Support the show to get full episodes and join the Discord community.
Check out my free video series about what’s missing in AI and Neuroscience
Hugo Spiers runs the Spiers Lab at University College London. In general Hugo is interested in understanding spatial cognition, like navigation, in relation to other processes like planning and goal-related behavior, and how brain areas like the hippocampus and prefrontal cortex coordinate these cognitive functions. So, in this episode, we discuss a range of his research and thoughts around those topics. You may have heard about the studies he’s been involved with for years, regarding London taxi drivers and how their hippocampus changes as a result of their grueling efforts to memorize how to best navigate London. We talk about that, we discuss the concept of a schema, which is roughly an abstracted form of knowledge that helps you know how to behave in different environments. Probably the most common example is that we all have a schema for eating at a restaurant, independent of which restaurant we visit, we know about servers, and menus, and so on. Hugo is interested in spatial schemas, for things like navigating a new city you haven’t visited. Hugo describes his work using reinforcement learning methods to compare how humans and animals solve navigation tasks. And finally we talk about the video game Hugo has been using to collect vast amount of data related to navigation, to answer questions like how our navigation ability changes over our lifetimes, the different factors that seem to matter more for our navigation skills, and so on.
- Spiers Lab.
- Twitter: @hugospiers.
- Related papers
- Predictive maps in rats and humans for spatial navigation.
- From cognitive maps to spatial schemas.
- London taxi drivers: A review of neurocognitive studies and an exploration of how they build their cognitive map of London.
- Explaining World-Wide Variation in Navigation Ability from Millions of People: Citizen Science Project Sea Hero Quest.
Hugo 00:00:03 I came away from this that I’m still very interested in, is this idea of creating algorithms that match the stupidity of humans, <laugh>, or the genius together, that genius and stupidity. So I have a cognitive map of my house. Let’s say I have a map of my friend’s house. These are, these are distinct cognitive maps. The spacial schemer is maps of people’s houses. What do I expect to have when I arrive in a friend, another friend’s house built from those experiences? You ask people, did you grow up in a city or some other situation? Suburb, rural mix. And then we’d simplistically asking with linear models or linear mix models, whether, um, people who grew up in those two different situations navigate differently. Is there a difference? And what we were really shocked when we first looked at the data, I was shocked, was that
Paul 00:00:57 This is Brent inspired. I’m Paul. My guest today is Hugo Spears, who runs the Spears Lab at University College London. In general, Hugo is interested in understanding spatial cognition, like navigation, um, in relation to other processes like planning and goal related behavior, and how, uh, brain areas like the hippocampus and prefrontal cortex, um, help coordinate these cognitive functions. So in this episode, we discuss a range of his research and thoughts around those topics. You may have heard about the studies that he’s been involved with for years regarding London taxi drivers and how their hippocampus changes as a result of their grueling efforts to memorize how to best navigate London. So we talk about that. Um, we discussed the concept of a schema, which is roughly an abstracted form of knowledge that helps you know how to behave, um, in different environments. Probably the most common example used for schema, uh, is that we all have schema for eating at a restaurant, um, independent of which restaurant we visit.
Paul 00:01:59 We know about servers and menus, um, and so on. Um, Hugo is interested in spatial schemas for things like navigating a new city you haven’t, uh, visited before. Hugo also describes his work using reinforcement learning methods to compare how humans and animals solve navigation tasks. And finally, we talk about the video game that he has been using to collect a vast, vast amount of data related to navigation to answer questions like how our navigation ability changes over our lifetimes, um, the different factors that seem to matter more for our navigation skills and many other questions. Learn more about Hugo’s work through the show notes at brand inspired.co/podcast/ 161. And if you value what I do, uh, you can help me keep doing it by supporting the show in various ways that will cost you very little, uh, go to brain inspired.co to learn more about that. All right, here we go.
Paul 00:02:56 You caused a, uh, a little, little friction in my household recently when I told my kids that there is this new game that they could, uh, download and play that they hadn’t played yet. We’re pretty careful with what they can play, you know, and their screen time is like precious. So, uh, any missed opportunities and wasted time doesn’t sit very well. And I blamed you, um, because the, I I, you know, after learning more about what you’re doing, and, um, I ha I was, I came to them and I said, I have this new game, and I th I think you’re really gonna enjoy it cuz they like Minecraft and they, they like, you know, these kinds of things. So I think that they would enjoy navigational games. And of course I downloaded, and of course it didn’t work. It asked me for a code, what’s going on with that? What, what am I talking about, first of all? And <laugh> and why can’t I play it?
Hugo 00:03:44 <laugh> <laugh>. Well, the game I’m guessing you, you’ve downloaded is called See Request <laugh>. Yeah. And the reason you need a code is that we changed over the way that game worked. It was completely open to play in the past, and that meant people could do stuff with it and give us data, but, um, we couldn’t tie it to anything else. It was all totally anonymized. So if I wanted to ask you a couple of questions or invite to meet you or anything, I wanted to link that to, you couldn’t do it. So we, we set it up with a code. Um, but anyone can get a code office. They just need to get in touch. Um, as they, they search, see your request on the web, they’ll find a way to, to get in touch. Um, and we are looking, I should say, later in this year to change that though. So that game will, I’m sure we’ll talk about it at length, but, um, this year we hope to relaunch it. So you do stumble across it. You can play it anonymously and do new things we’ve never done before with it. So, um, yeah. Watch this space for some, some new, new things we’ll be doing.
Paul 00:04:39 Yeah. We’ll come back to that later, but, but you, okay. So is it only researchers that can get access via a code right now? Or anyone can?
Hugo 00:04:46 No, anyone can. Yeah. So there are, there are at least 60 different projects ongoing all using this code system. So those researchers hand out codes to people and they contact them and they get ’em to do, typically they get ’em do other things, like other tests. They might get ’em to do running or fill out some other form questionnaires about their lifestyle, whatever it is those researchers are doing. Uh, and then they just tie it back code. But the neat thing for us is that code allows the game to work away in the background and send the data directly to a server Yeah. In a charity. So there’s no hit download to send the data to researchers. There’s none of that. Hmm. Um, so yeah, it’s working well. Yeah.
Paul 00:05:21 Okay. Well, I just wanted to make sure that I started off our conversation with a complaint and we will come back to see her request, uh, in a bit. Yeah. You seem, you seem busy and I, uh, I mean, well, may, let’s start with, um, your origin story, I suppose, because I’m not sure if you’ve ever had a grant rejected up to this point, but I know you’ve come in just under the deadline, um, or had very little time to write some grants to study what you wanna study. But, um, let’s back up even further. Uh, I know that you did a PhD with, um, Neil Burgess and John O’Keefe, um, and studying navigation, cognitive maps, um, et cetera. Um, were you always interested, did you know that that’s what you wanted to, uh, do from the beginning? Or how did you come into thinking, um, that being interested in navigation and cognitive maps and so on?
Hugo 00:06:13 Yeah, I, I got interested that I, I decided to do a, an undergraduate degree in neuroscience when it wasn’t sexy. Uh, this is back in 1995. I mean, brains becoming a big story. Like the, the, um, the key book that led me into a lot of this was Francis Crick’s book, the Astonishing Hypothesis. Mm. Where says were just a bunch of neurons. Uh, and it’s, it’s an provocative book and, and the you could take issue with the word just by the way. It’s, it’s a good book. Um, and, um, yeah, so I started my degree in neuroscience, so I was enjoying it, but, um, I sort of came across all sorts of bits of research, uh, and stumbled across all the work on memory and thought that was amazing. And then what happened to me was, was probably a lot, a lot of other scientists, I decided to do a, a finding or thesis project.
Hugo 00:06:58 And I, I remember going to see Samuel Zeki about Color Vision, lots of other incredible scientists at that time. Yeah. Um, but I stumbled across the project that was combining virtual reality patients who’d had brain surgery, um, and, uh, walking blindfold in the dark experiments, um, at an ins at the institute psychiatry <laugh>, and jumped on this crazy project. It just sounded amazing at the time. Hmm. And so that’s when I first foray into, into research on SVE Maps and Spatial, um, was with Mike Rache, um, when he was at ucl. He, he left, um, a while back and has been in the States for some time doing other other research. But, um, yeah, that, that was at that point. Then that’s after that undergraduate project, uh, I got a call from Neil Berg to say, come up to my office and I thought, what have I done? What have I done wrong? It was like the head teacher called to their office to, cuz I was on his course, and he like, no, if you’re interested in a PhD, we might. So, so that, that was my kind of route routine. Yeah. So
Paul 00:07:56 The project must have gone well for him to have called you in?
Hugo 00:08:00 Yep. Yeah, it was, uh, he wanted to know all about, he, he was thinking about doing something similar at that point, um, with patients who’d had, so these are patients who’d had temporal lobe, uh, lobectomy, so temporal lobectomies, um, either in the left or right hemisphere. And back in the sort of late nineties, early two thousands, that was a big amazing sort of idea of looking at brain, relating brain structure to function. And, um, yeah. Neil back then had been hacking away when I met him. He’d hacked into, uh, duke Newcomb, uh, and then started to create this entire virtual town that resulted in the science paper and lots of other papers. Um, but the idea om, was to combine that virtual reality, uh, game with testing patients. And that’s pretty much what I did with Johnna Neil, uh, Neil Burgess as my primary supervisor for three years, was putting patients into virtual reality environments. And I kind of not stopped. I still <laugh> Yeah. Still doing that in 2023. Yeah,
Paul 00:08:54 I bet. I think, think Duke Newcomb has improved. That’s, this is a, uh, first person kind of, how would you describe Duke Newcomb? It’s an, it was an early first person, uh, virtual kind of world game, right?
Hugo 00:09:06 Yeah. It, it was, um, and it, I think the thing that, that was interesting on Duke Newcomb was there other, there were like Wolfenstein 3d, there were other games at that time, so it was very soon after they created games you could move in 3D space through on your screen. And they were really novel. But that one kind of had more colors, more textures, more variety and kind of attempted to recreate cinemas, cafes things in the real world that the other games hadn’t done. So it was, it was Neil Burg’s spotting the, oh, well we could turn this into something that would test real life experiences as it was back then. Um, yeah. But yeah, it had a lot of violence, monsters swearing all that Yeah. That are in video games, you know, it was, it was definitely wouldn’t pass, uh, current standards for appropriate.
Paul 00:09:53 Yeah. I wasted a lot of time on Duke Newcomb, and I’m, I’m not a video big video game player, but I remember, uh, playing that one. Um, not with, not in a scanner or anything though. Um, so, so I thought I had heard that, um, you were going along your merry way and then read about the famous place sell and that, um, changed, uh, your trajectory. And I, I mean, I don’t know if I heard that incorrectly, but, but then just separately, uh, so correct me if I’m wrong about that, but then even separately, I wonder how many people this happened to because, you know, the idea of discovering these play cells and the idea of, oh, there are play cells, and that like, changed gripped people and changed their, uh, career paths.
Hugo 00:10:35 Yeah. I think that has happened. For me, it was grid cells. Um, I never, I’ve never recorded grid cells, but, but it was very similar experience of thinking, um, you know, years back. I, I wondered about doing that and thought I’d probably not cut out for animal research. It’s really demanding and thought, oh, I’ll do humans. And then working away in Neil Burgs and Johnny Keefe had a kind of joint lab back then. It was constantly exposed to the play cell recording mm-hmm. <affirmative>. And, and it was a really exciting time. Papers were being published every other week. It seemed major discoveries in that period. Um, but yeah, I certainly, I can, I can describe a number of other people I’ve worked with who, who were doing one thing and then saw the play cells and this changed, oh, dropped everything else went to become studying clay cells.
Hugo 00:11:21 Um, but yeah, the, the story of what happened to me was working away. I was in my third postdoc working with Ellen McGuire, uh, getting to the end of that, uh, thinking of what to do. And I spotted a fellowship, the welcome trust that provided that where it would provide you three years of funding for, for retraining. But you had to, to submit the grant to be successful, you had to show it was the kind of career change where you could, no one would hire you to do this late into your career. Mm-hmm. <affirmative>. So like, you wanna learn how to do genetics, but you had no idea how to do it. It was like, oh, you can, you can come and spend time doing that. I don’t think, think they’d factored in people doing single unit recording, cuz it was, it was like, start doing that and you’ve only got three years similar to a UK PhD.
Hugo 00:12:00 Hard to learn all those techniques when you’re late into your career cuz you’ve got other things to juggle. But the key, the key, like you said though, with that was I discovered the, the, it was there on a Friday and the deadline was Monday <laugh>. And so the main thing is you’re just getting the finance in order to say, I wanna, I need finance for three years sorted out. And it, it got done and I thought it was insane to submit a grant that, that quick. Um, but it was, yeah, whole weekend planning out three years of research and the stuff I wasn’t doing. So I was very grateful. It was Kate, Jeff who is the, the PI that helped support me and do that. I joined her at lab. But yeah, it was a fantastic thing to be able to do.
Paul 00:12:38 Would you recommend, um, waiting until three days before? I mean, you know, this kinda procrastination method that like forces you to do the work, you know? Yeah,
Hugo 00:12:50 No, it’s, there is that, I mean, there’s, if you’re forced to against the time world, then it can focus the mind <laugh>, but no, it’s not, not an advisable strategy. I also had one other grant where this is a, this is a crazy for me, another crazy story in my, where the James McDonald Foundation got on, got in touch with me on April 1st and got an email on April, April 1st April Fool’s Day saying, congratulations, you’ve been invited to apply for $600,000. All you need to do is, and I think I stopped at the point and said, all you need to do is, and then they emailed me a week later saying, we, we emailed you this, we not had a reply. You’ve only got two weeks to the deadline. Are you sure you don’t want to apply <laugh>? Wow. And I, I realized I missed this, this, so I, I, yeah. And I rapidly put that one together. So yeah, two grants where it was <laugh> very close to the deadline. Um, but yeah, I wouldn’t recommend it to, people do read when someone writes you saying, we’ve got 600,000. It, it may be true. You never know
Paul 00:13:45 Unless they’re a prince. Unless they claim to be a prince, perhaps. Yeah,
Hugo 00:13:48 Exactly. <laugh>. Exactly. Yeah.
Paul 00:13:52 So you’re, um, a lot of what you study is navigation. And I know that you love taxi drivers, uh, and you’ve done, uh, um, a lot of the famous work with, um, taxi drivers who have to learn this, uh, the, what is it called? The, the knowledge. Knowledge, yeah,
Hugo 00:14:07 The knowledge, yeah,
Paul 00:14:09 Yeah. Quote unquote, right. Which is like basically how to navigate all of London in the most efficient way possible. And they have to take this super rigorous test before they can become cabbies. Um, maybe you can just, uh, I, I know, I know this is like beating a dead horse for you, but maybe you can just describe the, you know, what, what, what the kind of, the famous findings from those studies. And, and I don’t know where you are because this was like tw almost 20 years ago. Right. Uh, and I don’t know where you are cause I know you’re, you’re doing longitudinal work and studying them as well. So maybe describe those results and then any updates.
Hugo 00:14:41 Yeah, absolutely. That’s been an, a real joy to work, work on that project. So, uh, the key, the key work with London taxi drivers, uh, all goes back to Eleanor McGuire, who’s the fellow of Royal Society at ucl. And I previously worked for three and a bit years with her, uh, as my supervisor. So back in 2000, she and others from the Brain Imaging Lab in London, Phil, this quite the functional imaging lab, um, had, she had decided in that they, this is quite a remarkable group of people. Why don’t we scan, um, and measure the structure in their brain, which was a new approach back in two thousands around that time to measure and use automated systems. But also they’ve like, looked through each of the scans and measured it. And the reason Eleanor decided to do this was that the, in a range of other different animals, like a food caching, birds, squirrels, and some other animals, their hippocampus, uh, changes seasonally when they need it to, to remember things.
Hugo 00:15:33 And the animals that have higher memory demands tend to a larger hippocampus. So there’s reasons across other species to think there’ll be variation in size. And, and as you said, taxi drivers on the way in explaining this, um, all they do their job like nine to five is solving a spatial navigation problem. That’s what they do all day, every day, using their memory. So most if you get Uber or some other, they don’t do that. They just follow a computer system. But London Taxi driver is the licensed ones referred to as cabs. Um, yeah, they, they have this exam to, to get this little badge to work. Uh, and the exam is very simple. It’s just ask a set of like 10 questions. It’s just they’re insanely hard <laugh> Oh. So they’ll say things like, right, take me from, um, take me from Barrack Street to, um, brick Lane.
Hugo 00:16:21 What’s every single street between those two very far away points? And if you make a single error like you, you know, you go, uh, then they, they say, no, sorry, you’ve made a mistake. Come back next time. Um, and so there are about 58,000 streets they could use. So from a computational perspective, it’s like a state space of 58,000 states, of which the exam is only two random ones. And there’s only one solution that they’ve got to retrieve. But it takes about four years. Typically. Some can do it in two years, but it’s often four years of study, constant study of the street names. So the knowledge is really that, that the, the knowledge is street names. It’s not, yeah. They need to know other points. So Eleanor found, I keep the, keep going back to the what. So this is un unusual. What she found was in 2000, that their posterior hippocampus bilaterally was larger than control healthy non-tax drivers.
Hugo 00:17:16 And the longer they had been taxi drivers cross-sectionally, the larger their posterior hippocampus was. And so when I came in to j join this was that when I joined Eleanor’s, uh, lab SS 2003, um, the, the aim of one of her grants was, let’s compel ’em to bus drivers. Because bus drivers do a lot of the same things taxi drivers do. It’s just, they don’t have to remember things. Um, they’re just taking the same route all the time. So it’s things like, you know, they’re, they’re getting visual motion, they’re dealing with the pollution, they’re having to deal with annoying customers. They live in London. All of these things are, are controlled for. And there was, as before, a larger posterior hippocampus in taxi drivers compared to bus drivers now. And the longer they had been taxi drivers, the larger their posterior hippocampus was. And then while I was working at, with Eleanor, we, we used to said to scan them doing their job in, in, uh, an insane study that I wouldn’t recommend anyone <laugh> to.
Hugo 00:18:08 So it was, this is a study we published in 2006 in, in Your Image and in that study, um, and I, I was discussing this with Dick Passingham on email the other day, like, Eleanor, I remember the event she said to me, so how we look at this said, why don’t you ask them what they’re thinking, <laugh>? And I, I said, what <laugh>? She said, just ask them what they’re thinking. And she had been wondering about this for a while about the idea that people might actually be able to tell you what’s going on. And if you ask people to remember a whole day’s experience is insane, of course they can’t, they can’t do that. But what we did with that experiment was to put the taxi drivers into a 4 million pound virtual reality simulation that we got our hold of. Cause it had been built for a video game called The Getaway.
Hugo 00:18:51 Uh, and then for extensive careful work with engineers, rebuilt the Sony PlayStation controller who would pass through a metal detector in an airport undiscovered, but for us, of course, would work inside an MRI scanner. And that was, that was a very hard task. But we then managed to put taxi drivers into a fully detailed simulation of Central London with pedestrians moving traffic, all of it, and hear the voices of people asking them to go to places in a whole ca cast of famous neuroscientists where the voices, these were young scientists then are now heads of institutes. Um, but yeah, no, we, we, after all this pace taken work, after they were endured this, all this driving through London, um, we sat them down with a cup biscuit and a cup of tea and said, here’s the, here’s a video replay. What’s happening? Like, what are, what are they thinking?
Hugo 00:19:41 Yeah. And they can tell you why they made that move. And it was the same with London Taxi drivers. They could tell you every reason why, like, I didn’t stop to let this guy go by to be nice. I couldn’t think about where to go. So I pulled over <laugh> or how I remember seeing Big Ben at that moment, I had eye tracking to verify, are these things they’re saying matching up for eye track <laugh>? 90% of the time they were, it was extremely reliable. Uh, but yet after all that painstaking work, the, the only time the hippocampus was interested, <laugh> was the moment they had to first plan. So like the whole rest of the journey to their destination, there was no increased activity on their hippocampus. Uh, a bit frustratingly, but that, that was the result that, that we uncovered there. Um, and then, and then yeah, that was back in the day.
Hugo 00:20:25 Uh, and in recent years I’ve decided, well, Helen has not done so much more with it. There’s a lot more to ask. Why don’t we go back and ask taxi drivers some more things? So we’ve just finished in the last year an fm I study that’s only focused on that moment. So instead of just taking a small sample of events where they have to plan, we now give them like a hundred planning events like they do in their exam. And we’re scrutinizing the F M R I signal for, for what’s going on and exploiting the, you know, advances in computational models for looking at state spaces. And, you know, it is, they’re, they’re planning over a network of, of spaces and it, it’s great data. And of course, unsurprisingly, we’re measuring the volume of their hippocampus <laugh>, we don’t know yet. We’re still scrutinizing that. So I can’t give you a, does it replicate 23 years later answer yet, but we hope so.
Paul 00:21:11 Are these the same, uh, drivers though, or is, is it long longitudinal in that fashion?
Hugo 00:21:16 No, it’s this, this one is, uh, cross sectional again. But, um, we hope we can hang onto these, these, um, people this time and scan them in another, say four or five years time. Yeah.
Paul 00:21:27 Gotcha. Are, are, are, are these drivers paid? Well,
Hugo 00:21:32 Yeah. Yeah, they are. Um, okay. I mean, you get an taxi and you wanna go seven minutes, um, you’ll have to fork out like, uh, you know, maybe eight pounds. It’s not, they, they d charge quite a bit to take a fully twice as taxi driver more expensive than Uber. Um, no,
Paul 00:21:47 You’ve quite, now you’ve question. I know, but wait, so you, you were saying that you were asking them, what was the point of asking them what, what they were thinking? First of all? I mean, while you were, uh, saying that I, I was thinking my wife is like constantly asking me about my feelings and I have like no access to that. I can’t describe what I’m feeling and thinking in any given moment. So, uh, so these taxi drivers, although it sounds like what they, what they were describing was more like the reasoning behind their actions kind of.
Hugo 00:22:16 It it was, it was a bit, yeah. I mean it, it, it was, it was one of those scenarios where if you came back into your house and found like a dog had gone crazy and uh, destroyed and then there’s a vase smashed everywhere and then you look like a whole sequence of events happen to you like that. And there’s a video replay <laugh>, and you see your face going, ah, like you could walk through what you remember and you might mm-hmm. Have they did, they had whole sections where they say, I have no memory of this. And, and they would say, and it would be, I’m just driving down a straight road, there’s nothing to see. Yeah, I might have been thinking about tomorrow, whatever, mind wondering, but I don’t know what I was thinking. And then they’d bits, they’d turn into a street and they’d say, oh my goodness.
Hugo 00:22:57 And it would look boring to me, but they would say, what’s happened to that shop? <laugh>, it’s gone. There should be a major, you know, supermarket there or they, they could anything that was sort of strange. Um, and, and we, we simulated one of the worst days you could have in London. So we didn’t wanna put ’em through an easy journey. We constantly gave them blocked roads, customers change in their mind, lots of things happening to them. So it created this flow of events and experiences that they were really able to say, that’s when I was thinking about this. Or, or that. And certainly when they were replanning their route, we would see lots of prefrontal activity areas. You, we, we know from lots of neuropsychological evidence and others that you would expect to see anterior frontal lobe activity when you ha have to replan. Um, but just no hippocampus.
Paul 00:23:48 So, okay, so their hippocampus like lit their posterior, your hippocampus lit up during the, during the, um, before the journey set off. And then it was not active while the journey was happening. But when they had to retool, then it was the frontal cortex that um, had a bolt, higher bolt signal. Exactly.
Hugo 00:24:04 And if we were to do it again, I’d have measures of what exactly how difficult the replanning is. Cuz it’s quite possible that maybe if it had been really difficult then, then we might have seen the hippocampus involved. That would be my expectation. Now, back in 2006, we just had replanning or not replanning, and I think we missed out on, but generally if it’s there, it’s only there for some replanning, but not a lot of it. That’s, that was my conclusion by then.
Paul 00:24:30 This kind of flies in the face of the classic, um, navigational studies and rodents, right. Where there the play cell happens when, well, in this analog, the play cell would happen when you turned on fifth Street or something, right. That the play cell would fire. Right. And then other play cells would fire in different locations and stuff. But is it, you know, you’re recording bold signal, is it that mm-hmm. <affirmative>, it’s just not sensitive enough to pick up the activity or, you know, could the, could there be that reason?
Hugo 00:24:57 Yeah, it’s a good question. I mean, we have to be, it’s very coarse, right? The bold signal is an indirect measure of the, the neural activity, that’s for sure. And um, here, you know, when, when you’re thinking about how the play cells would operate typically in a familiar space, pretty much if you looked at rats running through a maze or a box and one of the most impressive features, and there’s a very nice paper by ca Barry’s group on the monotonic, like it just maintains the same overall rate of excitation around the entire environment. There’s something incredibly beautiful about this homeostatic system that doesn’t really go up and down near the edges or the middle. It does something very clever. And so taxi drivers driving through London, you wouldn’t actually expect any moment to moment changes in the overall bolt signal if it’s that. Yeah.
Hugo 00:25:41 What you would expect is that if the taxi drivers hit a dead end, they weren’t expecting and have to now recompute what they’re doing, you, you might well expect shortwave ripples and to enter into some sort of offline, uh, a non-task focus, like, sorry, a non stimulus oriented into your own thoughts and planning state, and that you get a lot of shortwave ripples and if you’re getting shortwave ripples, you’d expect a lot more bold signal, uh, in hippocampus. So that’s for me, the theoretical link between, we think the hippocampus is involved in doing some memory-based planning through reactivating the maps, uh, recalling the information in those moments. Um, and so I think that’s why we saw it at the beginning of the journey, because you have to recall, you’re given a street name and you have to retrieve where is that? How am I going to get there? Um, but for the rest of this, we just don’t, that, that’s the link in my mind between what’s going on between the neural. But it’s speculation. We do need more data. We still need it. 2023. Yeah.
Paul 00:26:40 Uh, <laugh> rats. Rats are, are humans. Yeah. I, you know, I was gonna ask you this question, like, which, which is more difficult, right? Recording and, and planning, um, experiments and so on. But you already alluded to the fact that you did this insane study with humans, um, and that, that you wouldn’t recommend. So now I’m get, I was gonna guess you’re gonna say rats were harder to work with, but now I’m gonna, uh, guess and say humans.
Hugo 00:27:04 Uh, it all depends what you want the get ’em to do. That is the question that is the point. What do you want the rats of the humans to do? If you want rats to, to run around in a box, then they’re great. You, you’ve got them there. You don’t have to invite them in, they’re in a research laboratory <laugh>. Um, but yeah, so the behavior, if you’re getting something simple is, is is fine getting rats to do complex planning tasks or something more like humans. That is really hard. And that, we had a paper last year in current biology where it took a lot like five years to work up a task that rats really have to think about where they’re gonna go. That would also be hard for humans. The humans would struggle similarly. So we could try and match rats and humans and the, the rat side’s much harder than the human side. Um, dumbing down the human side is, is uh, making harder for humans is like, uh, the other flip side to that. So yeah. But overall rodent research is about four or five times harder than the human work, I would say <laugh>.
