Listen: Stella Christie on "Learning by Comparison"
Earlier this fall, Assistant Professor of Psychology Stella Christie delivered a lecture on "Learning by Comparison." In it, she uses both comparative and developmental evidence to discuss the process of comparison, the results of comparison, and the use of comparison in everyday life.
Christie studies cognitive development and her research examines how the mind acquires knowledge through analogy and comparison, the effects of language on cognition in development, and cross-linguistics and multilingual acquisition. Christie runs the Cognition and Development Lab at Swarthmore, where she and her students research how infants and children acquire concepts and language and how language development shapes children's learning about the world.
Stella Christie: Okay. Today, I'm going to talking about one line of my research, which is learning by comparison. I want to start with these adorable ducklings. Just this summer around July, my research community as well as the media got really, really excited about these new findings, and what these findings are actually ... show us is that adorable ducklings have abstract thoughts because they can spontaneously perceive relational, abstract relational thought, abstract relational concept like same and different.
Had the journalists ... This is really indeed exciting news for my research community, and I see the media actually pick it up right away. I would say, had the journalists actually dig just a little deeper and ask me, they could make the story even sexier and more sparkly, because just two years ago I actually published work that showed that two and three year olds humans could not do what the ducklings can do, okay? The ducklings and the humans are, were casted in the exact, almost exact same way. I'm going to walk through it so you know what is it like to test abstract concept among very young children and ducklings who could not speak.
This is the test. This test is the same for the ducklings and the children that I tested. This is called relational match to sample task. So, if I give you a sample AA, and two choices, BB and CD, and I ask you please pick the choice that matches the sample. And A in the letter can stand for anything. Which choice would you pick? BB, thank you. Yes. I will ask you to participate a lot so you really know what kind of test we do.
So obviously this is a very, very easy test. Right? And why is it called Abstract Relational Concept but it's called the passive Abstract relational concept because you match not based on things that you can see like color or shape but you match based on relation. There's a relation sameness in AA and there's a relation sameness in BB, so you match the relation to the relation. So this is called the relational match to sample task first coined by Primak in 1980 to test these very fundamental relation because you need to know same and different to go around this world very productively to test this very fundamental relations with animals.
So do animals have abstract concept? And only recently they tested this with ducklings and only two years ago, we actually tested this with children. So this is with the ducklings. In case you are interested, what they did is using imprinting method. The ducklings were imprinted either to a different concept or to the same concept, it doesn't matter and then they will be tested. So if they are, let's focus just on the side, if they were tested on different concept. I'm sorry, on the same concept, BB, then now they see completely new things. They see they are different color, they are different shape, CD or EE and if they know about the relation here then they should prefer this relation EE.
And what this study by Martin and [inaudible 00:03:14] and signs show that they actually do. If they've been imprinted to BB, they will go to EE. With the children, so that's how it looks like. With the children, the two and three and four year olds that I tested, we didn't have to do imprinting. In fact, humans do not imprint but we can just simply ask them. So I gave them a picture card that contains AA and then there's a BB and a CD, which simple shapes that they know and I simply ask them right away, "Which of these is the match for the sample?" And when we did this, we were absolutely certain at the beginning of the experiment that of course children would get this. Nobody has ever tested this since 1980s. The assumption is this is a really easy task and we just wanted to have this as a baseline and move on to something else.
But much to our surprise, and this is why all the children were tested, much to our surprise, this was the result that we got. On the Y Axis, you see I draw the proportion of the time you choose the relational match, which is the BB. Okay? And 50% is the chance line. So, you can see that the two and a half and the three year olds, unlike the four and a half year olds, were just choosing randomly. They were at chance and we are very surprised but we have replicated this study to get the same results, out of a lab. Including a lab at [inaudible 00:04:39] actually replicated this study after my study and found the same thing. So for two-year-olds and three year-olds they did not spontaneously perceive relational sameness.
We could either go say, "Well, kids are just dumber than ducklings." Or we could think about something else and say we were not giving up on our kids yet. So when we taught, we came up with a hypothesis that perhaps we couldn't spontaneously perceive it, but what they could do is they could actually learn this very quickly. So we were very interested in the learning mechanism. So one of the learning mechanism that I will be talking today is comparison. So, what we did in that original study is we gave them comparison. So we tested a new group of children but they are of the same ages. And this is important. It's not that the children grew up and become smarter like you do. No, these are the same age children except we gave them an opportunity to compare.
So same class as before but you see now there are two samples. Okay. And we encourage them to compare. See how this one looks like this one? And usually two and three year olds didn't answer anything because they're young but we still cast them anyway and we ask, Which of these two are the same like the samples? Okay. When we did this we got a very surprising result. Again, with just this very simple manipulation of comparing two samples, children are now, the two and a half year olds and three year olds, again same ages that you saw before failing this relational match, the sample match, are now successful. Okay? Just given comparison, they are now choosing above chance, reliably above chance, their relational sameness. So they could learn to perceive this relation sameness from comparison.
In other words, this study shows that comparison can give you, what this is called, abstract relational thought. If this preposition, if this hypothesis is true, it can be very, very powerful. Because abstract relational thought, as abstract as it is, paradoxically, they are very, very completely present in our lives. So relational concept is very pervasive. It's not only same and different, there are a lot of relational concept. You should know the same and different to go about this world very successfully but you also use relational concept in many things that you may have not realized before. So here's just a couple examples. In language, every time you use verbs, you actually are evoking relational concept. A verb gives you, the verb gives, gives you relations between the giver and the givee. If you understand a relation wrongly, you may not understand the verb. So all verbs are actually relational concept.
If you use prepositions, if I tell you that the book is on the table versus in the drawer, you are evoking a different kind of relation. The on relation and the in relation are very different from each other. In different languages this relation may be coded together like in Spanish you will have on and in as en but in other languages as in Dutch, you actually have the on separated into two different kinds of words. So preposition are also in code relations, you need to be able to know relational concept whenever you use the prepositions. Even nouns, things like bridge and barrier. You do not describe it in terms of how a bridge looks like but you describe it in terms of its real relational properties.
Obviously, things like in Chemistry and Physics and Mathematics, in many domains, relations are very prominent. In Chemistry if you have the same atoms, identical atoms but connected different in the same way but with different relations, you get different molecules like [inaudible 00:08:43]. Right? You get, and the relations dictate this. When you read maps, you need to know the spacial relations. You also need to know the spacial relations between the landmarks on the map actually corresponds exactly to the spacial relations in reality. So relations between relations are needed even to be able for you to understand map reading. To be able to read maps essentially and know what it means.
Lastly, even in art, right? You use relations a lot to judge, to produce art. You use compositional relations in order to judge whether the art is worthy or aesthetically enough and also to produce it. So relational concept is very pervasive. This is what makes studying relational concept, how do we get through the abstract relational concept, really fundamental, interesting and challenging.
So today, what I want to convince you is that comparison ... Or actually I'm putting it as a question, can comparison account or promote learning of all these abstract relational concept? In order to do that I actually want to look at the theory, about what is comparison? From a cognitive perspective what actually is that we are doing when we are doing comparison?
Traditionally or the traditional account of cognitive comparison is this, that basically comparison is a process of intersection of common features, much like venn diagram. So if you're comparing a bicycle and the skate board over here, you will list each of this object, it can be object, events, whatever, you will list features of this objects like their blue wheels, yellow, gray, so on and so forth. Same with the skateboard and then you count which features are actually overlapping with each other, like a venn diagram. And you will get blue and wheels as result of comparison. So this is traditional account intersection of common features. But you probably sense right away that this kind of account is problematic. So if I actually give this comparison that I gave to the children, there's none of the features are actually overlapping.
So using this traditional account of comparison, you will actually not get anywhere. You will not get to the common abstract structures. So my [inaudible 00:11:18] a long time ago actually proposed that comparison was not only intersection of common features but more importantly comparison is an alignment process where the objects or the events that you're comparing is not just a mere list of features. But more importantly, it has features but it also has relations that connect the features. So instead of representing the object as a series of features. I want to represent the object as features with connected relations on top of that. Such that I can actually describe it as wheel, as their features at the bottom, and then these two wheels can move together that's like things that connect them. So they became a structural canals. And when you do that you, you do an alignment process. You still like things that look alike together, so you still align the wheels with the wheel. So things that are similar go together.
And this alignment also promote alignment at the higher structural level. So to move together now is also match with the move together. So I'm doing this very simplistically and the theory. But the point here is that there is features and you align the features that gives you common structures. So the abstraction of common relational structure is something that you get from the process of alignment. Such as, this are both vehicles, which you will not get if you just merely doing [inaudible 00:12:56] of common features.
So, this is a nice theory, but this is work. We have to care about whether that works or not. And what do we mean by whether it works. So we remember the challenge these two ask with a comparison, which I now say is a structural alignment process can promote learning of all kinds of relational structures, not just same and different but a lot of different kinds of relational structures.
So I can go about testing every different kind of relations, compare and contrast and give you this and that, but probably my life as a scientist wouldn't be long enough to do all of those. So I have to go about a smarter way. And a smarter way is test the most logical stringent question. And one of the most logical stringent question to answer this test is to ask whether comparison can give you learning of new relational concept. If it can give you learning of new things, of new relational concept, then you can be more confident that really this learning mechanism can promote all kinds of learning.
So that's our task. We gave four year olds, in this case a three and four years olds, bizarre novel relational concepts. And what do I mean by bizarre concepts? Well, you can see. The relation here is black on top, white at the bottom. You don't have sir anything like that in english, it's not encoded in the lexicon of english. Here small on top, big at the bottom. Here is symmetry, which is not novel for you but we knew that it was novel for three and four year olds.
So this are all novel for the three and the four year olds. The objects themselves are not novel, the cats are not novel, but their relations are novel. So our question is whether children can actually learn these relations. So, the task is very similar as before, it's matching through sample past. However, this [inaudible 00:14:52] quite sufficiently more difficult, and why? It's because now instead of relational match and completely something else that is just distracting you so you can be at chance, I also gave another viable option. And what do I mean with viable option? What I mean is that, there's actually nothing really wrong if you decide that ... If I ask you, which of these two matches this one? You actually want to match this picture card, right? Because you can match the cat with the cat. And that's not a wrong thing but I'm interested about whether you actually prefer the objet match or the relational match. So you may actually have a different preference. You may not actually prefer relational matches.
So the three and the four year olds when we tested this, actually chose ... You probably could guess, the object match 98% of the time. So we gave a lot of trials to the three years and 98% of the time they say obviously this is the match, they were very confident about that. So this gives us a really interesting scenario in which we want to know, well, if they are very focused on this object match, how could they ever get to the relational matches?
So, our test is comparison. So in another condition, which has another group of children, we gave them the comparison condition. So, now there are two cats and the test is about the same for both conditions. Obviously, we need to have the same test. We notice something very interesting here, which is if comparison is merely intersection of common features like the traditional account, now, when I give you the cat picture card and the dog picture card, you may actually be more likely rather than less likely to go with the object match.
But our theory is that comparison can also highlight common structures. So even though you may be tempted to go with the object match, we actually predict that the comparison should give you the relational match. And indeed that's what we found. So the three and the four year olds, in the comparison condition in the red bar are more likely that the group in the solo condition to choose their relations.
Some of you skeptics out there will immediately notice that, hey, wait a minute, the comparison group gets two things, so maybe it's not maybe about alignment, it's not about the process, but it's about just getting two things, more is always better.
And this is a legitimate concern, and we have to answer this concern. Then indeed whether the advantage that we saw in the comparison group is truly due to getting alignment or it's just seeing more exemplars. So to answer this, we tested by saying that, if comparison as alignment is truly what really matters here, there if you get two exemplars, two samples without being able to align them, then the children should stay preferring object matches, just like those in a solo group.
But if alignment is irrelevant as soon you get two things, then they should behaving like the comparison group. So, to answer this, we created the new condition in the middle, sequential condition. So the children in the sequential group get two exemplars, but it's presented one after another. In other words, even though they get two exemplars, they could not spacial temporary juxtapose them. They could not align them. It's very difficult for our children to align these two.
US adults maybe to do, so it's not about the sequential-ness of the method, it's about whether you can align them or not. So the children could not align this we know that. But they got two exemplars nevertheless.
So they couldn't align them, but they get two exemplars nevertheless and we also re-run a new group of children in the solo in the comparison condition, we replicated our previous results that the comparison group is again much better than the solo group in choosing the relational matches.
And now, what about the sequential? Well, the sequential looks actually like the solo group. Facially they look high, but this is not actually significant from the blue bar. So, when children ... And significantly, and they're different from the red bar. So when children get two exemplars but are not able to compare them, then they behave as if they only saw one exemplar.
In other words, the alignment really matters. It's not just more is always better, but alignment really matters. And now we have also answered our stringent test that comparison can give us learning of two relational concepts. And is not merely intersection of common features, but more importantly of common relations. So it convinces us that it can really give us learning of a lot of different kinds of relations. And more is not always better. But you need the alignment for this to happen.
Okay. So, this is a very good learning theory at least at the beginning, but as a cognitive scientist you will not be happy at this stage. You also want to make sure that your learning theory is viable. And what I mean by viable is that in all experiments, I gave the children, the learners the comparison. So I gave them explicitly, here is what you should compare and what do you get from that.
In order to propose a learning theory, you must also account for the accessibility of the learning theory. In other words, the humans actually use this learning theory. And so, one thing we can ask is what initiate comparisons? So if the comparison cannot be initiated by other things except by the experiment they're giving you explicitly, then that's not a very satisfying learning mechanism.
So in the second part I want to ask what actually initiates comparison? Is it a viable learning mechanism. And I'm going to give you the answers right away, and then go through about trying to convince you that these answers make sense.
So, is comparison a viable learning mechanism? Do other things other than explicit experimenter are doing the comparison, initiate comparison? And the answer is, yes, for two things I have proposed. One, is that language can invite alignment, and second, attention to object matches can invite alignment. So these two things there are available all the time and in fact they can invite alignment.
So, how does it work? How does language invites comparison, invites alignment? It's actually not that complicated. So my theory is this, if you have the assumption, all of you have the assumption, that same labels refer to same things. And you probably do because we notice from studies, we also know that even young children have this assumption, that the same label refer to same thing. And that means that whenever you hear common labels, you actually are invited to compare. And once you compare you can get to the relation. That's the theory, I already said that.
So something like this, if you hear idempotent in elevator, somebody was in an elevator and say, I just did an idempotent procedure, and you don't know what idempotent is, doesn't matter, you probably ignore it. But then the next day you went to a computer science lecture and you hear the word idempotent again. You may not know what idempotent means, but because you assume that same labels refer to same things, you're being invited to compare.
What's the same between the elevator scene and the lecture scene that I just heard idempotent again? So this is where common labels invite comparisons to happen. And while a comparison happen, it can [inaudible 00:22:59] getting you the relational concept.
Notice something that's really interesting and powerful about this, that is you don't have to know the meaning of idempotent. Any common labels will do. And that is very powerful for learning. Because that means the learners do not need to come with a lot of background. The only thing you need is the assumption that same labels refer to same things, and you can be invited to do comparison even with labels that you don't know the meaning of.
So this works great for novice learners, for young children. Essentially, you don't need a lot of background. And that's really powerful. Obviously, so even novel labels should work. Obviously, we need to test this, and here's one experiment that showed that. So, remember this task where young children, two and three year olds failed, right? If they just got this, they just randomly chose. They chose at chance. So what we did, is we actually gave a label this time. This is novel label, they don't know what that means. Does mean anything, it's just something that we made up. This is [inaudible 00:24:07], and which of these two is a [inaudible 00:24:09].
So assuming that children actually know like names refer to alike things, the moment they hear which of these two is also [inaudible 00:24:18]. Then the question will invite the child to compare each of the alternatives to the standard. And when this comparison happen, hopefully, this will allow them to do the alignment process and get to the relational meaning. And that's what we found when the two and a half to three year-old to four year-olds we just did this to make sure that everything is good and nothing funky there.
And so when we tested the two year-olds and the three year-olds, we just did this very, very little intervention. Just gave them here is the [inaudible 00:24:49] which of these two is also a [inaudible 00:24:51]. They now chose relational matches above chance. Even though this word is actually completely novel for them. So the common word can invite them to do comparison.
The second mechanism that I proposed about why comparison can really be initiated within everyday learning is that attention to object matches can invite you to do comparison. What do I mean by attention to object matches? You've seen this, you've seen this in this example where when children were given these two choices, they strongly go for things that look similar. That's what I mean by object matches. Perceptually, they're matching. Things that you can easily see right away as a matching thing.
And this are very, very strong tendency in children. It actually even don't go way in adulthood, by the way. When we tested adults in this kind of things, we would get about 87-90% adult who've seen the relations, but they will be like 13, 10% adults actually choosing their object matches. As I said, it's not a wrong option. It's just the preference, the saliency change as you go [inaudible 00:25:58] more mature.
But for young children, this is very salient. And it acts like a double edged sword because obviously at this point they don't perceive their relations because they're really, really focused on the object matches. But it also actually works to actually invite comparison. So the theory goes like this, if I give you this analogy, a lioness to a cub is like hen to a chick, right. This is an analogy, there is a parent, child, relation. But if you are young you probably will never discover the relations because the relations that which is on top of there are really, really not salient to you. But if I give you this analogy instead, duck to duckling is like hen to chick, I can actually match the feature, the things at the bottom, the one that you can perceive right away to be similar. I can match this at the bottom, and once I match at the bottom, it's much easier to see the structure at the top.
So this kinds of ... Even though they have the same relational structure, the corresponding object matches will help to eventually get to the same relational structure. So that's why the object matches even though initially it can hamper your perception of relations it can also invite you to do comparison and alignment in the first place.
So, how did we test this? We test this by again using the relational match to sample test, and we wanted to precede it by given them an easy initial trial-outs of literal similarities. The same relations and same objects. The children, the two and three year olds are now getting this trial-outs. And forget about the relations, what you can see is that the triangle match the triangle. So when children were given this, then they 100% of the time will be able to deduce very correctly. And then they went on to the hard task, which again, without any kind of interventions, without any kind of manipulations, they all failed to do this. The question is, whether this kind of manipulations will allow them to see the relational matches.
Now, as a good psychologist you'll always have to have a control, so we don't want the conclusions to be, it's simply that they to more training, right. So we did a control group where the control group didn't get the easy matches, but in fact, they get just more of the actual test. So if it's just about getting more practice, in fact, the control group get more practice. They went on to the same test, two groups of children, one in the object match first, the other one in the control group.
We don't get results from the two and a half year olds because apparently their four initial matches is actually not enough for them. But for the three year olds, it actually made a lot of difference. So when actually are seeing the three year olds, when the three year olds who saw those simply four easy initial match just like this easy object match. They could actually went on to do the more difficult relational matches.
So we saw a big difference between the object match and the control condition, which is just a chance. See getting more practice didn't help them, but getting the right kind of practice. It's not a practice, it's getting the right kind of things, which is the ease in similarity help them.
More recently, we actually asked these very intriguing questions, what would happen with animals? What happens with other animals, and we got a very intriguing result, that I'm suggest that perhaps this strong attention to object matches may be unique to humans. So when we tested orangutans, chimpanzees, bonobos, and three year-olds, and we did this with identity procedure, we actually get very intriguing results.
So here is the study. The study is just a very easy hide and seek procedure. It's a spatial mapping procedure, and we wanted to do it very simple because we wanted the study to be identical for the three year-olds and the apes, you can't ask many questions to the apes. So you have to make it exactly identical because this has to be comparable. We're comparing the two of them.
So in the first phase of the study, basically I'm going to hide an object or food in the hiding box, and then the participant has to such it in the finding box. And there are two rules about how you can search correctly. So in the first phase of the study, which this is the first phase, I wanted to know about only relational abilities. So how do you go about relational in your relational ability. So the participants saw this, are either in this align condition, or initiative condition. So in the align condition, when I hide a grape, this is the favorite food of the orangutans and the chimpanzees, and the bonobos they like grape very much, so they're very motivated to do this that if I hid the grape.
So when I hid the grape on the top container, then you should find it in the align top container, top-to-top, middle-to-middle. Top-to-top, middle-to-middle, bottom-to-bottom. So, that's the align rule.
In the other conditions, different groups of subjects, I created an arbitrary rule that I called the shifted condition, top to middle, middle to bottom, bottom to top. And simply the question is that even though [inaudible 00:31:30] you should not expect one rule to be easier than the other, if you favor alignment then you should be better in the align condition than in the shifted condition. Do you follow me? Great.
So, what did we find? So we tested apes, and children, great apes and children in identical ways, and the black bar you see the align conditions, the checked bar is the shifted condition, and Y-axis is just proportion of correct searches. Chance is 33% because there are three locations.
What you can see here is that very clearly for the pan this is our chimpanzees and bonobos, we have to group them together because otherwise we don't have enough participants, it's much harder to get participants from chimpanzees and bonobos than from children.
So when we tested it, you see a very clear pattern, than the pan and the children are equally good in getting the relational alignment. So in this task, in this comparable spacial mapping task, even pan, even other great apes are really good at getting spacial relational matches. The [inaudible 00:32:43] was not successful, and we don't know why but they were not doing well in the task.
In the second phase of the study, we wanted to ask a question that I just post to you, which is about the object matches. So now, we kept the rule but we actually change the containers. So now there is green container, yellow container, red container. And you can see right away what does it do to the task. So if you actually are attracted by the object matches, you will ... If you are actually attracted by doing green to green, yellow to yellow, red to red. Then you will be really confused in the align condition.
On the other hand, in the shifted condition, if you like green to green, yellow to yellow, red to red, you should be doing very, very well. So we purposely created that way, the same participant state in the either align or shifted conditions. So they already knew the rule in fact but if they really, really favor object matches, if they found object matches salient, then they should do very well in the shifted condition but not in the aligned condition.
Because the presumption is that other animals, non-human animals, we're animals too by the way, right. But non-human animals are more concrete than us, than humans, then the field has always assumed that obviously other animals will do the object matches very easily. And there were some indications, like my collaborator and I just [inaudible 00:34:22] in some previous studies that made us doubt this. And for the first time ever, we directly tested this hypothesis, that maybe it's different in other animals, even though we always thought that they're more concrete than us.
So when we tested this, remember again, if you are, this is the really, really easy test. And if you are very concrete, you like object matches, you should do very well in this. And we actually predicted in the beginning that everybody should do very well in it. Everybody, meaning all the great apes including human children.
But what we found was not that, and this is the fun of doing experiments. What we found is only the children were very, very attentive to object matches. The great apes were not. It's not that they're not affected because you can see their performance actually went down, so they were somewhat affected but they did not find the object matches as salient as the young children.
This was a very surprising result, and again we have replicated this but in a very surprising so they give us a large implication. And that large implication is this, that the opportunity to compare does not come easily to other species. And when the opportunity to compare does not come easy to other species, at the very least using this route, you don't get a lot of relational learning.
So the lack of frequent comparison that other species are facing may actually be contributing to the fact that what we can do with our relational thinking is very, very far from what other species can do. Sure, the duckling have [inaudible 00:36:11] and difference, but they don't have language, they don't have verbs right, they don't have nouns. So this route can actually give you this gap between humans and other animals in terms of their relational thinking.
I'm yet to talk about my language work but I want to give applaud to Jordan's yes, yes work on how this attention to object matches may affect language acquisition. So this is coming up in [inaudible 00:36:35] post session on Thursday and Friday, so I encourage you to go to Jordan's post and learn about how object match is going to affect language acquisition.
So, so far what we've learned is that comparison is a structural alignment process resulting in learning novel relational concept and initiated by common labels and attention to object similarity. I provide very specific mechanisms about how it can be initiated in the first place. So it's a viable learning theory.
And in this last part, what I want to do is see about how is comparison used actively by the learners in everyday learning. And in particular I'm going to look at how this cognitive process of comparison manifests in the social learning and a social domain. I'm going to look at information exchange, social and friends, and also cultural learning. And most of this works are actually done by my students here at [inaudible 00:37:33] so it's very exciting a time for us to look at this question.
So the first kind of comparison information exchange came from my personal experience. And this is my personal experience when I was about to buy a car. So the salesman ... Saleswoman, over here. The saleswoman over here said, oh, yeah, that car over there has good mileage and spacious. Whereas this other car that you are considering is very safe and has great resell value. And I stop and think, what do you mean that this car is very safe, do you mean to say that the other car is not safe? And my bad is that I didn't ask that question to her. But my bad is like, if I ask that question to her, she would probably say, oh, no, this other car is also safe. But then I would actually counter back by saying, well, then you're not giving me any good informative comparison.
So in other words, these things that she gave me were not alignable. And because the comparison was not alignable, it was not informative for me, or at least it didn't help me to make a decision because they were not alignable.
So the alignability of the comparison gives you informativeness. And here is how it's actually even prevalent. If let's say, you're learning the adjective blue, by the way, all children have to learn the adjective blue at some point. So pretend that you're learning the adjective blue and I have these two kinds of comparison. I call them low align because they don't look like each other and high align because they look like each other.
And I say, this is blue, and this is also blue. And I know the objects, the only thing I don't know is I'm learning about blueness. I don't know what blue means. So if I give you these two comparisons, which of these two comparison is more informative for you to learn blue, the one on the left, or the one on the right. The low align or the higher align? Low align, great. You got the intuition right away.
That the low align condition is more informative. So now, if I change it now to a contrastive wording. So I say, this is blue but this is not blue. You can see right a way that this the higher align is the one that's more informative.
So these two situations is ... I created these two situations because I want to make sure it's not about two objects or two of the same objects, but it's about how the alignment give you the informativeness. So now you get this intuition that some alignments are good, and meaning that they can very good information, some alignments are less good in giving you good information.
So, in the experiment we actually ask, when children are learning from others, when they're ... Because children really like learning from others. We too, we also learn from others. So it's a very important process in our social learning. So when children are learning to others, do they actually prefer to learn from a teacher who use more informative alignment.
So the study setup is exactly what I showed you before, except that now we use novel adjectives. So there is Mr. Green, the teacher, or Mr. Purple the teacher. And Mr. Green will consistently use a low align cards, and taught using this is feppy and this is also feppy, it's a novel word. And he would use four different cards over here and all of them are low alignable cards. And Mr. Purple will actually high alignable cards, all of them are high alignable cards.
So here is in one condition, it's called the multiple exemplars because the sentence, the language is this is and this is also. In a different condition again, the same logic, I use this is but this is not. There is Mr. Green over there, and Mr. Purple over here.
So, our question here is this, can children actually prefer this teacher, who give better alignment in this condition, and this teacher who give better alignment in this condition. The test for the children after they saw teacher taught using two different kinds of cards consistently, is this, well, you know the teachers actually spoke the monster language, but I don't what to call this in a monster language. And clearly you don't either. And you need to ask one of the teacher, who do you want to ask, who do you trust to give you the information.
Obviously, children can just choose at random, which means that they're not sensitive to how the comparison gives you information. But that's not what actually we found. So when we asked them do want ask Mr. Purple or Mr. Green then this is what we found in the contrast condition, children actually chose the high align teacher, it doesn't matter whether it's Mr. Purple or Mr. Green, we counter balance it. Half of the children saw Mr. Purple using the high align, the other half saw Mr. Green using the high align.
But overall, they chose the high align, which is the more informative teacher. And the same thing in the multiple condition, it's not about the two objects or the different objects, it's about the informativity. So in the multiple condition, the more informative one as we have gone through together is the low align condition. And that's what the children chose when they ask, you don't know how to call this, who do you want to ask.
We also replicated the pattern when we asked them explicitly, overall who do you think is the badder teacher? Is Mr. Purple or Mr. Green the badder teacher. And we got exact the same pattern, which is really cool. And these are three year-olds, by the way. They're very young but they are sensitive to the comparison and they can actually make judgements about who will give you a badder information, and who to trust better based on the informativity of the comparison.
In the next study, what we did, when [inaudible 00:43:41] actually wanted to ask the question, well, what about this information exchange. So what about the children when they give information themselves. So we switch the role, the logic is identical but we switched role now it's actually the children are actually the one who's teaching. Okay. So everything is the same as before, except that because it's the children who is teaching, they have to teach something that they know. So it's no longer novel things, but they teach things like blue, shiny, round. And the children saw a video of someone who does not speak English. And they saw two different sets of picture cards. And they have to choose ... You need to pick the best card to teacher Anna the word blue. Which sets of cards do you to use? So now the children are the teachers.
And the question is, do they pick randomly or do they actually pick a card that is actually more informative? And what we saw here in the multiple exemplars, children pick randomly. So arguably when you are in the teacher role it's much harder, so we don't see evidence that they're actually using informativity.
But in a contrast condition, we did see that evidence. So, in the contrast condition, the children ... Three years-olds who are now the teachers are saying, I want to use this cards, and indeed those are the cards that is more informative. When they themselves are teachers they can also use the comparison to show which one is the most informative to teach others.
We asked about comparison also in thinking about evidence that we got from the world. So, this is work with [inaudible 00:45:30] was asking the question or noticing that, well, we have to make generalization sometimes ... Actually, a lot of the times we have to make generalizations. And guess what, the generalization that we make are based on limited data. That means, not a single one of you can encounter all the data that's available in the world, right. You have to make generalization based on the limited data.
And in other words, you need to be able to rationalize which evidence are good and which ones are less good. So the example is like this, if you're taking Bio one and Bio two, and Bio one and Bio two are both challenging, you're probably not going to be right away concluding that all [inaudible 00:46:14] classes are challenging. You will be a bit little careful to draw that generalization. However, if you took Bio one and Religion eight, and both of them are challenging, then you'll be more likely now to generalize that all [inaudible 00:46:30] are challenging.
At the very least, you'll be more likely to do that encountering Bio one and Religion eight, than in Bio one and Bio two. Do you get this intuition? So that means that the diverse set of evidence, you're more likely to generalize from a diverse set of evidence than from a narrow set of evidence. So this kind of inference, inductive inference for generalization may matter a lot in our social inference. So what [inaudible 00:47:01] is an actual application in the social inference.
So we told five and six year olds, here is a school that you haven't been to it's called the Geefy school, and this are the Geefy children. And you need to help me buy a toy that makes most of the Geefy children happy. So Geefy children like toy A, which is inside the back, we do not want to show them the toy because it's not about the toy. It's about the inference that you make about the people.
These Geefy children like Toy A, these Geefy children like toy B, so, which toy do most of the other Geefy children like? Which toy should I buy to make almost everybody happy, at least. And if you do the rationale that we just went through together, then you should pick toy A because this is the more diverse evidence that you race wise. And you should pick toy A and not toy B.
But this is what we found, children actually pick this only 30% of the time. So do you pick this 30% of the time, you could say, well, children were not rational being, they have to grow up in order to be rational like us, but notice that this is actual ... They were not choosing at chance. It's not that they were random, right. They're actually 70% of the time prefer that narrow evidence. So, it's not that they were random, this actually gave us a very interesting further question, why did they chose the narrow evidence? It's because they're not random, we can ask this question. And one of the things that [inaudible 00:48:44] proposed, is that, well, maybe they were actually doing this because this are children from [inaudible 00:48:54] community. And all our participants are white Caucasian. And maybe what the children are used to are also white friends.
So even though they are still generalized in the abstract sense, for them they believe they should be generalizing to the population that they know. And this population that they know are consistently most of white friends. And so it actually makes more sense to say, toy B, which are liked by this groups of people, by these groups of children are the ones that are going to be like, by the people that I know, the people that I have to be generalizing.
So, if this is true, then what if we actually tell them, explicitly that, you have to be generalizing to truly a general population. In the second study we tell them this, okay, so here is a picture of the Geefy children, make sure pay attention this is the picture of the Geefy children, and you notice right away, that the picture of the Geefy children contains mixed race. Have 50% black, 50% white, and when the children were told this ... This is a new group of children. Again, most of them come from [inaudible 00:50:07] community. All of them are white Caucasians. And we ask the same question, every thing is the same. Now, 60% of them actually chose the diverse evidence.
So they were actually not irrational. They were behaving on a set of believes, a set of [inaudible 00:50:24] that they have, and they can rationaly change their inference when you give them a different kind of evidence. And you can see how this actually makes a large implication. We haven't touched the adults in this. So don't tell your friends about this study because we're testing the adults from the [inaudible 00:50:44] community. Because it can make a lot of inference. When you make this inference, when you make this generalization, are you behaving on that familiar things that you have, or are you truly are being rational. Either about believes, or about in this case about the toy that you have, like about all kinds of things. Because we use this comparing of evidence for generalization in a lot of ways. So, again [inaudible 00:51:14] is also going to present some of this work in [inaudible 00:51:16] and she's doing a lot more work on this line. So if you have any thoughts, tell her.
Okay. But the widest comparison is not always informative. So this is work that Cynthia and I are dong together. And so, for example, if I tell you that I love the movie Foxcatcher and I hate Vanilla ice cream. You're not going to learn much about me. It's a very wide comparison I give you between a movie and a vanilla ice cream. But you're not going to learn about me. But if I tell that I like the movie Foxcatcher and I absolutely hate truly the movie the Martian, then you probably are going to be able to draw some inferences about me.
So, not all the time the widest comparison is actually good for you, or is actually useful. And so what Cynthia was asking was precisely this question. So we want to contrast between high alignment, and high entropy. In what situation you use high entropy comparison, in what situation you used high alignment comparison.
We just started this project, so I'm just going to tell you a little snip bit about it. We created two different conditions where we first tell Mom that Alex is a new student in your school, and during the last few weeks you gotten to know him well. You're excited to tell mom about your new best friend. But mom is really busy today, so she only has a very short attention span. And so she can only hear about one thing. So you have to tell her just shot. Do you want to tell her using the high alignment comparison, or you want to tell her using the high entropy comparison.
In the other condition, Alex is a new student, you know nothing about, we ask a six year-old. And you don't know him well, but his birthday is coming up, you want to give him a nice present and since he's shy you can only ask him once. And you don't know nothing about Alex. So do you want to ask him using the high alignment, or using the high entropy.
We predicted that in the pal condition, you should want to use the high alignment. Whereas in the fine out condition because you know nothing about Alex, you want to use the high entropy comparison.
So far what we got with a very small number of end because Cynthia just started using this study, we got completely the opposite. This is the joy of doing experimental studies. We got completely the opposite, so it gives us a very interesting puzzle to deal with right now. But also it might change when we run more studies. So stay tuned for more of these results. You can ask Cynthia in a few months what we got out of this.
In the last study, I want to share with you how comparison is actively used in a cultural learning. This was inspired by this kind of phenomenon where we learn much of our cultural idiosyncrasy through imitation learning. What the child is doing here is learning a cultural act, and the child learn this through imitation.
And you notice that the child is actually using ... The imitation actually happens in a group. So when the imitation happens in a group it raises the question of whether the relations among people in the group is going to affect why and when you actually imitate. Because something that's really, really salient from looking about the group is about the relations among people in the group. And our question is whether this can actually affect how likely it is that you want to imitate a particular group.
Think about fashion, for example. You see a fashion of using the hat backwards, if it's only one person, you probably not going to imitate it. But you see in a lot of people doing that, you probably will imitate it. And so our question is that, what kind of relations within those people that allow you to imitate it even more?
So we look at the most fundamental relations that you know very well from the very beginning of these talks. Sameness and difference. Which really in it something that leaped out immediately when you are looking at a group. So you see a group, they're all kinds of relations, hierarchical relations, and so on and so forth. But something that leap out to you is whether the group is diverse, very different from each other, the people inside it. Or whether the group is very cohesive and very similar to each other.
So Sona Kumar asked this question, and she's still doing this for her thesis. And what Sona did is asking whether you're more likely to imitate a diverse group or a similar group. So Sona created this video, which I'm going to play for you, I'm going to ask your opinion because I'm kid of ruining Sona's study in this case. So hopefully, Sona can get some data from this audience. So the children, like you are now, are seeing this video.
And then the child will see this video.
So, noble gestures, gestures that we made up, Sona made this up. And it has the same meaning, but the forms are different. The question is not about which form, but the question is about which group are you likely to imitate. Obviously, the groups give you a different form.
So, we asked the child this question. Now, show a puppet that you are a friend. So, which gesture do you want to use? And I do actually want to get your idea. Your sense. Who among you will imitate this similar group?
Okay, raise your hand, one, two, three, four, five, Sona we can take this as a data. All right, so who among you will imitate the diverse group? Okay, there's definitely a lot more diverse group. And l think that's because you think that diverse is more generalizable. But at the cohesive, the similar group probably think because when you're similar to each other you're also more cohesive. So think about clicks. A group that were very popular tend to be similar to each other, they're very clickish, and that may convince you to actually imitate them, rather than the other way round. In other words, both are possible. And this is why it's fun to do this study because both are possible.
Now, most of you are wrong in terms of five year-olds. So the five and the six year-olds, doesn't matter whether it's a chain or the fore hat, overall, they prefer to imitate the similar group. So, at least with five year-olds they prefer to imitate the similar group. We are now uncovering why exactly, what is the rationale for this. We're the first study who showed this, and there is a lot of follow up questions that we have to do, including doing the study with adults as well.
So I'm going to close by saying that one of the take home point that I want to convey today is that, we became sophisticated relational thinkers through learning. We may not have ... Or the two and the three year-old may not be able spontaneously perceive relational sameness like the duckling do, but they could learn to do so very quickly. And what I hope I have convinced you today, is that one of this learning mechanism is comparison. They showed in detail about how this learning mechanism can work to highlight common relational structure so you can learn a lot of new relational concepts in the world. And that it is very feasible because it can be initiated through things that are available to us humans, through like common labels and through attention to object matches. And that it really is a pervasive mechanism that we just started uncovering, my students and I started to uncovering about this is used actively in the social domain for social learning. How discovery the process really manifests in everyday learning.
And lastly, obviously, I want to thank many, many people. Especially, my students, who do a lot of the works that you saw here. Abigail Dean, who has been very, very helpful in getting a lot of participants, all the day schools and nurseries parents and children, and the fundings that make this research happen. And also all of you for being here today. Thank you.