177 - Alison Gopnik: How Can Understanding Childhood Help Us Build Better AI? (REAIR)

0:01Alison Gopnik:Welcome to the Stanford Psychology Podcast, where leading psychologists come to share their most recent work. I'm Anjie, and for this week's episode, we have Dr. Alison Gopnik. Alison is a professor at the Department of Psychology in UC Berkeley. She is also an affiliate professor at the Philosophy Department. In addition to her many groundbreaking findings, she is also a great science communicator. Allison authored multiple best-selling books, including The Scientist in the Crip, The Philosophical Baby, The Gardener and Carpenter. In this episode, we discuss one of her recent review pieces on the role of childhood.

0:36Alison Gopnik:We talk about how childhood can be viewed as a solution to the explore-explore dilemma. Allison also shared an interesting observation that made me think a lot. In some sense, grad school is kind of like an extended childhood where you get to explore many new ideas. Well, without further ado, here's our conversation.

1:17Alison Gopnik:Thank you so much for joining on the podcast today with us. So we will be chatting about one of your recent pieces on philosophical transactions called childhood as a solution to explore, exploit tension. And if you don't mind, how about we start with this relatively technical term here, explore dash exploit. What does it mean in the context of childhood?

1:39Anjie:So an interesting problem that people in computer science and neuroscience have discovered is this tension between trying to solve a problem, trying to get a particular outcome, and trying to get information that you need to be able to solve the problem. And this goes all the way back really to the beginnings of computer science. It was during the problem first came up in World War II when people were first beginning to try to figure out how to design computers. And there was a joke that they should drop the problem on the boffins in Germany because it was such a difficult problem that they would spend all their time trying to solve it and they would never do anything else.

2:20Anjie:And it hasn't gotten to be less difficult. It's still a really difficult example of a trade-off. So, you know, if you think about it, the kind of classic example is imagine that you're trying to decide, should you go to the restaurant that you know that always gives you a good meal or should you try going to a restaurant back in the days when we could go to restaurants that you haven't been to before but that might actually provide something that might actually provide a better meal. And that's a kind of classic explore-explore trade-off. And it's actually quite hard to figure out exactly how do you solve it?

2:50Anjie:How do you play off the possibility that you won't do as well against the fact that you'll get some more information? And one of the ways that typically computer scientists try to solve this is start out by exploring and then narrow in and exploit. So start out by just trying to get as much information as you can, consider as many different possibilities as you can, and then as you know more, narrow the range of options that you're going to choose. And the argument that I've been making is that you can think about childhood as a kind of evolutionary biological version of that solution. So we get this early period where we can look around, we can explore, we can do all sorts of things that look kind of random and crazy, which is what little kids do, where most of our energy and motivation is about finding things out about the world, getting information, trying out hypotheses.

3:44Anjie:And, you know, our immediate needs are being taken care of by these caregivers around us. And then as we get older, as we know more, as we have more information, we narrow the range of possibilities that we consider and we're more oriented towards exploitation. And, you know, the way I illustrate this sometimes is if you think about the curriculum in any psychology department, I'm sure this is true at Stanford as well. If you take the adult cognitive psychology class, you'll learn about things like attention and decision making and planning and memory. Those are the things that you'll hear about.

4:18and maybe there'll be like one class about learning at the end of it whereas if you take

4:24Anjie:the cognitive development class it's just the opposite most of what you'll be learning about is grammar induction and statistical learning and concept formation those are the kinds of things we think about in cognitive development and and I think it's because kids are not very good at planning if you if you were giving a course about kids planning and decision making most of it would be about how bad they are. But they're really good. In fact, in empirical experiments that we've done, they're better than grownups when it comes to learning.

4:58Alison Gopnik:Yeah, if you don't mind, I guess I want to kind of go back to this, like the starting point, which is about this idea of childhood as the solution to this tension you just described. Because when I was reading your piece, I can't help but wondering this concept of childhood, it reminds me of a book by, I think it's by Neil Postman on the disappearance of childhood, where basically he was arguing that the concept of childhood is basically a social artifact and it really kind of varies across culture as well as the historical times. So for example, in medieval times, a kid might be considered already as an adult when one reaches 12 years old.

5:33Alison Gopnik:So I wonder when you're kind of talking about childhood in the context of as a solution to this problem that sounds very universal, So have you encountered any kind of cross-cultural differences in terms of how people conceive childhood and what kind of role it should play?

