176 - Elizabeth Bonawitz: How to Have Fun While Studying How Children Learn so Much From so Little

0:00Adani:Welcome back to the Stanford Psychology Podcast. I'm Adani and for this week's episode I had the pleasure of chatting with Elizabeth Bonawitz. Elizabeth is a professor of learning sciences at the Harvard Graduate School of Education. She completed her PhD in cognitive science at MIT and her research spans many topics and methods. She bridges computational approaches and neural methods with cognitive development theory to understand how children learn so much from so little. Elizabeth also cares about leveraging the science to inform educational practice. I've been working with Elizabeth for many years now, and I would not be a researcher if it wasn't for her.

0:42Adani:So this conversation was a special honor for me. Maybe most remarkably, Elizabeth's enthusiasm for research and her humor and levity are really contagious and make everything she talks about really engaging. So I hope you'll enjoy hearing from her and learning about her work. Without further ado, here's our conversation.

1:29Adani:Thank you so much for joining us on the podcast today, Elizabeth. Thank you so much for having me. I'm excited to chat with you today. The listeners have already heard a little bit about all the fun different things you study and your research, right? So play, curiosity, exploration, and just learning and development writ large. And you use a range of methods. And nevertheless, to start with, I just want to ask you, how would you describe what you study and why you study it? And also, I feel like there's always a question here around like, why study this with children? Don't just children say a bunch of random stuff, do a bunch of random stuff?

2:02Adani:What can we even learn from doing research with them? What's your take on that? Children do say a bunch of random stuff, but that's part of what makes it super fun. So I study children because the primary question that I'm interested in is how are we as a species so smart? How are we so intelligent? How do we learn so much from so little data? So what's really fascinating to me is that, you know, we come into the world, whether or not it's a blooming, buzzing confusion, as William James might have said, or there's some knowledge that's in there. Nonetheless, as infants, you know, there's so little that we can do.

2:41There's so little that we seem to know about the world. And yet by the time we're four years old, by the time children have just circled around the sun four times, so they're just starting preschool, they have these incredibly rich intuitive beliefs about other people. They can reason about people's desires, their beliefs, their actions. They have intuitive beliefs about how forces work. So they have expectations about how a ball will fall or what kind of mass or exertion you have to put on something in order to get it to move. They have intuitions about biology so they can reason about growth and illness and disease.

3:16They have all these built up intuitive theories, we might call them, about how the world works. And they can explain and predict and counterfactually reason. And even we might, as scientists say, intervene, but explore the world in these incredibly principled ways. And so to me, that's incredibly rich amount of knowledge to build in a very short amount of time. That's the learning problem that I think is so cool, how we learn so much from relatively so little in such a short amount of time. And it's something that's unique to our species. Only human beings can explain and reason totally abstractly about things.

3:53Can we see, you know, the wind blowing the trees and design sails and sailboats and sort of build on these kinds of observations that we have early on in childhood to build intuitive theories that allow us to predict and explain and reason? So if you're interested in this learning problem, how it is we learn so much from so little, then studying children is really the natural place to go because you're getting right at the heart of who are the most intelligent creatures that we know of in the universe. and can we sort of understand what their predispositions are, what the starting points are like, and what are the mechanisms that must be in place that allow them to get where they're eventually going to get, which is, you know, this sort of rich knowledge that we have as young children all the way through adulthood.

4:39So I think children are really the core participants for my studies because I want to answer that core question about what makes us so smart, and they're the smartest things that we know of in the universe.

4:52Adani:Yeah, that's such a fun summary. I wonder actually listening to all the things you just shared about sort of these rich intuitions even young children have in these various domains, having intuitions about the physical world, about biology and all of that stuff. So what actually is there left for children to develop in terms of their thinking and reasoning? It almost sounds like in many ways they are pretty similar to adults, maybe, like some of the intuitions they have, but I imagine there's stuff that are like domains in which they are like ways in which they aren't quite like adults. So what is there left for them to develop in?

5:23It's a great question. So there's a question about enrichment, like how are we building on knowledge that we already have? And there's different kinds of knowledge. We can memorize facts about the world, but we can also form intuitive beliefs that are causal and abstract. act. And there's also questions about what are the mechanisms that allow us to learn? And so what I think is probably not changing throughout development is sort of core learning mechanisms. The way that our brain processes information, the way that we're integrating new knowledge with prior beliefs, those core mechanisms, I think, are just part of the human condition.

5:59They're what allow us to be so smart. But we have a bunch of stuff that builds on top of that. So for example, even in terms of learning mechanisms, as we go through childhood, we become more aware of our own uncertainty. We're able to reason about our reasoning, sometimes called metacognition. We form the ability to plan actions, to switch beliefs quickly, to inhibit prior beliefs, things that are associated with what we'd call executive function. Those things are coming online and can also help us continue to revise and reform our beliefs. And then a lot of the beliefs that I talked about, our intuitions about physics or biology or theory of mind, many of those are instantiated in early childhood and kind of stay with us as adults, but we also enrich them in lots of really important ways.

6:45So children develop intuitive theories, but often those intuitive theories are wrong. And it's because the world might give us misinformation that makes it hard to form the right beliefs, or it might be because they're too simple. It might be because they're not fully connected up. And so there's lots that's still developing as well.

7:05Adani:Yeah, there's actually another aspect I've been thinking about a lot lately when it comes to developmental science, which is, so a part of what I heard you sharing is that, you know, children have intuitions about certain things, and these can be further enriched and can be further revised as they go through education, but also as they just go through life. And in one sense, also, we as researchers studying development, we may have had intuitions about what children are capable of in various ways and how they learn. And in some cases, we were wrong and we had to revise our own understanding of those things.

7:33Adani:I'm really curious what you have seen maybe in your own time just in this field, but also before what you've seen as some of the major, I don't know, conceptual revisions among people studying these processes. In which ways maybe were we really wrong about what we thought about children before and how is that different now? Yeah, I think there's a classic tension that existed in developmental psychology for many years, which was a question about whether or not children's beliefs are relatively intransigent. So whether or not, you know, that's a fancy word for saying whether it's hard to change your beliefs, whether like it's hard for me to revise my beliefs and learn and whether kids are just sort of stuck in their wrong thinking about things on the one hand versus sort of other camps that sort of focused on this idea that no, no, no, no, children's beliefs are fast and flexible.

8:24They're revised really, really quickly and really effectively. And so there was a sense in which, you know, the focus was on sort of characterizing, is it that they're stuck, right? Or is it that they're fast and flexible? And so one thing that I think was a big change or a new way of thinking about that problem came up sort of as I was coming up as a graduate student some, oh goodness, 25 years ago at this point, was the idea that we could use these kinds of computational models and we could use ideas from other fields like linguistics and cognitive psychology to understand that learning problem in a different way.

9:04And so the computational model in particular that really informed our thinking was using Bayesian models to understand learning. And the core idea, there's a couple core ideas behind Bayesian models. One is that it's capturing the idea that our beliefs are relatively probabilistic. So data is ambiguous. We can't always know the truth of the world. And as a result, we might not know exactly what's the right model that gives rise to the data. And so as a learner, we see some data and our job is really hard. We have to figure out what's the right rule that produced that data? What's the right theory or intuitive causal story, the right explanation to have for that?

9:42And so Bayesian model says, okay, we can capture the idea that our beliefs are probabilistic. We might not know for sure that one belief is definitely right. Instead, what we can do is believe that a couple of beliefs are possible, and we can believe them with sort of relative weight. So I might not know for sure that I can throw out one belief, and I might not know for sure that another belief is right. I can still believe it more than something else, but I maintain this sort of probabilistic nature. And because they're probabilistic, the Bayesian model is just a mathematical prescription for saying, look, the way that learning works is I start with some prior beliefs about the world.

10:17So I have some explanations that I think are already more likely than other explanations. Then I observe some new data and I say, how likely would that data have been if explanation A was right versus explanation B? And I thus update my beliefs in accord to however much I previously believed in this belief and how much the data now supports some alternative belief. And that integration that just combines to form what's known as the posterior probability, the probability that a particular explanation is right given some data. And so this is really important as a model to come back to this opening problem in developmental psychology because it actually helps us understand why in some cases it might look like children are really fast and flexible learners and in other cases why they're really intransigent.

11:06So if I'm a kid and I think, oh, I think I have a good intuition, but I'm not totally sure what the right explanation is, and now I see some data that feels really unambiguous and really compelling, that Bayesian model says, go ahead and update your beliefs strongly, and you can really believe certainly in this new explanation. Instead, if I'm a learner and I have a really strong belief in a particular explanation, and the data is kind of ambiguous or noisy and isn't sufficient for me to overturn that explanation, then it's going to look like I'm really intransigent, that I don't want to change my beliefs, that I'm really stubborn.

11:43And so Bayesian models basically just provide a simple framework to think about the ways in which our prior beliefs and the nature of the data that we observe are going to interact. And so what I think has been a sort of revolution to the field is rather than getting into debates, no, they're fast and flexible, no, they're stubborn and slow, right? It's to say, okay, let's try to actually characterize the nature of the learning problem and the nature of the learner in each of these cases. And we can predict and explain when and why, in some cases, they might look flexible and fast, and in other cases, look slower and less flexible.