Paul 00:28:01 And so you ha you’ve never, you haven’t worked with non-human primates. I know this is, this is an aside, it’s probably only of interest to me, but I’m always like questioning cuz I worked with non-human primates and, and I’m always, you know, questioning like, what would it been like to be happy, you know, <laugh>. So
Hugo 00:28:16 Yeah, no, I haven’t, but I can imagine that that that like three or four or five times harder becomes even harder with non-human primates. Um, depends what you’re trying to do. Of course. It’s always the issue. Yeah. Yeah,
Paul 00:28:27 Yeah, yeah. True. So I mean, let, let’s jump into this. I mean, you, you developed this task and, and you already talked about, uh, alluded to the paper that, um, came out last year. Um, and this is specifically to test navigation, um, skills and the neural underpinnings thereof, um, in a setting where rats, um, and humans are doing the same tasks. So why was that important to do? And then just, you know, tell us about the, the wonderful successor representation reinforcement learning algorithm that we’ve talked about a little bit on this show before, but, uh, and, and, but why, you know, what was the question that you were asking and why was it important to compare rats and humans?
Hugo 00:29:07 Yeah, uh, no, that’s a great question. Uh, it seemed to me in the field we’ve got lots of, uh, research, like review articles and papers that are trying to combine rodent and human work together and say, oh look, these tell us about the same sort of things. I just did that with the taxi drivers. I said, I think it’s because we see this in rats, it’s probably what’s happening in humans. But these are, you know, it’s, it’s always a bit of a marrying up very different scenarios and situations. So we took the view of what if we were to actually directly linked this so we have rats running the same kind of trajectories that humans would actually run. Um, would you see similar patterns in their behavior? That was the, that was the core question. But the, but the, the key with that then for me was, um, if, you know, we often also think about rats and humans as having these different abilities, like humans have a large prefrontal cortex.
Hugo 00:29:55 And again, I’m just trying to link together these two species, like how different are there, how different is their, their way of navigating or are they actually quite similar? So one of the core questions in that experiment was, um, how are they very similar or very different when we’ve tried to equate, is it always gonna be different cuz you could never quite match it perfectly, but when you really work hard, can you get that? And the answer is yes, there are very similar rats and humans in the way they, the way they navigate mazes, um, like this. Um, but the other, the other reason to set up this experiment really for me was, um, that then we could test the, their behavior against models in the maze. We designed it this experiment to make a maze that would be useful for testing against reinforcement learning, learning models.
Hugo 00:30:37 So again, in the reinforcement learning field, you talked about a lot on this show. There’s fantastic examples of like forum mazes where you can look at the how, you know, applying, applying some hierarchical representation of sub goals and things in the models really helps involve improve the match to behavior and, and things like that that, that people have worked on in reinforcement learning. But, um, you know, I’d known from working in rats for the, you know, from start going back to that training fellowship of just how hard it is to get rats to do something interesting behaviorally that you can match onto these reinforcement learning. It’s amazingly, you know, rats can screw things up in the most amazing, elegant ways you would believe <laugh>, they’ll make it hard to run experiments. Like they just get interested in other things. And the key thing you’ve gotta remember is that they’ve got a, for a coat, they’ve gotta keep pristine and they need to go to the toilet.
Hugo 00:31:25 They’ve got all sorts of other needs are like they prey, they don’t wanna be eaten. So there are all sorts of things going on that aren’t happening for undergraduate students and experiments. Um, and they, they, they’re quick, they run fast, they like to move quickly. So all these things are going on. But the, the idea is that if we could build this experiment where you would see comparable, um, behavior where rats are not succeeding all the time and humans are not succeeding all the time, they’re failing, they’re not failing so badly, they’re lost. So somewhere in between where you’ve got lots of errors, the purpose, you learn nothing. So took a while to get, and the key innovation in that experiment was having a maze where the rats have to plan and go around obstacles to get to target locations to get to the goal.
Hugo 00:32:07 Um, but the, when the, when the maze changes, when you move those obstacles and those barriers visually, there’s not a big change. And that’s cuz the, if you have a big visual change, rats will get confused by where they are and spend time resent marking cuz they like to pee on things and bite at stuff. So we had these canyons, so it was a maze full of canyons. They run around on this surface trying to get to the target and we could reconfigure the maze layout to, to make that happen. And other groups have used similar approaches. Uh, John O’Keefe’s group, for example, have this, um, hex, um, honeycomb maze where things move up and down and ours was much cheaper. We just moved blocks between trials. But it was, um, I was very pleased by how it worked with the rats. But the, the other key reason for why you’d want to put rats and humans through the same tests is that yes, you can, you can apply them against the similar, um, reinforcement learning algorithms set, but down the line, as we’ve done since, you can then apply both the F M R I and the rodent single unit recording or other rodent methods and then sync these up into the ch same trajectory space and the same computational model.
Hugo 00:33:12 It’s seem like quite a powerful idea to have these cross species methods. Um, then you can exploit human experiments and rodent experiments in one kind of paradigm. So we’re, we’re still working on that. That’s, that’s hard to do as well. Um, but yeah, we’re good. Going back to what you said coming in, we were looking at the, the rats and humans from the perspective of humans or thinkers, we plan, we mull over what we’re gonna do. So you might expect humans to fit more with let’s say a model-based planning system. Um, you know, where you, where you simulate possible trajectories and you choose from a set and this wonderful literature in the reinforcement learning field, people will be familiar with and model-based, uh, reinforcement learning. And you might think rats having watched them, I have, one of my expectations is that really habit, they like to build a habit quickly and run <laugh>.
Hugo 00:34:00 And so it’s quite possible that a model three system that just really emphasizes the optimal actions the rats are taking might capture rats. And then as you discussed, there’s this, um, successor representation, uh, model, um, reinforcement learning, uh, model where, which was devised by Peter Diane in 1993 that kind of sat there for what happened and became a big thing in the field. But the idea there was that this might be, and Kim Feld had, uh, her paper with, um, several prestigious co-authors, um, in in nature neuroscience arguing that the at THEC campus has Susan a successor representation from looking at the cells. So again, we wanted to look at that with the data and uh, Kim Stack Andel was probably very happy when we’ve found that, yeah, the rats data and the human data in the way they move, the errors they make best fit the kind of errors that a successor representation makes.
Hugo 00:34:53 And it’s certainly a, it has a very sensible system to do this flexible planning, to not have to simulate every single time you move where you might want to go, but to actually just cash, uh, a structure of the environment. Um, and so yeah, it was, uh, but the person who deserves a lot of credit in that project is Will Dhi, who’s the first author, who is a PhD student in my lab, who then is now a postdoc in Kawell Barry’s lab. Uh, and he really threw his heart and soul into all the modeling, helping collect the own data, helping with the human’s data. It’s a hell of a PhD.
Paul 00:35:28 Hmm. What do you think about these different algorithms for, well, for example, we’ll stick with reinforcement learning algorithms, right? So in, in like the study you tested three different very clean algorithms and then, you know mm-hmm. <affirmative> the, um, successor matched the behavior the best. But as we know, uh, let’s say we assume that the brain does this cleanly as well. It has clean systems for, you know, separate systems for successor representation type for model-based reinforcement learning and for model free reinforcement learning. Yeah. But the reality is, you know, we don’t actually know how, you know, separate these systems are in the brain or really if these clean algorithms map onto anything in the brain, you know, ma have this like corresponding algorithmic mapping onto processes in the brain. And, um, if they do, they’re all interacting and they’re all would, would it be fair to say that they’re all probably going on at the same time? Um, and then yeah, vying for control or being controlled, which one’s being used and so on. How, how do we, how do you think about that in terms of like what the results that you got out of that study?
Hugo 00:36:33 No, that’s a, that’s an excellent perspective. I’ve given the kind of very basic perspective on that experiment, but you’re absolutely right that, um, it does, it doesn’t make sense not to do that. So, um, just like in everyday life, if you had to think about it strategically from a model, everything you would do, I need to get the milk from my fridge that I do every ti day. And I think it, it’s just, you would, you would burn out from exhaustion and you, you need a lot of glucose for your brain to just do that. So we have to have these, these habitual systems. And there are other times where yeah, if you, you, you can’t, your roots blocked you, there’s a block road blockage or what in the real navigating, you can’t use your habit system, you’re gonna have to do something else. So, so you’re right, in the real world there are all these scenarios where it doesn’t, it doesn’t make sense that there’s one system they’ve got to, there’s a lot of evidence there’ll be these are competing cooperative, potentially systems for solving problems.
