies learning something at like six months, or 10 months or 12 months, what is the standard deviation on that, on those numbers?
LAURA SCHULZ: You're asking me the kind of question that I [INAUDIBLE] developmental psychology. It varies from experiment to experiment. If you care about the [INAUDIBLE], you'd have go back to the specific papers and the specific question. Sometimes [? agents ?] are very meaningful. They really do show that [INAUDIBLE] But for very many experiments, what you find is that they've tested it [? in this age ?] but people just haven't looked to see if earlier [INAUDIBLE] absence of evidence for earlier doesn't necessarily apply [INAUDIBLE] absent earlier. They're only developed at that age. It's just that they ran the experiment on that age.
Some of the reasons why you might run an experiment on one age and not another are irrelevant to the conceptual content of the experiment. you may run it because your dependent measure was reaching. And your babies can't reach and you haven't figured how to design the experiment without reaching the dependent measure. Or you might need to give the babies a fair amount of exposure to a kind of complex scene. And you're going to lose the younger babies before your can even get to the interesting question.
It doesn't mean they wouldn't have the ability if you could figure out a clever way to show it to them but you haven't figured it out. So how much weight to put on particular agents to give a finding. There are some case studies where there's real precision and they'll care and [INAUDIBLE] but it really depends on which experiment you're talking about.
You might have asked me a different question, which is what age range do we mean? And the younger the age that you're testing, the narrower a range is permitted to be in the standard deviation of it. So if I'm testing pre-schoolers, I might mean 48 months to 60 months. But if I'm testing 12-month-olds, my [INAUDIBLE] is two weeks. And if I say I'm testing a newborn, I [INAUDIBLE] three days. So there's a [? titration ?] with the age of the babies In infancy studies, if you say six-month-olds [INAUDIBLE] two weeks from their birthday.
Other questions. Yes.
AUDIENCE: A study where [INAUDIBLE] when you are just drawing the balls blindly or are you actually looking at them selectively choosing them?
LAURA SCHULZ: Not in the [INAUDIBLE] processes study, but there's a series of other, quite lovely studies about babies expectations about the relationship between samples and populations, they have shown that, babies [INAUDIBLE] if they have a box of red and white ping pong balls, and there are more red balls that while balls, they expect the samples to resemble the population. If you're drawing blindly, if you're drawing selectively, all bets are off. If you're not drawing to the box, but from [? your lab, ?] all bets are off. So it's specific to the sampling from the population. Babies care about the physical properties. If they [INAUDIBLE] Other questions? Yes.
AUDIENCE: I think the Hamlin study that you mentioned about the helper, the hinderer, when we discussed it, we framed it in the moral understanding of it. And then I think I studied somewhere that the [INAUDIBLE] because it's probably just a simple association between like the positive correlation and negative correlation. And then I think there's some problem with the [INAUDIBLE] study at all. Basically also like their googly eyes seem to be [INAUDIBLE] recognize. It seems to be like a very debatable issue. So I just wondered do infants have a good understanding of morality?
LAURA SCHULZ: There's a lot in that question. [INAUDIBLE]
LAURA SCHULZ:Hamlin had come under a huge amount of heat for this study. But it was suggested very early. And you'll notice it says the title of that is not about moral reasoning, it's about social evaluation. And I think that's there's a slight shift in how [? Kylie Casaul ?] talks about that work. She talks about social evaluation a little more than moral judgment. Because they got so much flack from moral judgment. But I think, my assessment of the evidence is that overwhelming favors Hamlin's interpretation of this.
There's a huge body of work and I'm actually going to show you some more. [INAUDIBLE] real and robust. If you do control conditions with not an agent going up the hill, it's an object. If they're just associating the motion trajectories, you should show the same kind of phenomena. What I think you may have discussed in the class is an interesting point though.
There's some argument about the googly eyes. So if they're looking like their goals were just the same, just spinning in space. And the argument is that it needs to be represented as an agent and have that header. You saw that effect yourself in looking at [? Kaagaw''s ?] wolf pack studies, right? Where he [INAUDIBLE] really seeing this as agency in one condition. And very similar stimulator but with a different heading representation no longer looked like agency.
I think the point in Kylie's favor that they represented this agency when it had the header but not when the eye was free to spin and it doesn't look agent-like anymore. So I don't think like [? content. ?]
Now, there's been some interesting arguments in the [INAUDIBLE] about whether you really want to take this in the full, rich interpretation of moral helping of it it's something like goal congruence, where you're taking literally [INAUDIBLE] taking another agent's goal as your own or understanding helping and hindering in a moral sense.