5:48Anjie:Well, I think it's important to point out that biologically, childhood is a really important biological concept. Now, it's true that we have a notion of childhood as a cultural concept and people's ideas about childhood have changed over the centuries. But it's a biological reality. So if you look across many, many different species, one of the most distinctive things about species is what biologists call their life history, how they develop over time. And one of the really striking biological phenomena is that the longer a period of childhood a species has, the more likely they are to be flexible, to learn a lot, to be smart from our anthropocentric human point of view.

6:27Anjie:And humans are kind of at the end of the spectrum. We have a much longer childhood, even in forager cultures, for example. we have children are not producing as much food as they're consuming until they're about 15 years old. And among chimps, by the time chimps are about seven, they're producing as much food as they're consuming. Now, of course, in our culture, 15 still feels like it's kind of a child. So our 15-year-olds aren't producing anything. Right? Like they're not contributing to the family income. So from that perspective, you might think, oh, well, you know, children, 15 year olds 12 year olds are going out and hunting and producing food and it's true that's very different and in lots of cultures by the time kids are seven or eight they're beginning to be apprentices to the kinds of things that adults will do and they're actually starting to contribute but it's still I mean it kind of makes you realize just how striking childhood is that you know nobody thinks that 18 month olds are going to be useful workers low expectations so all the stuff about cultural uh you know differences in childhood is talking about you know seven eight nine ten year olds that it's true that they're working more in other cultures but even then if you're preparing that to primates um comparing that to say our closest chimp relatives nobody thinks that seven-year-olds are self-sufficient like there's no culture in which you'd assume that a six-year-old is going to produce as much uh food as they consume it's going to be able to just take care of themselves with no other kinds of support.

8:05Anjie:Whereas if you're looking at chimps, chimps really are self-sufficient by that age.

8:10Alison Gopnik:But I guess from another perspective, one thing that I was considering is that I feel like when we are comparing human children and like chimps children, I feel like these two species, they seem to be facing different problems. So for chimps, it might be like it's their higher priority to gather food, to produce. But I think especially in the children in the Western or modernized society, what they're facing, their struggles might be like schooling. They need to like succeed in their test and they need to kind of deal with a lot of like the stress associated with like the education system. So in that sense, I feel like a lot of the pressures on the kids today is kind of, I don't know.

8:49Alison Gopnik:Do you think that it's kind of in a work against with this idea that childhood should be a period where kids explore and have like extensive play period?

8:59Anjie:Well, one of the things that we can do is compare, again, you know, the life of children in say, foreign cultures versus in contemporary industrial cultures. And it's an interesting, complicated story because on the one hand, so the first thing is that, you know, the kids that I've studied are babies and top, so up to about preschoolers, up to about five or six. And I think it's funny that we actually call these children preschoolers, that period preschool. Obviously, school is a very recent invention, but it does seem as if there's this kind of transition around six or seven, where the, I think it's that early, early period, the infants and the preschoolers who are really doing the most exploration.

9:43Anjie:That's really all they're doing. And they're not doing very much goal-directed activity at all. Now, Now, as you get older, then you start seeing this switch to a kind of apprenticeship where their kids are starting to learn the important of older techniques of the adults. And you even see that in the neuroscience. So you see more and more prefrontal control, for example, in that school age period as kids are getting older. And we know that that prefrontal control is very associated with things like planning and focus and decision-making and all the executive kinds of capacities. Now, in our culture, it's very strange because we have this idea that we have a much longer learning period.

10:25Anjie:And I think to some extent that's true. So even up until graduate school, right? You're not producing a lot of food, right? You get to have this really extended. And some of us, our scientists get to do this for our whole lives. So it is true that that period has been, that period has been extended. But it's also true that the way that our educational system is structured, there's this kind of strange things, which is that children are learning to be good at school, which is a very goal-directed activity that requires a lot of focus that isn't like free exploration and play at all. It's just that the goals are getting a good mark on your task rather than the goals being getting a, you know, succeeding in a hunt or gathering as much food as you can.

11:16Anjie:So I think what's happened is that especially for kids at school age, but now increasingly people are pushing this model down to preschoolers. The pressure, the kind of narrow goal-directed activity is even greater than it is in, say, forager cultures, where even though kids are doing useful things, they're also wandering around, looking at lots of people, getting lots of chances to explore. And that we're narrowing it, even though that goal-directed narrowing, it's kind of interesting paradox, it's supposed to be in the service of learning, but it's not actually the kind of exploration that's best for learning.