Hugo 00:37:26 And I think maybe one of the, one of the ways of looking at our experiment is to say, and this is a challenge of doing the work I do on navigation <laugh>, is we set up, we set up the experiment test a particular scenario that we thought was interesting. And it probably was oriented towards success with the successor representation approach in that you had to do flexible planning, but it wasn’t pushing you. Like, my God, you have to rethink everything all the time. So I think that is the context there. Um, and that is the challenge of publishing research papers is you try and say, we do this work and we think this is a, we see this and it makes sense that the successor representation can explain this pattern phenomenon. It may be used to guide, it may be part of the way we navigate, but you change the paradigm and do something else, you would find very different results. I think there is no question in my mind. Um, so, but that’s, that’s an empirical, uh, in a question, something we need to look at. So I’m not claiming that the, it turns out the way we definitely navigate is a successor representation. That would be a very false perspective. And I’m pretty sure we didn’t say that in the paper. We had quite a long, long discussion in there. We tried to partake and eat it, as I said to a, someone that wrote to me afterwards. Yeah,
Paul 00:38:37 I mean, do you think though that um, something like the successor representation would translate to non navigational domains in like in the same proportion or, you know, if, if you made a study and it kind of lended itself toward, or, or was biased toward a successor, successor representation kind of, uh, strategy, uh, with, you know, with the relative different frequencies, let’s say, of these three separate algorithms. I’m sorry I keep rolling my eyes about these things, but um, you know, what kind of proportions are we using and in our day-to-day lives?
Hugo 00:39:11 Yeah. Uh, and yeah, I mean that comes back to a nasty question again, which is for each person what is their daily life? Because some people’s lives are really unpredictable. And if we take you and me, it’s probably a lot of the <laugh>, a lot of model free <laugh>, thank goodness. Cause again, it’s like a super, you know, um, so there are beautiful experiments of people taking these, these models and we, you know, when we were looking at our experiment, we weren’t at all the first to try this by any means. Nathaniel DOR had been doing a wonderful job of comparing you’ve had him on your show, you know, talking about his wonderful work comparing algorithms to people’s behavior. So mm-hmm. <affirmative>, yeah, this experiment was where, as far as I remember, he hadn’t, he hadn’t done rats and humans and we wanted to combine these two literatures together.
Hugo 00:39:54 Um, but, and I should add, I mean like, it’s not in the paper cuz it, it’s a very long paper as it is. But when we were trying to create the experiment, it was very hard to get rats to show, uh, successful navigation behavior in these mazes. Cause I was changing the maze around and they would just keep going down the dead ends or hang around near the goal, but not getting to it. So they looked a lot in the early experiments, like the model free algorithms. Oh, because we made the mazes so hard that you probably need a lot of model-based planning to solve them and rats just don’t naturally, you know, do that. Potentially humans might, but rats certainly weren’t. Yeah.
Paul 00:40:30 So you, you, yeah. You said it took like five years to, um, to get all this working. So was that just titrating for the rats behavior or did you have to worry about humans, like making it too easy for humans or anything?
Hugo 00:40:41 Both. Both. The rats are much harder to get that to and it’s all, you know, bits, projects do just take with different people coming and going and, and to building equipment. Yeah, I hadn’t a friend who’s an architect designed the maze getting it all. So all those things just take time. So it’s not like, um, we were just held to, you know, a fire, you know, team of people for five years solid. But yeah, the rats, the rat just getting a, a scenario where the rats would run and show these wonderful flexible route where, uh, for me, the data’s in that study is beautiful. Where I see rats start a journey. They go into the may subtly find that they can’t walk towards the goal. They’ve been doing that again, again, it’s blocked. And then use their knowledge to run right round the maze and find their way into the goal where they’d never done it before.
Hugo 00:41:24 So it, it fit these kind of narrative. Rats don’t tend to do that normally, but I think we, we built up a scenario where they would, they could, they could learn how to do that through the, the environment we picked for the humans. It was really hard to, we had to basically build a lot of fog, you the humans in a foggy environment because their vision were got fantastic binocular vision to spot changes in a maze far away. But once you’ve got the fog in there, humans become quite dumb. Like rats. They just don’t, trying to remember where all these edges and the MAs are and they, they show the same sort of error patterns.
Paul 00:41:56 <laugh>, dumb like rats. I thought the story was that we’re supposed to appreciate, uh, animal intelligence a lot more than we do, but, but you’re saying perhaps <laugh>. Yeah, perhaps that’s not the case. <laugh>,
Hugo 00:42:07 Yeah, maybe I’m a bit overly like dumb, like rats. Like we, we were trying to get rats to do human things, whereas if the rats have built Macs for us, we’d look like unbelievably dumb. We couldn’t, we couldn’t believe we couldn’t smell the cheese at a distance, you know,
Paul 00:42:18 So Yeah.
Hugo 00:42:19 Right. We do have to <laugh>.
Paul 00:42:22 So, uh, before we move on, just what, what’s next on that project? I mean, do you have, does this tell you anything about the way an AI agent should be built or how you know how to build intelligent agents? Or is it more a story about, um, the nature of our own navigational abilities?
Hugo 00:42:42 It tells us that last part partly about how we, we what what sort of algorithms might explain the way humans would solve a maze in a scenario like this. So that’s one framing. But your question about AI I think is a good one of, um, what does this tell us about AI and, and developing those? And you did use the, a really nice phrasing in, you say you took three very clean for that exact reason. Very, the simplistically creative models we could with no bells and whistles. We really wanted to have very simple algorithms, see what they did. Um, the takeaway i I came away from this that I’m still very interested in is this idea of creating algorithms that match the stupidity of humans <laugh> or the genius together that genius and stupidity. Um, a lot of the work, if you look at the beautiful work at a deep mind that I’ve been following, cause I know that I had my, my, uh, period in Ellen McGuire’s lab overlap with Demis Deis ais.
Hugo 00:43:35 So I, I saw the rise of Deep mind and a lot of that’s incredible showing these like better than the best world chess expert, better than the best go player now. So we can show AI can do things at a super high level. Um, what I’m quite intrigued by is the idea of building AI is that can simulate dumb people <laugh>. Um, because there’s huge advantages to, to predicting people being dumb. Uh, and I’m sure we’ll see more of that work going on, but it’s not something I see a lot of, I dunno if you’ve had guests on your show talking about that, but I don’t tend to see a lot of discussion about building AI to capture stupid behavior and then what do we do to rule this out or, or solve it or, so I think that’s, that’s a really nice thing. I I think we’ll see out of, out of some of the, and this project did that, it helps spot, you know, it shows that humans aren’t, like in this scenario, fantastic model based planners. They make errors that are not consistent with someone constantly planning. It’s not what we do. Uh, although some humans in the experiment definitely were their data matched the model-based plan. So I think, I think that’s one of my perspectives on this relation to ai and what, what’s happening in reinforcement learning
Paul 00:44:41 A related topic is ecological validity, right? So there, there’s this craze right now that, um, experiments need to be more ecologically valid and in a sense you strip that away from both the humans and the mice because they have very different affordances in the world. Their, um, ves are very different. So, uh, an argument could be made, and I’m not trying to be, you know, critical, I, I’m just, you know, wondering about your thoughts on this that, you know, by, by putting rats in an taking them out of their ecological environment, taking humans out of their ecological environment, what is really to be gained, um, because maybe what you should be doing is putting them both even more so in their ecological environment, but then it’s not a controlled nice experiment where they’re doing the same thing, I suppose.
Hugo 00:45:25 No. So, so I think the, so that’s a lovely question. I think that’s great. Uh, and I think the answer is that you need, you need a whole range of data, right? So it’s great to get fully ecological data on humans hand rats when you’re looking at what they do. So this was a really nice study, uh, 2021 looking at using runkeeper, this app for running whether a team in m i t got hold of the data very carefully through lots of discussions and we’re able to show algorithms for how humans making navigational choices in the wild in their running in like New York. So that’s one way of going away from what we’re doing that shows you how the behavior in the lab relates to the, the real world. Um, but another way is, and that’s just wonderful studies of rats in the real world as well.
Hugo 00:46:05 Um, but then, then, yeah, what’s the, what’s the rationale of bringing this in? And it’s exactly that, that you can’t quite so easily get rats running out in the New York subway with optogenetics. People are gonna scream, it’s not appropriate, not ethical to do that <laugh> uh, suddenly, you know, recording from people, you can’t move MRI scanners currently around, there’s movement to do more mobile brain imaging, but there’s just methods that mean can’t do this and what can we do. Uh, so I think, you know, going back to that, if we had all the best kit and abilities in the world, it would all be out there. But we don’t <laugh> we need control. We need some lab scenarios and we don’t have the time to collect a thousand years of data to, so, so I do, I I I would also say that experiment we ran in the lab, we have humans play video game, a video game that’s very much like what millions of people are doing right now. They’re playing video games like we did, and the rats are maze experiment, I think is more ecologically valid than most of the box or plus mazes or, you know, thing. It’s, it’s much more like the kind of scenario you’d find in burrows, uh, the rats would crawl through or, you know, it’s, it’s got that kind of, um, irregular feature layout. So that’s my, that’s my pitch <laugh>.
Paul 00:47:15 I want to come back. Okay. So I know you’re doing work, um, recently on schema and, uh, I realized I don’t really know what a schema is, but so, so we’re gonna talk about that. Um, and I, well, I’ll bring in the, uh, uh, taxi drivers later in the, in this little conversa this part of the, the discussion I suppose. Um, but in this paper you distinguish between a spatial schema, cognitive maps, and like event schema. What, what is a schema? What, and do I have it right that are, are schema making, like a comeback in the neurosciences or in psychological sciences in general, or have they always been there? It, it just, I see more and more about schema these days.