For instance you might wonder if you take someone's goal as your own but their goal was also [INAUDIBLE], You've got that [? talent ?] and you're going to be blown up, I think you have a moral sense than the hinderers. The moral agent and the helper is not. You can ask a lot of richer questions, but I think the evidence that babies do really distinguish agents who take on another agent's goal is and-- first, it's hinder another agent's goal.
They do it specifically-- conduct that's specifically when they're [? caught ?] in the event. They do it specifically in context where the agent has reason to know what agent's goals are. It's at this point overwhelming. We've replicated it, I think some 20 times. So, anything can be wrong. Maybe it will fall apart, but I haven't seen any convincing evidence that convincingly undermines her interpretations.
AUDIENCE: So if the object of her experiments, so if the argument that [INAUDIBLE] is that a claim of--
LAURA SCHULZ: That was Piaget's very old plan, [INAUDIBLE] So the old plan of [? acting ?] [? permanence ?] is that babies-- I always have to be careful when I teach this. When I teach this to the undergrads, I should put a big red x over it because this is such an overlearned thing and so much [INAUDIBLE]. You take one this away from you.
Babies have object representation. Object representation [INAUDIBLE] They represent to an object. It's not true that they don't have object permanence. They do have object permanence. Whether they can reliably act on it in all conditions and they try to perform it [INAUDIBLE] acting on it and then searching for objects, clearly there are limitations on that. But there's no question, I think, about their ability to represent objects that are out of [? phase. ?] Again, newborn chicks can do it.
AUDIENCE: Right. So this is more of a representation [INAUDIBLE] Like if you were to do the experiment that you saw with the--
LAURA SCHULZ: Chicks.
AUDIENCE: [INAUDIBLE] the frequency of that [INAUDIBLE]
LAURA SCHULZ: If they did, you wouldn't know how to [INAUDIBLE] I think it's one way to say whether you can-- if it's an adult, might not know what to think in that context. It becomes a little bit [INAUDIBLE]. Suppose they had an object behind a screen. An hour elapses and then that screen black, well, you may or may not think that object got moved in the interim of was still there.
You may have forgotten about it because there were other things you were thinking about or not but it wouldn't tell me-- it would tell me much more about that other kind of information processing demands on you than it would about your potential object representation. So, yes, you could match with the phenomenon. but how richly you would want to interpret something within that kind of a setting would be very hard to interpret. Does that make sense?
AUDIENCE: It's more of a psychophysical sense. So what is the longest [INAUDIBLE] and still have that [INAUDIBLE]
LAURA SCHULZ: I don't think any of us [INAUDIBLE] like what's the whole psychophysics of the [? communications ?] [INAUDIBLE] just as for causality [INAUDIBLE] agency, [INAUDIBLE] really can affect how you would interpret that event. I think it speaks to the kind of [? precepts ?] that give rise to certain kinds of representations. And that's an important question that [INAUDIBLE] but it does ask under what conditions do you, in fact, have those representations? [INAUDIBLE]
PRESENTER: I'll talk a little bit about those kinds of things later on today. Those questions are part of what motivates a certain [INAUDIBLE] To say, well, just like Laura was saying, you're expectations might only be aspects of object knowledge. It depends on general [? tiers ?] which can be graded about how the world works. Like when you say that objects are permanent. They don't [? weaken that ?] resistance, what you mean is all of us are meaningful but you can't [INAUDIBLE] you expect they're still going to be there.
You can imagine, of course, various ways not to be there, but I think [INAUDIBLE] and those can interact in time. I don't know if that's what you were asking about. But those are going to be [? attracted ?] with time. The more time elapses, the more room you think there might be for other things to come in to move that object out of the way. And I'll show you--
PRESENTER: --afterwards in a slightly different paradigm. Yes.
AUDIENCE: Yes. In terms of a connection between agents and goals. It seems like agents do a lot of things. Do the things that they do usually have a goal? Or do they attribute goals to agents they see in general? Like if you're talking about an agent, could you [INAUDIBLE] Does the baby assume that that spontaneous movement probably had a goal or that was just a spontaneous action without a--
LAURA SCHULZ: Well, as when [? Susie ?] had that experiment, where there were no other [INAUDIBLE] agency moved spontaneously and insofar as this kind of [INAUDIBLE] tension, the babies didn't treat it as having a goal. They didn't see that as intentional action. So there's lot of reasons to think that spontaneous movement by itself isn't quite enough.
But again, if these instances are probabilistic and great, then after some contact with some type of prior knowledge or there's some other type of agency, they might very well might treat it as n intentioned agent. In some contexts, [INAUDIBLE] So again, if you think of these as graded representations where there are a lot of different things you might use--
Associated Research Thrust: