 All right, thank you guys for hosting us. So just to quickly introduce our panelists, I'm Josh Tenenbaum, this is Leslie Cabling. Leslie is a professor in ECS and CSAIL and one of the world's leaders in robotics and machine learning and the interface of that, founding editor of the Journal of Machine Learning Research, just really one of the most interesting people that I know thinking about intelligence, especially in an embodied context. This is Laura Schultz, who is in the brain and cognitive science department and she's in charge of the early childhood cognition laboratory. She's one of the leading researchers in the world studying children's learning and not just learning, but all the kinds of activities of human intelligence, which you can see in adults, but you can maybe see them in their most essential and interesting way in young children. And this is Polker Agrawal who's just started at the MIT faculty just this year. He is a leader in robotics and reinforcement learning and has done a lot of interesting work around curiosity. And I think that's one theme that actually all of the speakers here are curious about. I'm sure that will come up under the broader theme that we're gonna be exploring of how we might go towards more human-like forms of learning in AI. We might not just restrict ourselves to learning, but that's certainly a theme that all of us here share. Also, we have this question system that you guys have had a little experience with, the Slido system. So the way we're gonna structure the discussion here is just go through a couple of structured questions that all of the panelists might have something to say about, and then at the end, we'll take audience questions, although I'll keep an eye on the questions that are popping up here. If any of them are particularly germane to our topic, we'll try to introduce them, or if any of you guys see them, then point to them. But we'll definitely make sure to leave some time for those at the end, okay? All right. Has this timer hasn't actually started yet? So, okay, there it goes. Perfect, all right. So the first question I thought we'd start off with is where do we need more human-like learning in AI? Where are places where current AI approaches might be failing or running into dead ends? Where learning in a more human-like way would be valuable? We might also think about what are aspects of AI that where maybe we don't want more human-like learning? Or where do we want more human-like learning and where are the complementary strengths and weaknesses there? So, we don't have to go in this order, but. I mean, the thing that, for me, I'm interested in robots that learn, and for that, experience is very expensive. And so one of the things that I'm super interested in is how to learn from less data. So people talk about big data as kind of an important case of learning, but also for people who make systems that interact with the real world, small data is important too. So being able to learn from less data is a big deal. Yes, okay. Yeah. Hard to argue with that. Well, so I'm stuck with humans. I work with the human learning. And one of the questions I've sometimes posed to my class as we think about what it will take to get AI to the level of human learning is if we succeed, what do we know, or what are our guesses about the possibility that if you really develop machines that can learn optimally from very sparse data in a very rich, noisy environment from just a few examples, as quickly and accurately as say human children, that we are not actually the optimal solution to that with all of our abilities and all our liabilities. And there are things that our machines have always done faster than us, like we've had calculators for a long time that are much better at us at a single kind of task. And now they're better at many different tasks, including playing go for instance, at least one task. If there were to be better at all the things that humans are, are we confident that there wouldn't be costs built into that also that would resemble human learning? So that's not really an answer to the question because that one version of the answer is, of course, we'd like better more human learning across the board. But I think that is the natural challenge to that question. Yeah. Yeah, I mean, so I guess, I mean, we all understand that humans don't start from scratch and they're actually using what they already know to learn new things. I guess like, I mean, one can take two views of that. One is we need to build in maybe a few things that are required for learning to happen. And the other could be, I mean, how do you develop algorithms which can use experience and kind of learn? And you can view both of them as kind of two different kinds of priors. When in one scenario you are building in a lot of that information. On the other hand, you are kind of trying to learn from your experience and what is the interaction between those two to actually enable us to do things like few short learning, learning from few examples and so on. I think that part would be... But like, oh yeah, good. I was just gonna say, I mean, some ways though that we might want to do non-human type learning would be, I mean, so I guess humans learn through culture a lot, right? So lots of information is transmitted across generations through culture. Through robots, we can transmit information by melding their brains, right? I mean, we could take the weights or the learning of this robot and of that robot and combine them in some way. So there's those things that we might wanna do that we have the opportunity to do with robots that we don't have the opportunity to do directly with humans. I guess the other point also is that when you're doing learning with robots or systems, we kind of approach different problems separately. For example, imitation learning and so becoming a separate problem. Reinforcement learning is a separate problem. Supervised learning is a separate problem. But all of us are kind of learning by these things together, you know? I mean, you are observing people, you are learning some skills from them, but sometimes you also get a reward. So I think having frameworks just don't focus on like one technique per se but actually making use of all the rich observations which are around is an interesting thing. At least in aspect of what we can learn from humans and learn in that particular manner. I think the word learning is overloaded, right? I mean, in