Hugo 00:47:57 Yeah, I mean they had to, they, they were first broad, so, so defining a schema, um, you know, off the top of my head, uh, I’m sure there’s a very nice, the definition out there on Schema and it’s in our recent Nature Reviews neuroscience article, I’m sure of that. Um, so, so in terms of working on schemers, uh, this is the, the recent paper review I wrote with, uh, Shane Rosenbaum, uh, and, uh, Morris Moscovitch and Andela as the first author. She’s fantastic. So, so the team wrote this familiar, this, this review on how can we think about distinctions in cognitive maps and schemers. And you’re right in the sense that schemers, um, were, were, as far as I’m aware, going back to Bartlett in the 1930s, psychologists thinking about how you can remember events and things, but actually you abstract over knowledge. So not that you learn, like you can learn the Paris as the capital of Frances semantic fact, but, but gathering a knowledge base about what’s likely to be the case or likely to happen in scenarios.
Hugo 00:48:56 Um, or if I go into a restaurant, what do I expect to be in restaurants or happen in restaurants? So an event-based schemer is things like, what do I expect to happen at birthday parties? So it’s not like a semantic fact. Your brain retrieving goes, oh, the answer is this. It’s a set of predictions, really. So a part of the idea of schemer is predictions instead of a set of representations, representation of the brain. Of course, the whole, you heard a whole podcast on what, what’s, what representations are. But this idea of something that allowss you to map on, have expectations. So the key point in that article we wrote about cognitive maps and spatial schemers, is that a, there’s been a lot of talk about event schemers. What happens if, what do you expect to happen in a birthday party and less so on?
Hugo 00:49:38 What do you expect in a, in a, in a spatial environment, in a city, for example? Um, although there’s been a lot of work on rodents and spatial schemers, um, and the distinction between the cognitive map and the spatial scheme we draw there is that, um, a cognitive map as described by Johnna Lin Nadel and their, their key book and building on Edward Tolman’s famous, um, work on where he proposed cognitive maps is the idea of an environment. You, you go into a new environment and you build a internal representation. A map doesn’t mean it’s a vertical map. So people get really fixated on this. It’s how, what extent is the map, but it is a system of representations of distances and angles and, and information about the spatial structure. But it is all vast environment. So I have a cognitive map of my house.
Hugo 00:50:24 Let’s say I have a cognitive map of my friend’s house. House. These are, these are distinct cognitive maps. A spatial schema is maps of people’s houses. What do I expect to have when I arrive in a friend, another friend’s house built from those experiences? And what we’ve pointed out in this article is a surprisingly little note about this where, where <laugh>, where, where is the neural system for this spatial schemer really represented. Uh, and this is a, an article in, in nature of these neuroscience where the first figure kind of outlines where this may be in the brain. And it’s a guess. It’s, uh, really be honest and say the data is not sufficient to really be confident about where, where that might be. So with events schema, so what do I expect at a birthday party? It seems like the ventral medial prefrontal cortex in humans plays an important role in providing that, that information through your imaging data and patients.
Hugo 00:51:18 But for spatial, we don’t have as much information. So, so that’s really the key distinction to draw in your mind, um, between a map of an environment you have learned versus a, a gathering of lots of environments that becomes a schema. And, and we just linked that into lots of, lots of ways of thinking about urban analytics. There’s been a lot of wonderful work on how we could represent cities and how cities are structured. Um, so it’s, it’s a wide ranging review trying to talk, talk to that, but, but it’s something, yeah, we, we still don’t know much about. I’d love to know, uh, I mean, extending this to taxi drivers, there was a case when I was working with Ellen McGuire of a, an amnesic taxi driver. It’s like a Venn diagram of amazing memory loses memory. Let’s study this one person who every hundred years may come forward and he could navigate a lot of London really well. He could tell you an enormous amount of spatial information. He retained a lot of semantic knowledge about London, and he had a lot of what the, like, schematic ways of thinking about the, the space. It was just when he had to navigate really intricate roots, he couldn’t do it. And you definitely wouldn’t have want paid him money to take you somewhere, um, in London.
Paul 00:52:28 Well, he was he still taxing?
Hugo 00:52:31 No. <laugh> he wasn’t. So he, he was someone you’d, you’d, uh, you know, shake his hand and say, hello, go and get a cup of tea, quickly bring it back. And he would’ve forgotten who you were and introduced himself again. Um, and when he were navigating, so he used the
Paul 00:52:42 Classic h kind of
Hugo 00:52:44 Yes, exactly. Likem. He had severe bilateral hippocampal damage. Um, and he was densely amnesic, but able to navigate London. A lot of it, but not perfectly. Um, so yeah, I think, yeah, that’s the story there from my side.
Paul 00:53:01 Yeah. So I’m asking this because, uh, after I read your paper on spatial Schemas, um, I read another review, or I think it was a book chapter on schema’s, um, written, co-written by Alison Preston. Anyway, she, uh, like at some point in there there’s a sentence that says something to the nature of a blah blah, blah, a cognitive map, which can be considered a spatial schema. Mm-hmm. <affirmative>. So in that sentence, she was like, equating spatial schema, schema and cognitive maps. What you’re saying is a cognitive map is a specific, um, uh, navigational, um, map, right to a spec, to a specific environment. And schema is, are is an abstraction over many different kinds of environments. Like in the literature, I don’t know what your sense is, but is there agreement on what is schema is do like I have trouble conceiving, like it’s so abstract? Or is it, is it abstractions or is it an abstraction? And like I I have trouble visualizing like what is schema actually is and how be represented.
Hugo 00:54:02 No, I think, I think you’ve got it right. Your confusion is, is the way you should be if you’re reading about this, it’s that, the reason that’s right is that it is confusing. There isn’t to across people the way they’ve thought about it and described it gives rise to, cuz it’s not precise. It’s not a, there’s not a computational relationship between these things. We can’t say, this is what you would define for sure as a schema. Uh, and, and in our review, we try to pick up on the fact that we draw examples like, um, you know, uh, in cities I expect the big wide streets to be important streets. And I can tell you a lot about what I would, and you can have an awareness and you can describe it, but at kind of lower level, your, your brain is just drawing on information to make predictions about where you should go in a space.
Hugo 00:54:47 Dirt is building on those, those learned representations of, of the environment. Um, so another way of looking at it is that if you were, you know, ideally you’d build a perfect home to map, like you wouldn’t make an error ever going in a new place. You’d go on holiday to The Bahamas. You don’t make a single mistake. Whether you think about the space is perfect, but we’re not like that. We do make errors. And part of that might welcome come because you have a schema of what you expect in The Bahamas and how the streets should be laid out. Uh, and it’s wrong. Like you think everything should be 90 degrees, all the turns, everything’s recline and it turns out <laugh> it isn’t and you get caught. So I think that’s one of the ways we were thinking about the way in which your prior knowledge, these representations, other people might say, who goes wrong? I don’t see them like that. That’s not how I think about schemas. And I’d say <laugh>, that’s where we are. It isn’t a term that everyone would think agree on. And Ali Preston’s comment exactly fits that where, where, um, in our review we didn’t, we didn’t, uh, we didn’t agree with that. Um, but I, I wouldn’t say there was some huge disagreement with us. It’s a, I can see where she’s coming from.
Paul 00:55:54 One of the reasons why, um, I wanted to ask about schema is and, and bring back the, uh, the idea of the taxi drivers and the increasing posterior hippocampus, um, is cuz we think of the brain as hierarchically organized right? From like lower level kind of details to more, uh, abstract concepts to ches now, right? And <laugh>, um, and in the paper, and I’m gonna throw the word gist in here also, and ask you how, what, what’s different, what is a gist relative to a schema and then how those map onto, um, different parts of the hippocampus. And then, um, you just said schema are located in, and we’ll use this term loosely, um, ventromedial prefrontal cortex. Um, and you can kind of, I don’t know, can you think of that as just higher in the higher structural hierarchy of the, uh, brain from posterior hippocampus to anterior hippocampus, and then that’s the end of the ability of hippocampus and it needs to offload then on prefrontal cortex. What is that the idea there?
Hugo 00:56:51 Yeah. Um, yeah. Um, there’s a, there’s a lot going on <laugh>. There’s, there’s certainly lot. The brain has to have hierarchical structure. There’s you just anatomically it has <laugh>. It’s, it’s not, it’s not a, that’s not a, you know, a hard, hard thing to, to argue. Um, and I think the, the spacial gist is a bit of a, that was something we suggested in the article that you could go to full schema of what do I expect in French cities, right? I go to France, there’s particular things to French cities. It’s quite a high level representation of my expectations that can go from right the top, there’s a center, there’s a cathedral right down to how the streets are. That’s different to other countries. Um, whereas a spatial gist might be something smaller. We argued that just fragments of that where you’re capturing, um, commonalities over experiences, say possibly even in a single environment about the relationships between things.