12:17And so I think that was a big change in the field, at least in terms of my research program. And I think there's been a lot of other research that's come on. It's called rational constructivism is one way that people have named this idea. That's really, you know, in the last 20 years or so, there's been an explosion of literature re-understanding and redefining the learning problem in terms of those ideas. And so I think that was a big shift in the field.

12:45Adani:Yeah, there's a lot in there. Also, I'm glad I learned a new word, intransigent. I actually did not know that word before. So that's cool. One thing I am curious about, though, with this, and I imagine some of the listeners are thinking about this too. I think there's this popular notion that we, as people, as adults, especially as we grow older, we become more stuck in our ways. We become less flexible in some ways in our thinking. We kind of stick to our beliefs. So is there something to that sort of beyond childhood that we become more transigent, less transigent, more transigent as we become older?

13:17More intransigent. Intransigent.

13:19Adani:There we go. Yes. So Bayesian models are basically capturing all the data that we've observed, not just in this one learning instance, but over the course of our lifetime. And so if you've seen lots and lots and lots and lots of data over the course of your lifetime that has given really compelling evidence for a particular belief, then you should be more stubborn. It should be harder to overturn your beliefs because you've formed more certainty. So, for example, you know, if I walk into a science museum, this is an example I think Laura Schultz once used 20 years ago when I was her graduate student.

13:54You walk into a science museum and you see that there's a hat floating in the middle of the air, right? You don't want to look at that and think, oh, my entire theory of physics is wrong, right? This one little piece of evidence is enough to say, actually, there's no such thing as gravity and no such thing as any of these things, right? No, instead you want to look at it carefully and say, okay, maybe something about this context or something about the hat, maybe it's getting blown up by air, maybe the magic show is actually on, maybe it's being held by a string, right? So it's actually rational that it should be the case that as adults we're a little bit more stuck in our beliefs.

14:32That being said, it's also the case that as we mature, we also develop certain kinds of cognitive abilities that might make learning more fast and more flexible. So if I have a more developed executive function skill, so I'm good at using sort of frontal low planning areas to quickly switch from one rule to another or to be more inhibited when I'm thinking through things, that can be really important when I'm in, say, a novel learning scenario. So if I learn a particular rule, A, and then I start seeing data that there should be a new rule be, if that's well-developed, then I'm going to be better at switching to this new rule.

15:15And so that's something that actually is the reverse of being more intransigent as we get older, that children who have more developed executive function skills and in general adults are more developed than children, those children are actually better able to switch or to revise a previously learned or stuck belief, holding all else equal. But your point about adults being a little bit more stubborn? Absolutely. And that can be rational. The other thing that we do as adults is that we tend to limit the kinds of hypotheses that we consider. So Conwin has a way of talking about this thinking fast and thinking slow, that we have this sort of system one, which is, you know, just reasoning quickly, reasoning from probabilities.

16:00And then system two is the critical one that notices inconsistencies and that tries to pay attention to these wrong things. And so if you have a belief that you think is generally right and generally doing good enough work for you, not only are you more likely to just use that as an adult, but you're also more likely to sort of seek out confirming evidence for it. You're more likely to sort of remember the evidence that's consistent with it. And so you can actually sort of trick yourself into generating evidence that's more consistent with your belief and thus lead you to a sort of more polarized or stronger belief in the data than you might otherwise have.

16:35We're really good at humans explaining away things because we don't want to spend the mental energy of having to go back and find a new right to belief. And so that's a sense in which adults can be a lot more intransigent and less optimal or rational than children who might be less likely to do that or less socially or emotionally committed to maintaining a particular belief. Maybe it doesn't associate with their identity or something else that matters for believing something.

17:03Adani:Yeah, I think one of the things that is coming through with all this that I also found really cool and exciting when I first started learning about some of these ideas is the notion of learning really being about finding ways of making sense of the world around you and finding ways of navigating the world. Because I think a lot of us students, you know, we typically know learning as something we do sort of in school, like kind of academically oriented learning. But this notion feels much, much broader. That said, I know you are also someone who cares a lot about how we can take some of this developmental science and our findings to translate it into real world settings like education and just more generally to benefit children's lives if that's fair to say.