AI, it often just means any system that comes to have knowledge that it didn't already have before, let's say, right? Somehow it gets better from experience. But in humans, human learning, what we normally mean by learning is like a thing that you do in your lifetime. Like a child learning or I'm learning in school or I learned this yesterday, I didn't know it, today I know it, so I must have learned something. As opposed to other kinds of processes in biology that build knowledge like evolution, biological evolution and cultural evolution. So when you guys are talking about culture and I think definitely to me right now, a lot of the especially like big data, big compute kinds of machine learning, they're effectively trying to, they put all these things together, right? And a lot of what they're doing is building in machines more of the kind of things that we might call evolution or what evolution has given humans. And that includes cultural evolution too. So we might, one thing we might want when it comes to more human-like learning, I think, might be to distinguish these different things and to say, well, what are the kinds of learning that you can actually do right now in the moment from very little data like a person can? Or what are ways of getting machines that could participate in culture? Either learn or somehow contribute to knowledge in addition to just say trying to evolve some of the very, in some sense, redo some of the very basic things that evolution has given humans and many other animals. Another place which I'm curious what you guys think is there are other things where it's well-known, lots of other people, lots of other talks here have been trying to address. Things that are, that problems with today's AI, failures of robustness, failures or just lack of explainability. You know, do you guys see those as places where learning systems that learn in a more human-like way might be more robust? Well, I mean, I think the lack of incrementality is the thing that bothers me a lot, right? So generally speaking, you get some data or you get a simulator, you get some environment and you train up your thing and then you say, here's my thing and it's great. And there's no way for that thing to actually just add on new understanding, right? It could augment its data set and retrain in some sense, but that doesn't feel to me like the right mechanism. Like, if I tried to teach you a new skill in the kitchen, I could probably teach you that and it would probably not interfere with all the rest of the stuff that you knew how to do. So how could we make systems that, I mean, to kind of build on what Paul Kitt said, do different kinds of learning and in particular, are incremental. I think that I'm kind of really hung up about that. Linking that incremental but also sort of modular or compositional, right? Right, well, modulating might be what gives us. Right, well, exactly. So it might be that instead of just learning in a sense of end to end, the system is trained end to end actually saying, okay, well, I'm gonna learn this part and then keep the others fixed or add in a new module or something like that. Maybe it's a way to get that kind of lifelong incremental learning. Yeah, I mean, the other aspect of it is like learning on a single task versus learning on multiple tasks, right? I mean, many times you end up having systems which learn on a single task. And in that case, the incentives are aligned for you to learn some heuristics or some kind of very particular decision boundaries which will overfit you to that particular task, which in turn will lead you to issues with robustness or in terms of you don't know whether they're going to fail or not. But if you have a system which can learn multiple tasks, when you would hope that some of those issues will get resolved to some extent, I mean, not fundamentally, but at least the impact of those issues might end up being lower in those scenarios. Yeah, yeah, and this, how all these things go together into, you know, call it lifelong learning or general learning. Yeah. I mean, it's so remarkable that humans, we learn so many different things and learning a new thing usually doesn't get in the way of all the other things we've done. Sometimes there's actually positive transfer, but the ability to just actually learn to do so many different things. I mean, I think it's fair to say at this point that we've had great advances in machine learning systems, but they all basically, each one basically just learns to do one thing. You know, a system which learns to play Atari games might learn one Atari game or maybe it learns multiple Atari games, but it basically just learns to play Atari games, right? Or a system might learn to play Go and chess, but basically it learns to play board games. And when you think of the space of all the things that a human can learn and does learn over their lifetime, at this point, any of the actual systems we have is just picking out one little small teeny corner of that space. Yeah, and I guess picking up on Leslie's point when she was saying about this discreteness in the learning, right? I mean, for example, when we are training our systems on ImageNet, we download this data from ImageNet. And I mean, at least for a long time, people said we want to build in all these kind of invariances into a system that we should be invariant to lighting and pose and everything. But if you end up considering like a baby, I mean, it is not exposed to that much diversity in terms of courses. It lives in a similar environment which could be composed of a couple of rooms, but it gets other added advantage that it can see the same object in many different lighting conditions from many other different viewpoints. And there is certainly information which might be available there which one could use to learn the representations and those representations might have a different kind of a nature than when you just throw a lot of data which is very diverse. But maybe, you know, you can't, you don't have the flexibility to kind of see it from many different perspectives. The returns a little bit to the earlier question, but one of the areas where you might want AI to have more human-like learning is to help us understand a human mind and human cognition, right? So in so far as one of the functions of AI is to do cool things in the world. But another potential function of AI is as a scientific tool to help us understand our own brains and our own minds. And to the degree that AI is gonna succeed at that, it's gonna have to consider the constraints, the abilities, all of the things that human learners do. And I think, as a scientist, that's what I'm most excited about AI. I think it'll do lots of cool things in the world, but what I would like it to do is help me understand how human children learn, how human minds work. And so many of the things that can result in human-like performance or even super human-like performance are very useful as tools in the world, but not useful as scientific models for understanding what we can do. And I think we all have reason to want that because one of the hopes and promises of AI is that when we end up with an injury or an insult or a stroke, it would be awfully nice to know what computations were happening in that area, such that we could recreate them in a way that would actually interface with the rest of our minds. I don't think we're there, but that is a possible future for AI. And that is part of, I think, what I would like to see in more human-like AI, not just for its own sake or because it might make our machines better, but it might help our science understand more. So yeah, so this is a good opportunity to segue to our next set of questions. I'm curious for the different panelists, what questions you might have for each other or each other's field. So from the people who are working most in AI, what would you most like to know or what do you think we most need to know or would like to understand better about how humans learn if we want to build more powerful, more human-like learning in AI? And then from people studying humans, would Laura and me to some extent, or me? You know, what questions would you have for the AI side or do you see the AI and machine learning world being most potentially valuable? You already gave one good answer to that, but feel free to give others. So I don't know, we could start from Polka to move this way. Yeah. For fun? Yeah. So I think that, I mean, I think this question can be answered at like multiple levels. I mean, the first level would be, for example, I mean, we end up using back propagation to train all of our networks right now. But I mean, there are many reasons to believe why that might not be biologically plausible. So what are the kind of the learning rules that we're using to learn, right? My second question could be that, I mean, if we imagine how learning is happening today, I mean, we have some data and we are trying to fit a model, but from the first day, we are trying to fit a parametric model. We have some parameters that we want to tune. And in some ways, I mean, you could imagine that, that if I have like three or four different tasks, I might first want to overfit to them or kind of memorize them. And only when I see that there's some similarities between the tasks, I try to find some relationships and learn a model which explains them. So I think this interface of memory and learning and how do they evolve with each other? I think that would be an interesting question to know. Then I think, I mean, the other question ends up, I think at the exploration versus exploitation trade-off. Like, I mean, when do we explore? What are the incentives for us to explore? I mean, which would kind of relate to things like curiosity that you were mentioning. And if you end up learning by exploration, by being curious, how does that compare with learning just based on rewards that you might get? Then I think a fourth question, which would be, like, I mean, I think there's like this, I mean, the debate which ends up happening would be model-based and model-free approaches. I mean, is it possible that if you are training a model-free system just based on say, rewards or some other criteria, do some things like models emerge automatically if you end up training on like multiple kind of tasks or are there some kind of primitives that just need to be there to encourage that to happen? Those are all really interesting questions about machine learning. I'm wondering which of those are questions for human learning, but let's have Laura weigh in on that. When I've been following the conversations and one thing is clear that machines do extremely well that humans cannot do, which is handle very vast amounts of data extremely efficiently and return answers that humans cannot. So this is hugely, hugely helpful if you're doing epidemiology or studying the cosmos or anything else. It's great that we have systems that can see things, can find patterns in data that in a reasonable lifetime we could not without them. What we do, however, is often think and learn without any data at all, right? We can propose new ideas. It's not just that we learn from very little data, which we also do, which is an interesting problem. We can get less and less and solve problems in data fiction, right? So children can learn words from one or two examples. They can learn causal relationships from a single example. That is impressive, but you can also ask them a question and they can generate an answer that they do not know to be true and that might even be false and that is nonetheless a pretty good answer to the question, right? And it can be a novel answer when they have never heard before. And you can look at your students and they're your colleagues and they can propose an answer to a question that you know won't work and that is wrong and false. And you can say, great idea. It doesn't work, but it's a great idea. It's not great because it's fitting the data. It's not great because it's prediction in the world. So why is it great? It's great because humans think, right? And we think of new ideas. We think of new problems. We do it in ways that are deeply and I think vastly underestimated how data-independent they can be. That is, in some ways, our distinctive adaptation of many animals who can do a lot of things with the observed data in the world. They can see, they can move, they can navigate, they can solve problems, they can manipulate things. They don't do what we do, right? So if we wanna understand distinctively human cognition, what I would like to understand from our machine learning and from our AI is how we can make what feels to me a qualitative shift in the ways we think about machine intelligence so that we can train and learn machines that are not just dependent on large amounts of data or for that matter, large amounts of reinforcement because we can stay curious and engaged in problems when we are failing for years, right? We can stay motivated when we are making no appreciable progress towards a goal. So then would you kind of advocate for, because some of these things that you said could be explained, for example, if you're a very strong prior. For example, if I find that the idea might not work but it is still kind of interesting, could be because the idea is implausible under the prior and therefore I might find it interesting and I might be able to analyze it and say it doesn't really work. So I guess I'm wondering, I mean, is it about like getting more priors in that would bump us there or is it something fundamentally different that you would say we are kind of missing over here? I think it's something fundamentally different. But what idea would be, I mean, we all kind of know it's harder to generate a good solution to a problem than to test one, right? And so if you generate even a sort of good solution to a problem and I test it and say, oh, that's almost a good solution, then I can be happy about that. I can appreciate the fact that your search method found an answer that I didn't know. So I mean, it could be like that. But actually defining what's a good measure of almost a good solution, actually that isn't trivial. Well, objective functions are a huge issue. Yeah, but what is the right solution? Because often, I mean, we've done this in some robotic control things and something where sometimes it's easy to see like some kind of physical problem solving or tool you setting where like your goal is to get the ball in this basket and like it almost got it to the basket. But you can tell that just doing that same thing is never gonna actually get there, right? You have to do some fundamental different thing. Or another time you can tell, oh yeah, it's just a small adjustment. I just need to drop it from higher or throw it a little harder or something like that. So yeah, so even just defining, what does it mean to say a task was almost successful? Although you're also getting at something different. It's not just about generating solutions, but more like generating questions or something, right? That's it. Part of, I mean, yeah, I mean, again, a lot of human learning starts from people asking questions, whether it's scientists or three-year-olds, right? Somebody asked on one of the questions that was going past here, is it possible to get human-like learning without solving human-level natural language? And that's a really good question, right? Because I think it's fair to say that depending on how you count, most of what we as humans learned, we didn't learn from sensory motor experience or a direct immediate experience in this moment of space and time, but we learned from people telling us things or reading things or asking questions and getting answers and then asking more questions. And that's how we get access to culture and all the knowledge that accumulates from others through culture. So I think that's certainly gonna be important. Well, there's a big question about how much you have to have, how necessary is it to have scaffolding from the learning that you got in the physical world to understand that or do you not? It might be that that's like, ultimately the main contribution, I mean, again, feel free to disagree, but to be controversial. One lesson, maybe, from studying human children is that for at least a lot of the things we really talk about when we're talking about intelligence, they might start from our embodied experience, mostly in the sense that that helps us build the things like language that then help us get beyond our bodies. And you could be born missing an arm or a leg or missing both arms and both legs and also blind from birth, but as long as you have some way to access language, the history of development shows that you basically come out to be a cognitively normal person and you might even contribute materially to human culture as Helen Keller famously did, right? So whereas if you don't have language, it's very hard to really be a full participant in the human world. But sometimes I guess the hard problem becomes this, I mean, how do you get like the symbols on which the language is based? Right, what I'm saying, maybe that comes from our embodied experience, that's what Leslie was suggesting, right? Yeah. But I guess, I mean, just kind of going to your point, right? I mean, sometimes you can get to an environment where you might not even have a way to get that bodily experience and you might still be able to speculate, I mean, what are the important things in this environment or how things might be working if you just passively observe them also. It could be that, I mean, the scaffolding that you're learning is not really specific to one or two condition, it's kind of a more general purpose, it provides with a tool, which helps you to scaffold in many, many different environments. Yeah, Leslie, do you want to weigh in on other things that you'd like to know about human learning that would be valid able to take them in? Well, yeah, I mean, the thing I really want to know is, so right now there's a kind of an ethos in some parts of the robot learning community that if you build in anything, it's either guaranteed to be wrong or it's cheating, so those are two different criticisms, but they're both criticisms. But I think you have to build in some stuff and I think probably the fact that my robot's gonna live in three space, I can't imagine making a robot that doesn't live in three space because that's here we are. So the question, but the question that I'm super interested in is, what is it that we can probably safely build in? What do we know is built into human systems? Now I recognize the building anything in is building in some limitations, right? So evolution built some stuff in, it helps us in the niche we live in and it might not help robots in other niches, that's very important, right? So human-like priors and stuff might actually not help robots in a whole bunch of niches. The other thing that I'm interested in is, what corners is it plausible to cut, right? So anybody who does AI tries to solve impossible problems. All of our problems are computationally hopelessly intractable if you look at worst case stuff. On the other hand, most of the things we try to do, humans can do actually pretty sort of well, probably not optimally, but there are some kinds of relaxations, some kinds of approximations that it's like pretty okay to make, like you can make them and things don't go terribly wrong. There are other kinds of approximations that are really, really bad. And so understanding the ways in which humans are apparently suboptimal in their reasoning and decision making, I think can give us license to be suboptimal in our reasoning and decision making. If one of my favorite examples is, if I watched one of you walk down the hallway, you would not do a perfectly beautiful controlled trajectory. And if my robot went on the same trajectory that you went on and I showed it at a robotics conference, people would yell at me and say, well, that's a very bad control system. And I'd be like, well, you know, it got down the hallway. So how do we make those, it's interesting to think about how we make those trade-offs. Yeah. I would say, yeah, both of those, very interesting. I started reading all these very interesting questions there and that linked my mind about the thing I was going to respond to. Corner cutting, or what can we build in? Well, yeah, yeah, and what can we build in? Oh, yeah, and you said, whenever we build something in, it imposes limitations. And yeah, that's certainly true in the sense of like, if you have a prior, then it focuses your inference. Some things are more than other things, right? But actually, I think a lesson from biology, some both classic and very recent work in human cognitive neuroscience and cognitive development, is that building in doesn't necessarily mean like hardwired unchangeable, right? So building stuff in is useful because it can give you a head start in certain places, like building in that the world is three-dimensional. But we have various forms of evidence actually at this point that, well, in fact, probably human brains and other animals are built in with that kind of information. If, for some reason, they don't have access to, there's a lot of plasticity, basically, too. So, for example, we're built to have a visual system. A lot of our brain is visual cortex, but if you are congenitally blind, then that part of your brain seems to get, sort of taken over by language, for example. Worked from Rebecca Saxes' lab here in MIT originally and Marina Bedney's lab. She did that work and then now is at Johns Hopkins. Has shown all sorts of interesting ways when you study the brains and minds of congenitally blind individuals who don't get normal visual input, then things which were, in some clear sense, built in, in that you can see them even in the youngest babies, newborns, well, then actually that structure is still plastic and can be reused for other purposes. There's all sorts of other very interesting examples of places where I think, in some sense, pre-wired doesn't mean hardwired, right? Yeah, no, okay, that's a good point. That's a good point. I think other, did you want to respond to that? Yeah, but I guess the mechanisms of how you end up changing those built-in things are also very interesting because sometimes the nature of how we know how to build in are things which are hard to change once we end up building. Right, so one question is, I mean, yeah, when you look at human learners, some things that are built in are very hard for us to visualize five dimensions or something like that, right? But other things are, that's not true for us. So I'm just understanding what are the kinds of learning mechanisms that can change things very quickly. You know, again, just in responding to some of the things you said, I think there's some implicit assumptions when you say what are the real biological learning rules, right? You know, for instance, which, or there was another one where you said, well, how can we go from learning a bunch of, you know, you start off learning just specific things and only then later on form generalizations and abstractions? Actually, you know, when you look at human children, it might not work that way, right? It's not clear that there are any learning rules. There's certainly mathematics that governs learning, like, you know, Bayes' Rule, it's hard to have a learning system that doesn't, in some form, kind of respect Bayes' Rule, even if it doesn't explicitly do that. But it might be that the real mechanisms of learning aren't like any simple, you know, that, Yoshua mentioned this earlier, like, I mean, it's a great idea, and it's one that's inspired a lot of us, that there are a small number of simple general principles of learning. But it might be that there's, like, some mathematical laws, which learning has to respect, but the actual learning mechanisms are, you know, maybe much more like practices, you know, Laura's talked about this, like, some of the ways, like, take science, so there's this classic idea that Laura and others have worked in the field, sometimes called the child of scientists, that children's learning is kind of like science. You have abstract theories, you have different competing hypotheses, there's things you don't know, and you do experiments, like just playing around with, in the lab, in the crib, to test out these ideas and let your curiosity drive your theory building in some interesting loop like that. You know, in science, we don't think there's a single magical, or even, there's no, like, science rule, right? There's culture, there's practices, there's a very number of different kinds of criteria, some of which can be formalized in objective functions, not all of them so far can, but somehow science still works, and it might be that human learning is kind of like that, and that doesn't mean we can't understand it from an engineering point of view, it's just that it may, the right way to understand it may not be one or a small number of equations or something we'd call a learning rule, or this idea that, you know, when children learn, there is one classic idea that isn't necessarily always supported by experimental data, sometimes it's actually the opposite, that we start off learning concrete specific things and then only later come to more abstract generalizations, but actually sometimes it seems like infant cognition, they, you know, they have the big picture first, they have the abstract understanding, like even, you know, very young babies, six month olds or younger, might distinguish even three month olds in some form, might distinguish the difference between an inanimate physical object and an animate agent who acts to achieve a goal, and only much later do they learn