Hugo 00:57:46 Um, and the confusion here is these things are a bit like they, they fall between that episodic memory and semantic memory where you’ve got one shot, you’ve experienced this thing, you’ve encoded the way the part benches were laid out, or you’ve got a lifetime of knowledge about how part benches are always laid out. And a spatial gist might be somewhere between, oh, in this particular scenario they tend to be doing this. So it’s gathering that kind of a knowledge that’s not based on a single episodic event. Um, that might be the spatial gist. Then going into the brain anatomy, I think it’s important. There’s a lot of debate as you’re seen recently about localist, the interior mecampus is doing dysfunction of, you’ve seen this debate. Yeah. And I fall in the, the camp of where I think people are talking across as, so for me, there is undoubtedly localization function.
Hugo 00:58:31 Oh yeah, there ha there has to be, um, because of the low, just the lower level structures about how, how we breathe. There are dedicated neurons localized in your brain for allowing me to breathe. What for me right now, they’re not distributed. They are set there <laugh>. It’s much complex, I hope not complex for the neocortex as to how that’s localized. Um, but I think the really key thing is to think about these as the hippocampus does this. Uh, and I think it’s, it’s certainly the case. The neurons are transforming information through the network. And the key thing that keeps going up again in this context, right? So under this, we just talked about it in my mazes, is that’s a particular context and experience. So when we say hippocampus is doing this, it seems like the evidence is the posterior hippocampus is more engaged in fine grain detail. There’s a very nice review by Lynn Nadel and Morris Moscovitch and others, and Jordan Ang, um, looking at that. Um, so, so we in our review are building on that idea of, of more posterior detailed areas to more broad towards the anterior parts. And there’s also some evidence in the prefrontal cortex of this.
Hugo 00:59:40 So I think that’s where that comes from. But it’s important not to think in my mind that it is a localized, oh, it gets passed and then it gets offloaded and it’s much more of a, you’re in a particular situation and these circuits will operate and use the word circuits as in pathways of, of, um, axons game between neurons. And there’s huge, uh, feedback and dialogue and um, and a lot of processing going on. Um, but yeah, getting to the bottom of it, but it doesn’t mean you can’t look and say, where is this tending to happen? You know, where is this localized to, I think is a reasonable que it’s just how you draw inferences around it. But sorry to go off onto a tangent about the recent, um, no, yeah, I keep coming up every so often. Yeah.
Paul 01:00:21 I mean I, I think, what is it, 99% of, uh, scientific disagreements are people talking past each other. It’s gotta be close to that, right?
Hugo 01:00:30 Yeah, yeah. Absolutely. So yeah,
Paul 01:00:33 I’m sorry for my naivete about this, but so, so you brought up like semantic versus episodic memories, right? And does that kind of map on, so, so semantic is more like generalized knowledge, um, thought to be, you know, accrued over time via many episodic memories. So in some sense, episodic memories are like the, uh, nitty gritty, detailed, um, navigational, um, memories, navigational memories like cognitive me, cognitive maps, right? Found that, that we can kind of localize to the posterior hippocampus. Um, is, is there a, is there a, um, posterior to anterior episodic to semantic, um, indexing in the hippocampus as well? Do you, do we know that
Hugo 01:01:20 Question? Um, not that I know of. I don’t think people are conceiving it, so is a really nice question. But I don’t think people are conceiving it of that way. Uh, if they are, I haven’t seen that argument made. Um, it’s, it’s much more, there’s lots of different perspectives of course, and you’ve had lots of wonderful guests on your program talking about how these things may work. Um, but the, the impression from a lot of research across the field would be the hippocampus is important for, um, helping you acquire, maintain and, and hold onto recent experiences, um, both the anterior and the posterior. Um, and there’s a tendency for the, and the anterior is doing interesting things as well, like the, the, the way it’s organized is involved in suppressing, there’s a lot of like, um, different problem roles in stress and other functions. There’s not just a, a system purely for spatial memory.
Hugo 01:02:07 That’s, and certainly humans. And if you think about the wonderful concept cells, the hall Halle berry cells and so on, they’re distributed in the, in the hippocampus. So I always think back to that work whenever I’m thinking, you know, are the hip, I, I know I, I’m well, well raised in the cognitive map field from John O’Keeffe’s, um, you know, UCL perspective. Yeah. But that is just, it’s really important to think about those concept cells. So, um, you know, how the hippocampus is involved in semantic memory is, is much less clearer, I think. Um, but there’s a wonderful work from Farne Acardo back in 1997 when she discovered that the children with dense amnesia who could learn vast amounts of facts despite not having any episodic memories cuz they had, they had suffered hippocampal damage at birth or just after. Um, so we do see these amazing dissociations, the system is able to adapt to some things and not others is telling us something. Um, but yeah, I still to this day, I think there’s a lot of debate. There’s some very interesting reviews out there on how do we think about episodic and semantic memory and how distinct are they really, you know, they are a continuum now. So I I think it’s this rich interesting to think about these things. Yeah.
Paul 01:03:18 Do you think of, uh, like a cognitive map to j just to schema? Are those three separate things or is that a continuum?
Hugo 01:03:27 Yeah, that would be it. Well, uh, yeah, if you think about it as an active process that you, um, you grow up as a, a person in an environment. So the, the moment you start moving around, I think would be the experience. You would start to build maps of your environment, right? A cognitive map. But you also start to acquire repeated experiences of what streets tend to be like, how wide they are. So you’re, they’re constantly in my mind, going from, from the point as a child, when you’re developing your hippocampus and these circuits to help find your way and remember things in parallel, you’ll be building up that type of knowledge. Um, and that happens straight your whole life really. And then if you choose a career as an architect, you’ll get even better at certain things if you choose a career. So your, your experiences in life are going to shape the extent to which you construct schemers from my mind. These, these expectations, these predictions that your brain has built.
Paul 01:04:17 Okay, so this is what I’ve been wondering about. The London taxi driver increased posterior hippocampus and, and then thinking about if we buy into the idea that the posterior is more concerned with the gritty details, and then as you move more anterior, you, you, um, are more concerned with more abstractions, higher level concepts, quote unquote. And then I thought, why would I want my posterior hippocampus to be bigger? I would want my anterior hippocampus to be bigger. I should not memorize cities. And, and that’s actually could be detrimental because these taxi drivers also had slightly smaller anterior hippocampi. Yeah. Yep. And I’m wondering what the, if I don’t remember, I don’t know what, uh, the taxi drivers IQs are, if there’s a correlation to IQ and their ability to pass this test. Like are low IQs? Uh, no, they’re, they’re more,
Hugo 01:05:10 I don’t think you’ve ever had the IQ tested. It would be interesting to do that, but I think they’d be quite, quite high IQ actually. Uh, I wouldn’t say they, they’d probably be very typical, but if anything’s slightly higher because they spend all their time solving a, a complex puzzle. But, but the interesting thing with that, like you just mentioned, they in the early work and, and replicated, they had a smaller anterior hippocampus, but that replicated with the bus drivers who have the same kind of background of London, same, you know, level of education. So I don’t think that shrinkage is to do with, um, the other, the other things. It seems to be, I mean this may relate. I mean it was, it’s in the original article as a suspicion that it’s some homeostatic process that you know, Kevin Mitchell who’s, you know, you know, it may well have been on your program. I wouldn’t be surprised. Fantastic. He said, well
Paul 01:05:59 What’s this? Yeah, he was <laugh>,
Hugo 01:06:00 You know, he said like, you know, if you keep doing this purely it’s gonna, you know, explode out your skull. It’s a tiny change in the size of this, but it may, there may be some conservation that you don’t just keep adding here. There’s some equilibrium in the distribution in the hippocampus that it’s, it’s getting altered. We really won’t know. I mean, if somebody manages to do a postmortem, uh, collection taxi driver brings, at the end of my life career as a professor, that would be amazing to discover that. But very hard to ethically and, and chase up and do. Um, yeah. Yeah. I I would, I’d still like to know, we, we still dunno in 2020 to 23, let years later, we’d like to find out we still can replicate that effect because it may be that it’s, um, for all sorts of reasons may not still be the case. Um, it has been replicated three times, so this is not an, not a result that has just a one-off and in a high profile journal risk. So, so we’ll, we’ll find out. I I’ll let everyone know when we fi finally get the result.
Paul 01:06:56 If, if, okay, if you, but if I were to, uh, force you to answer this, if you could shrink and grow, uh, one area of your hippocampus, one region and, you know, grow one region and shrink the other, which would you shrink and which would you grow?
Hugo 01:07:14 <laugh>? I think I’d go for the, I’m thinking back, so I meant to say earlier that the taxi drivers have the shrink, they also are worse, uh, story recall and the ray figure copy, which is, you see this abstract picture and you have to reco copy it out, and then they come back with a blind page 15 minutes later and say, could you redraw that picture? You just drew from memory. And they’re, they’re, they’re not, they’re not clinically bad. They’re not, you know, impaired, but they’re significantly worse than London bus drivers and non-taxed drivers. So there is, so if he’s going to answer your question, um, I’m thinking I probably would prefer the detail to be retained cuz that’s what I do as a scientist. There’s a lot of picking through detail, um, then that sort of episodic recall. Uh, but you know, my wife might say are you could do with a better episodic memory for what we’re supposed to be doing tomorrow. She might have a different, yeah, yeah.