11:54Anjie:It's really just learning how to do another, just learning how to do another goal, just not a goal that actually is very useful to anyone. I can't even spray it on your chest.

12:06Alison Gopnik:I do want to clarify on this concept of Go-directed a little bit because I feel like in the development psychology literature, like people talk about Go-directed all the time. And we were considering like, there will be talks about how infants will do this kind of Go-directed reach. And I feel like the kind of the Go-directedness or the lack of Go-directedness in play seems to be using a more abstract level of Go. So I guess I'm wondering about, Because when I'm observing kids when they're playing, sometimes it looks like they have something in their mind. They want to put blocks together. They want to build a tower.

12:39Alison Gopnik:They want to build something. They are not just like randomly like touching all the blocks by random. So I wonder like when you're talking about play as something that does not intrinsically have a goal, what are the goals that's in this context?

12:53Anjie:Yeah, so that's a really good question. And I think one of the big debates, one of the big just empirical research programs now is how much is what's happening that the children's goals are different so that their goals are things like getting information. And we have a lot of recent evidence, a lot of beautiful studies showing that if you sort of and this is part of the point about the Explore Exploit tradeoff is getting information is a goal. So you might, you know, you have to get up and go out and do things and try things to be able to get information. but it's a very different kind of goal than maximizing utilities think about classic economic description trying to trying to just get you know achieve something that will be useful to you that you know will be useful to you so one question is how much are the children's goals actually oriented towards utilities towards getting particular outcomes versus how much are they oriented towards getting information but another question you could ask is just how much random variation is this?

13:53Anjie:So one of the things with kids is that it is really fascinating and we've been looking at this and this comes out in computer science literature as well. If you want to solve this explore-exploit trade-off, one way to do it is to have a kind of meta goal, which is, all right, now I'm going to go and find out about, now I'm going to go and get information and I'm going to use my executive function to do that. But another thing you can do is just say, the hell with this. I would just try something for no reason and just do things in a kind of more random, stochastic way. And both of those are ways of trying to solve this explore-exploit problem.

14:31Anjie:And when you look at kids, there's this really interesting sort of balance where you're right. They're not just behaving randomly, but they're behaving with more variability than adults would. And there's a lot of noise and variability in what they're doing. We don't yet know exactly how it is that, you know, what do you have a kind of meta idea? This is this is like a not coincidentally something that comes up in and say mindfulness traditions. Right. What does it mean to have a goal of not having a goal, for example? What does it mean to have your intention be? I'm just going to let my mind wander wherever it wants to go.

15:08Anjie:So so I think that's a really interesting question about what's happening with the children, how much of the children, and how much stochasticity and randomness is there and how much just being directed towards information is there. And there's another really interesting piece of this. We've got a grant from the Templeton Foundation to look at play. And one of the ideas that my colleagues are thinking about, computational colleagues, so again, play is another example. So I think kids do much more than adults and young animals do more than older animals. And yet almost by definition, the definition of play is that it's not useful, that it's not work, it's not producing something.

15:45Anjie:So there's this puzzle about why kids are doing it. And one idea is that if you look at something like goal-directed behavior, what play lets you do is try out lots of goals when you're not really seriously trying to get the goals. So if you look at the child who's playing, as you were describing, the child has ideas and the child will even sort of set up problems for themselves to solve. But the point seems to be the activity of solving the problems rather than actually having the problems be solved. So, you know, the example I give this in our house is what we call Addy chess. So Addy is my youngest grandson, Atticus.

16:25Anjie:And he, his big brother likes to play chess. And Addy, when he was, he's a little older now, but when Addy was about three or four, he liked playing chess. But the way that he would play Addy chess is he would take the chessboard and he'd do something like put all the pieces in the wastebasket and then take all the pieces out of the wastebasket and then put them on the board and then he would put all the you know white pieces in one corner and put all the black pieces in another corner um so he was setting up these little chest-like goals for himself about what can i you know i can see that i could see that augie my my big brother loves doing things with these pieces what kinds of goals could you have about taking these pieces and moving them around on the board.