about specific kinds of objects or specific kinds of goals that people might have. And that's an interesting challenge for learning, we've sometimes called this in some of our work the blessing of abstraction, you know, either it's built in or maybe somehow we have learning mechanisms that can learn the big picture first and then fill in the details, yeah. Yes, I guess I was not really advocating for like that there's one single learning rule for everything, but maybe if we are able to figure out exactly what you said, right, what are the kind of learning mechanisms? Yes. Which are there, I mean, that might influence us in how we are developing our machine learning models, right? I think some of those principles would be very interesting to know. But I think it is a few important points. One thing is that there are things that are true of the world, right? There's physics out there, there's sources of sunlight, there's agents who act in goal-directed ways, and if you start building in some things that we know to be true of the world, you might be able to elaborate on them, expand them, maybe in some weird context, defeat or change them with evidence, but most probably those kinds of things are actually gonna get the world right because they actually are true of the world, right? There are objects, there are forces, there are agents, there are goals, and we have very good reason to think that humans and lots of other animals have a built-in representation of those. What's interesting about humans is we can take some of those built-in representations and especially through possibly natural language, learn and change them in very dramatic ways so lots of animals can represent approximate number, right? Approximately, how many berries on that bush? Twice as many, half as many. They can also sometimes represent exact numbers. How many eggs are in my nest? Our babies can too, they know number. It's not a concrete thing, it's an abstract thing. They know approximate ratios, they know small exact numbers, but what they don't have is 13 or 17, right? Or the idea that 17 is exactly four more than 13. To develop those kinds of concepts, you also cannot learn concrete information. If you just learn about these M&M's or these pine cones, you're not gonna have what it means to have natural number. Children do learn natural number, they learn it before kindergarten, right? They can count, they know what four means, they know what six means, right? At least in our culture in Cambridge. It's possible. That is not the kind of learning that can be explained by sensory motor learning systems. And it's not the kind of learning, I think, that is plausible without starting with other systems of number representation, right? Where you then begin to bridge the gap between those other kinds of systems. So, when I say there are very different kinds of processes in learning, I mean very different. There are things that we know are trivannals that we know are humans that are not being captured, I think, by the kinds of approaches in AI today, that I think are really important to the remarkable flexibility and richness of the ways that human beings learn. That was... Was it you? No. Objects and gravity. One more of our questions, and then turn to some of these very interesting questions which you guys can all see behind us. So, what are concrete steps? If we wanted to build AI that learned in a more human-like way, what are concrete steps we could do? So, one suggestion that's been made, some people are investigating, right? And it sort of relates to questions that Yoshua asked also earlier or ideas put forth. We know that human babies learn in an interactive way. They learn from actually interacting with their environment and especially interacting with other people. So, one idea that's been proposed is we should make some kind of virtual simulation environment, especially as we have AI methods that are kind of good for learning in simulators and games like that. So, one idea that's been proposed is we should make a simulation environment that's like the world of the baby and then unleash our AIs in it. Does that make sense? Is that gonna work? How could we do that? Is that maybe that doesn't make sense? Maybe we need a different approach. Any thoughts on whether, how to do that or whether it makes sense or what else could we do? Yeah, I mean, I guess the first comment would be, I mean, just comparing and knowing that two systems can perform the same tasks does not really kind of mean that they're solving the tasks in the same way. So, I mean, then I guess we'll become a game about how do you come up with those right tasks that will actually tell you whether the learning is happening in the same way or not in which in itself becomes a non-trivial question to answer. So, I mean, I guess, I mean, I wish there was like a simple answer to this. I mean, I guess simulators would be a good start, at least, you know, kind of trying to go towards the question. But, I mean, fundamentally we have to realize that, I mean, turing tests kind of scenarios where we have a turing test between AI and the human might not kind of really help us come back and say, okay, this is what I need to change in my system. I mean, I think it's, so simulators are a huge step up from data sets, right? So, because at least simulators are about interacting with something. So, that's like really important. And then I think variance is the next thing that's important, right? You know, Polke's been talking about multiple tasks and there's domains that have variability. But to me, right, the hallmark of, and I don't know, as a robot person, I don't actually care if the solutions I come up with are like the solutions that humans have, but I think to get to the same kinds of competences, what is critical is to have an enormous variability in the domains and in the goals or objectives that you ask the system to achieve. And that by having a big variability, then you have to arrive at representations and algorithms that are reasonably generic in general. And that's what gives us a kind of radical generalization to new problems. That's, I think, essential for testing human level intelligence or human-like intelligence. No, no, for training. We need the variability. But it is so interesting that, you know, you can grow up as a perfectly normal human baby and never leave the room that you're born in for the first few months of your life, more or less. Like you