17:42Adani:I'm curious what you have seen as some of maybe the exciting parts and challenging parts of trying to bridge that gap or do you even see it as a gap? I don't know maybe you don't but what has that been like for you to work at that intersection? Yeah, I mean, I think the core question from cognitive science, the one I opened with is this question of how we learn so much from so little, what makes us so smart? I think in education, a core question that often comes up is, you know, why does the same evidence not work equally well for all learners? Or why, you know, how do I optimize teaching in my classroom so that all 20 of these very different kids that are coming from very different backgrounds with very different expectations, with very different starting knowledge states and capacities, distractibilities, whatever it is, you know, hunger for that day, whatever it is.

18:31how do I find a way to sort of teach for all of them? And this is a huge challenge of the sort of public school system in many, many countries and in particular in the U.S. And so part of my research program has been trying to think about, you know, what is it about what we learn from cognitive science that can help explain and understand these problems that teachers face? So how do we, you know, given that we have better models or understand how learning works in general, all, can that also inform this question of why learning doesn't always work equally well for everyone and thus how to optimize or how do we choose evidence that's going to work well in classrooms?

19:10And so, you know, just to start with the initial example that I gave about the fact that you can have different prior beliefs. And so in some cases, your learning is going to look fast and flexible. And in other cases, it's going to look more intransigent. I'm going to use that word a lot today. It's going to be totally encoded. Everyone will have learned a new word at the end of this episode. Okay. More stubborn, right? So from the perspective of schools, well, understanding what the child's knowledge state already is when they're coming into the classroom and recognizing that that's going to make a big difference.

19:39Susan Carey has a wonderful quote in one of her papers, and I'm not going to get the quote right because I cannot memorize quotes, but she basically captures the intuition where she says, you know, the hardest thing about teaching is not the blank slates, not the learners who come in not knowing, right? The hardest thing is the learners who come in with the wrong beliefs, because it's really actually hard to overturn the strongly held wrong beliefs rather than instantiate in a fresh tabula rasa, right, a new belief. And so this is a challenge for schools, is recognizing what beliefs my learners already have and how are they different?

20:17And so, you know, something that maybe schools can think about and teachers already do this a lot in their classrooms. So I always feel like there's a little bit of hubris as a cognitive science trying to pass on messages for educators because, you know, they have a whole wealth of knowledge that goes well beyond just this piece of the science of learning that's really important for classroom management. But one thing I think that's important is this idea of how do we understand and assess where our learners are right now? And how do we think about the evidence that's optimal for learner A versus learner B?

20:49And recognizing it may not be the same for each of them. So that's something that I think is important for educators to keep in mind just on that first idea. I have lots to say about play in classrooms and being exploratory and promoting curiosity. And there's all sorts of other things that I think are really important from my broader research program that I think I would also like to see more in more classrooms, but that's a whole other set of questions we could maybe get into or not.

21:17Adani:Yeah, no, absolutely. That's wonderful. I was just trying to remember, because I think I know the paper you're referring to, and I really like it too. One version of it, I believe, is the one called Science Education as Conceptual Change. If you blur here, I'll link it too in our episode description. And yeah, I think this idea of if you want to meet the learner where they are at, I mean, there is this sort of compelling idea that maybe they're a blank slate and you just sort of start from zero, but they also bring their own ideas and it's really worthwhile knowing what those ideas are. So you can sort of go with them from there.

21:46Adani:Actually, maybe a somewhat related question. So I think one of the things I always enjoy in hearing you give presentations, and I'm not saying this just because you're here, is I think you have a way of making your research really sort of fun and engaging and relevant to different audiences. And I think a lot of that matters if we're trying to sort of take our research into other contexts, right? Like practical contexts. And I wonder if you think there is any sort of secret or trick to communicating your research in such ways to sort of make it compelling for different kinds of audiences outside of just academic research.

22:19Adani:And if so, what are they? I'm going to get really meta in my own head now and realize that I haven't been doing any of the things I'm going to suggest. That's kind of me to say, first of all, thank you. I think the first thing is study interesting problems. I think it's important for anyone that they're answering a question that I think people find intuitively interesting. But what makes things intuitively interesting, I think, is the fact that it might be things that are a little counterintuitive or a little bit surprising, I think, make things a little bit more engaging for people. Concretize, so give specific examples, is something that I think makes it easier for audiences to immediately latch on.

22:58We, in our own fields, there's a lot of jargon that we don't realize we're using all the time. We say words without knowing that intransigent, right, that people might not have come across them. And so when you concretize things, you say, okay, here's a specific example of the kind of test we could give a kid. Now they can really visualize it. They can be put in the study, in the experiment, and they have a good intuition of what's going on. And I think, again, that makes it engaging. And the other thing is being pithy. That's P-I-T-H-Y. just in case my audio didn't come across clearly there, succinct.