17:10Anjie:And there's some very interesting work that we're doing with our colleagues in AI about how that kind of setting up internal goals, if you look at something like meta-reinforcement learning, in a way that's the way that computer scientists solve that problem, by having a system that sets up a whole lot of goals for itself and then goes and tries to accomplish a variety of goals and then has the kind of apparatus in place so that it can deal with a new goal. And that's one way, a good example of the exploit problem is if you just take an AI, for instance, an AI agent, you just give it a goal and, you know, classically, like the way that deep learning or something or reinforcement learning works, you just get it to solve that goal.

17:56Anjie:You can get the agent to do a really good job at achieving that goal. But now you give it a new goal and it can't generalize. It does what computer scientists call overfitting. It just is really good at doing that one thing, but it can't deal with something that's new. And if you do this kind of meta task where you give it lots of different arbitrary goals, then it's going to have a repertoire that will let it deal with novelty and generalization better. And, you know, when you were saying the thing that from an evolutionary life history biological point of view, The thing that humans are better at doing than anything else is exactly being able to deal with novelty and generalization.

18:33Anjie:And you could think about childhood as being this period when you get to do this kind of meta-RL operation of trying out lots and lots of different goals, including completely arbitrary ones like putting all the chess pieces in the wastebasket and getting the skills that you need to be able to accomplish those goals and then generalizing them. So a thing that we're thinking about with play in general is this idea that play is about all the different kinds of operations, cognitive operations you need, whether it's accomplishing goals or having hypotheses or considering possibilities. And play is a context in which you can try out many, many different ones without having to worry about their actual outcome.

19:15Alison Gopnik:That's so fascinating. And it also kind of like make me wonder if adults in the society should be given more opportunities to play. But I guess in addition to like all the other things that adults should do, like we don't have a lot of free time. We need to like fulfill our like social roles and all the obligations. But looking at the adults, like what kind of play behaviors can play the same like equivalent roles as what the plays are doing for kids? Because I'm thinking, even if we're given the same amount of time and same amount of leisures, we'll probably not be engaging in the same type of play activities as young children do.

19:52Alison Gopnik:So, yeah.

19:53Anjie:So I think that's a really good question. And we know that adults, again, across many different cultures can shift back and forth from being in this kind of exploit mode to being in something that's more like an explorer mode. So when we tell stories, one of the things you could think of storytelling as being is exploring a set of possibilities the way we do when we're kids. When we have sports, you could think of that as being kind of like this childhood thing of set up an arbitrary goal and then see how skilled you can be in accomplishing that goal. Um, so I think it's for, for, for the, for your audience, I think, um, graduate school is a really interesting example, maybe one of the most striking ones for adults, exactly of this explore-explore tension, because what you want with graduate school and what graduate school sometimes is like, is it's a real period of exploration.

20:50Anjie:You're protected. You've got, you know, you're feeble. But you're feeble postdoc salaries enough to keep you, you know, alive. And you can just try out lots of different ideas. You can find out about new areas. You can do new things. And I think it's probably true that, you know, 40 years ago when I was a graduate student, that was more true than it is now. On the other hand, you could think of graduate school as being, oh, no, I need to get the papers out so that then I can be competitive on the market and I need to be able to get grants. And I think there's a tension that you can really see with science.

21:29Anjie:The whole point of science is doing this kind of broad exploration and play. play. But because we're all adults, we turn it into this exploit task of, am I going to get the papers out? Am I going to be, you know, the famous scientist in the National Academy? Am I going to get tenure? We sort of turn it into this exploit process. And I think there's some real questions about how we keep that balance, even institutionally, even for scientists, so that scientists can feel enough freedom to be able to play. And it's interesting, when you think about an institution like tenure, the idea behind tenure is exactly supposed to be that tenure is like a return to infancy, right?

22:09Anjie:So the idea is supposed to be once you get tenure, mommy will take care of you, right? Once you get tenure, you can play, you can try all sorts of ideas because mommy's going to keep feeding you. But of course, in practice, because we're human adults and we're very well-tuned to exploring, that isn't quite how it turns out.

22:28Alison Gopnik:Yeah, that does remind me of the piece on children as scientists is the idea that children are not small scientists, but scientists are this kind of grown-up child. So I think this is a very interesting idea echoing about what the play and exploration should play a role in the current science institution.

22:47Anjie:To be fair, I think scientists do do that naturally. That's the best part of being a scientist is when you're in a lab meeting and you have a million different ideas and you don't know which one's going to work and you can get to play around with them, even if our toys tend to be more expensive than the kids' toys.