can, I mean, humans can have, in their lifetime at least, you know, a quite limited amount of variability compared to what they're actually able to then process. So evolution had to operate in the distribution. But the individual doesn't. Yeah, I know, okay, that's fair, yep. Okay, any other thoughts on that, or? I'm just thinking that the idea that if you just had the rich environment and you put a general purpose learning mechanism in that environment, they would just learn everything. In some sense, it was behaviorism, right? Like, okay, control the environment, control the reinforcement algorithm, you'll get learning out. And I think the history of cognitive science has been to say, funny, but in the same environment with the same reinforcers, different organisms learn different, very different things. Children of different ages learn very different things. It's not just about what's in the environment, it's about what's in the mind, right? It's, you know, what kind of information can you access? How do you represent it? How do you relate it? How do you combine it, right? So in exactly the same environment with the same learning mechanism, I mean, with a simple reinforcement learning mechanism, you can't explain why you all will learn one thing, your six-year-old will learn something else and your dog will learn something else, right? And one of the things we have to capture when we're thinking about intelligence is who's intelligence, right? What are we modeling? What aspects of intelligent learning are we gonna be able to get from that? We will get something, right? We may even get what we want for some AI things, but that doesn't necessarily mean that we're capturing the scope of human behavior because, again, I'm a developmentalist. The whole point about developmental change is that in the same world, you know more and different things than you did when you were 12, than when you were 10, than when you were a baby. It's a very good point. I mean, I think it's understandable why it's tempting for engineers and AI researchers to focus on the environment, because that's an easy thing to study and model and simulate, and simulators are making a lot of progress, but it's not just about the environment, it's about the mind that you bring to that environment, and that's a place where, again, I would make a call for, and this is something I've stood for a lot of my work and a lot of all of our interactions, is the AI people and the cognitive development people like having long-term engagement. I think the field of cognitive science and cognitive development has built techniques to study minds. You can't directly observe them, but you can study them the same way you can't directly observe black holes, but physicists can come up with amazing ways to study them and even now actually observe them, but you have to build a whole theoretical empirical measurement apparatus, and this is a place where AI and cognitive scientists can really have a lot of common purpose long-term. Maybe we should go to some of these questions or feel free to respond to that if you wanted to. Yeah, I just had like one quick question for Laura. I mean, so for example, I mean, in the scenario that you mentioned, that there are many different behaviors you might end up learning, depending on what mind you're bringing in. So I mean, if there is like an AI system which does display many different kinds of behaviors, learning with the same reinforcement, would you then kind of start calling it more like human-like or would you, I mean, when would you start calling it? The devil's in the details, right? So a dog displays many, many different behaviors. None of those behaviors are the same as a two-year-old or the same as a four-year-old, and again, that would be true. We all in some sense live in the same world. We all evolved in the same world, right? So I think there is a real question of why the kinds of representations and structures that we build given exactly the same data often are quite different in our minds, right? How are we thinking, what are the concepts that we can work with? So I do think, I mean, I'm an optimist about AI, and I do think that the point of this cooperation is to try to build models that both do cool things and capture human-like learning. So I think it is possible to artificially engineer systems that can capture some of the richness of human behavior, but one thing we know it can't be is a basic skin area and reinforcement learning system because that can't actually explain this kind of diversity of behavior because lots of animals can do that and lots of animals can't have the conversation we're having. Let's turn to some of these great questions, and maybe since time is a little short, the first two questions are kind of the flip side of each other, so maybe a panelist want to address those, right? So on the one hand is, are we at the point where AI is explainable, or what do we need to be able to do that? And then the other question is, there are limitations of deep learning. I mean, you're saying many of you are using deep learning. I think, well, Polket uses a lot of deep learning, Leslie and I use a little bit of deep learning. Laura mostly makes fun of deep learning. So, but questions of, but what's wrong with deep learning? Well, again, there are many things that our current deep learning systems don't do, but the ability to know what they don't know, for example, that goes, I think that goes hand in hand with being able to explain. Both of these are kinds of metacognition, if you like. So I don't know, what are your takes on either the abilities of current systems or what else would be needed? I'm not sure knowing what you don't know is the same as explaining, right? And I think they're interestingly different, right? And there's a lot of work on now, techno-work on trying to at least make any system that makes predictions also make some kind of commitment about how certain it is about them and so on. So that many things that I've done. But it's more like explaining is like knowing why you know what you think you know. Well, right, so that's different. So that's like nailing down the chain of evidence or something. That's like saying, no, I know this because I saw this thing or I read it in the newspaper. And that's much trickier. For some kinds of reasoning systems, and we do, there was also a question for me, I know the interlocutor too, about planning long-term, can you learn to do long-term plan? Well, but I think it's relevant, right? So long-term planning, right? Consistence learn to plan, so that's a question. I think the answer is absolutely yes. I mean, if AlphaZero does anything in the world, what it does is it learns to plan, right? It starts with a model and it figures out how to plan more efficiently. And in lots of other work in model-based reinforcement learning, then the strategy there is to learn models and then learn how to use them to plan. And a system like that that has the models that it's using available to it can also make explanations, right? You can explain a move in AlphaZero or you can explain, my robot can explain why it decided to pick up a particular object in terms of what it was trying to make true and how and why. So I think- But to some extent, I mean AlphaZero can say, well, I did this because it had the highest expected value. Well, no, but it can trace it a lot further, actually, right? I mean, it can say the long-term dependence is actually quite clear. If you do the search, you'll see it. Yeah, in the sense that you could say, I did this because, but I think it's still gonna come down to the thing. I did this because most of the time, the parts of the game tree that I predict are gonna happen if I do this and you do what I think is recursive- We don't have good- Value iteration. What we don't have is good abstract concepts. Yeah, exactly. We don't have to say, oh, you were in the whatever situation and therefore I did this one. Yeah, that's right. But that would probably be wrong. I mean, to- The explanation would probably be wrong. Well, I mean, it's an interesting, I mean, again, I think human explanation, you can focus on glass half full or half empty. Like, famously, there are things that humans know how to do, which no human knows how to explain. Like, for example, most of intelligence. But there are other things, which like, it's only our ability to explain, explain to each other that allows, humans to learn from school, to accumulate culture and expertise. Like, nobody would be able to do AI if you weren't able to explain the basic principles of how to use TensorFlow or how to use graphical models or how to do computer programming. Like, nobody learned that from RL, right? There's all sorts of great, you know, instruction manuals explaining how to use the tools of AI or any other tools of human expertise. So, you know, I think we shouldn't be, we shouldn't be skeptical of explainability. I think we should just be mindful that, that it's not always the thing to start from. Yeah. Other thoughts on either explainability or how you go beyond, how you build systems that know what they know and know what they don't know or at least are interested in that question. So, one interesting consideration is if you end up having a dog, right? I mean, there's no way you can communicate with a dog and it takes you some time to figure out what the dog might be trying to do. So, there's a period of where you are kind of... You can't communicate, just it's, you can't ask it what it's doing and why. You can't ask the question. But there's this co-evolution period. You're spending time with the system and then you're able to kind of, maybe make some judgment of how the system might behave. I think having some of that component where instead of going from one task completely end to end and then you introduce a human and the human is asking questions to a system, if there's a process where you can kind of grow up together almost, I think in that way, you might end up having more intuition about why the system is doing what it is doing. But again, it will just be intuition. It won't necessarily be the correct explanation of why that is doing. But I guess in many times when we deal with animals or those things, we are fine to extend to some set of problems. Yeah. Yeah. The extent that we understand anything, it's a limited understanding. Question is, Laura. Yeah. I mean, I think that there are lots of processes that we cannot introspect on that govern the kind of behavior we're actively working on. And I can't explain how I just picked up this thing. I can't explain how my visual system works because I'm not a vision cognitive scientist. Somebody can. But even there, we're still working to understand it. What's strange is, as Josh said, there are a large class of things that not only can I explain, but in so doing, I can get the same computation from my head into your head. And all it takes is that ability to communicate symbolically, syntactically in language, and I can create a full representation in your head which you can then take off and use. So it doesn't work for all things. I can't explain to you how to downhill ski and have you downhill ski, although I can maybe give you a few tips, right? But I can probably explain to you how to cook a souffle and have you cook a souffle, right? So part of the question is, for what kinds of intelligence can we do those kinds of things? And how does that interface exactly with the parts of our mind in economics? We have one mind. We do all those things. We combine those. So we combine symbolic, abstract reasoning with things that are not subject to introspection, which is really ground out in all kinds of species in all kinds of ways. And I think it's the interface between those two systems where we're really gonna understand a lot about the future of intelligence. Yeah, I think it's an insight. I mean, you've emphasized this in some of your work, Laura, and others in cognitive science that even if our explanations are often wrong, there seems to be a link between our drive and our capacity, our interest in explanations and our capacity to construct explanations, right or wrong, and our abilities to assess why they might be valuable or satisfying. Again, almost independent of whether they turn out to be true or not. That is part of how we learn. It's part of that mechanism that we don't yet and maybe never will cast into some simple, elegant learning rule, right? So understanding explainability and the capacities of explanation might be essential for systems which learn and are intelligent in more human-like ways. I think we're being told we're out of time, so probably it's a good time to wrap up. Anybody have any other concluding thoughts or? Okay, well, let me thank the panelists and thank all of you for your questions and look forward to engaging more today and the rest of the week. Thanks. Thank you.