Paul 01:08:06 Okay. Well, I mean, you know, I, first of all, I mentioned IQ earlier and I I probably shouldn’t have even said anything cuz it’s not like if, you know, IQ is something, something that we created as a measure of intelligence and, but I tend to kind of think that intelligence is correlated with the higher order aspect of thinking, which then would be correlated more with the anterior hippocampus function in this little story that we’re telling. Um, but I, I mean is that an
Hugo 01:08:33 Yeah, and I think that’s a little simplistic in the sense of, uh, like you’re thinking about how, so to go back to that model that was provided by Lynn Nael and Morris Moscovitch are kind of like, the detail is stored. There’s a more, there’s a kind of re it’s the storage, but the processing of more fine grained information alongside more global information. And I’m not sure that that maps to like the way of creativity or the way you could big big picture thinking. I’m not sure. Sure. It’s that, that, um, just thinking about the kind of evidence, but it is an interesting idea to think why are some people good at big picture thinking and others too detail focused? Is it to do with the way they’re using these, these circuits? I I could be wrong. You could be right. It’s, uh, it is something that you could in theory test and find out are big creative people.
Hugo 01:09:21 Um, you know, I do remember like an experiment I would love to rum but never go around to. But a, a famous author, um, well at least famous to me was Will South, who’d written a lot of, is an English writer, said his brain was just exhausted with all the detail at the end of finishing a book. And it’s like burnout for, for weeks. And they would come back and he said, oh, you’ve gotta scan writers when they finished submitting like a 400 page book. Cause their entire world is in their head. And then a bit like writing a PhD, but somebody’s probably gonna do that at some point or may have done it and I missed it. But I think there’s some interesting questions there.
Paul 01:09:57 Yeah, well, I mean, I, I know that the, the story was like too simplistic and if you set it up an experiment, it it seems rife for if I did it, it seems rife for just confirming my biases Right. By asking such a simplistic and dumb question. I’m gonna get a, an answer and probably a nature publication out now. I’m just kidding. I won’t, I won’t, uh, I’m not gonna criticize the <laugh> depending on what I claim. Right? Yeah. Okay. Um, so, so what, what, moving forward, like what do, what’s the next step in the schema work? Are you testing, um, how are you testing it and what, how are you looking at the nature of the representations that is schema is or are you?
Hugo 01:10:32 Um, we’re not, I mean I, I, it’s a great question. I, I was involved in that review and I think we will carry on mm-hmm. <affirmative> and do some, some analysis. The nearest thing to that is I work with the London taxi drivers, looking at how there, it’s not so much the, so the near, yeah. One way to look at that is to think London, if you were planning a route through it, you could just plan all the possible states you’ve gotta go through. Alternatively you can structure it hierarchically like, this is the area I need to get to, let’s get to the area and then plan within it. And so that gives rise to different predictions in, in how they plan. And that falls under that kind of schemer of how you are representing the structure. So we are actively, we, we have evidence of that and we’re just writing that up in two different manuscripts. Um, where, where we’re looking at it. So that’s where I think we, we’ll be reporting on something schema like, like, um, which is about the hierarchical structuring of knowledge in, in London. Yeah.
Paul 01:11:28 What’s higher than schema?
Hugo 01:11:31 Uh,
Paul 01:11:32 Or what’s, sorry, what’s more abstract than a schema?
Hugo 01:11:37 Maybe a, um, it’s an interesting probe in que cause I had to really think about that one. Um, and I don’t, I mean it is from the literature, I don’t think anybody’s suggesting something. So here I have to make more of a joke about it. I think <laugh> then draw on, on actual art. But when I think about it, it may be your life story. Like you tell me who are you, then it’s like, oh, I have to abstract over lots of knowledge. But that’s not so much a scheme of for what to predict, but gathering different bits of all sorts of experiences to explain something. And you have an expectation there of what it is. The schemer is what is that person expecting from me? It’s not, when I was two, this happened to me a few months later when I was two and a half this happened, <laugh>, that’s not, that’s not the level we want. It’s
Paul 01:12:23 It’s identity. It’s like
Hugo 01:12:25 Identity. Yeah. Your identity of yourself might be higher. Um, but it’s a great, great
Paul 01:12:30 Question like that. Well done coming up with that. Well done. Coming up with that so quickly. <laugh>. I was, I didn’t have an answer to it, so I like that answer though. Okay. Well, um, my schema, my podcasting schema tells me that, uh, it’s time to move on to see her quest. Um, so this is a video game and we started off talking about it where, um, the game, uh, someone memorizes a map and the map has a few locations that they’re gonna navigate in a little boat to, and it’s like a virtual reality kind of game, right? Um, and then, and you helped, like you had a company helped develop the game and then they advertised and you had something like 7 billion people play the game. And so something like that, right?
Hugo 01:13:13 Yeah,
Paul 01:13:13 I think it’s 8 billion, but a lot of people played the game except my children. And, um, so, so, and, and then from this game you’ve been able to extract, uh, lots of different information, um, even about people’s, uh, sleep sleeping habits, their why men are so much better at navigating than women. You must have gotten in trouble for just having that data come out. Um, yeah. So give us the gist <laugh> of the, of what a see her quest and what, what you’ve found so far. And just the power of video games and using video games in, in scientific inquiry, I suppose.
Hugo 01:13:51 Yes. I’ll start with that. Maybe. Um, uh, yeah, I’ll try and give the life story of it in a, not in the, kind of keep it concise for the, for this, this story. Um, but yes, the power of games. Uh, and I obviously, when I started the podcast, I said I was a draft as an undergraduate to a video game type. I did my PhD with video games. So I’ve been working for 20 years exploiting video games, um, to test, to put people into situations that allow me to rigorously test what had happened. In, in, um, around 2015, maybe 2015, I started working with, uh, Michael Hornberger, who’s a dementia research expert, a cognitive expert in, in, in dementia. And he runs the Dementia Research Center in U e a, uh, at University of Lia. So he and I have been working together thing could we help develop some diagnostic tools for, for, um, Alzheimer’s, which he’s an expert in, uh, involved in video games, but it’s really hard to do.
Hugo 01:14:43 And then one day he got, got an email, he got a, someone contacting him, say, there’s a company out there with an unlimited budget who would like to do something like this kind of crazy thing we talked about earlier. Uh, and he contacted me and said, oh, I’ve got this very strange request for a one page outline for probably come to nothing. So we filled it out, helped Michael apply this, this, it went back and then a couple months later he got an email saying, yeah, you’ve been picked. That’s it. We wanna work with you, uh, to do this thing, God. And it was at that point it became apparent that the, the funder was Dge Telecom or T-Mobile whose avenue their, their revenue per annum there to turn over 68 billion <laugh>. So they have some money, they’re not a, they’re not a poor, they’re the largest telecoms, richest telecoms company.
Hugo 01:15:27 But their idea was just, just to take their, their advertising budget for one half a year or whatever it was for a year and do something really innovative. And they employed Sachi and Sachi, this creative agency who are famous in the UK for crazy advertised, employed them to come up with a way to use the budget. And their idea was let’s make a video game that can do something, um, fun that, that, that’s, that that gets your message about UK to people. Uh, and the idea they picked was dementia, doing something with dementia. Uh, and then we got picked Michael and I to, to then create, um, a test that would hopefully help with Alzheimer’s. So that’s how we got there. It wasn’t, uh, we got some games designers to come in and Oh no, it was, uh, a funder came to ask for the specific project, specifically asking, and we gave them a project.
Hugo 01:16:13 But it was incredible. You know, we, we got a commercial games company, we were able to pick, uh, and work with this company Glitches then for a whole year at high speed developing less than a year. I think we went from idea to, to out there to produce, cuz that’s the way a telecoms company would work. Um, but right, it’s, it’s a video game launched in 2016 on the app store, on Google Play, uh, and you play a little boat and, um, you just, you go around trying to find sea creatures and take photographs of them, uh, and share them with your friends. And, and because it was a telecoms company, you know, we had a, we had a good advertising budget to promote it on all sorts of platforms. Um, and one of the key things that budget after a lot of negotiating from the PR team was they, they, they stocked and tracked down and harassed PewDiePie, the number one YouTuber in the world at that point. I think he still may be the number one YouTuber, but they convinced him to advertise it. And so he pitched on his YouTube channel. Oh,
Paul 01:17:08 Is this like the, uh, where people watch other people play video games?
Hugo 01:17:12 Exactly, exactly. So he’s one of those people that plays video games and talks about them and does a lot of other things, but he, um, he had 5 million people look at our game in two days. So it was an incredible, but he, he was paid a lot of, uh, funding to do that <laugh>. Um, and he donated all of it charity though, so I have to say very helpful guy and, and was good <laugh> in that scenario. Um, but that, that’s, that’s some of the background to how we ended up with 4 million people taking part. So we have people from 195 nations enter their data. I mean, you know, and one of the big things you should be asking is that, are we really, you know, did everyone who said they were from the Vatican really from the Vatican, which you, we doubt, you know, so we, this is a great, we, we look, we’re looking forward to publishing an article.
Hugo 01:17:58 It’s just like, how much bias is there <laugh> in a data set like this? How do you spot it and what is it? Um, so we, we’ve seen the bias in excruciating fantastic detail, but it’s good. Like it’s always there. It’s in every lab experiment, you know, if you wanna test people in your lab, not everybody, not everyone’s gonna come off the street and do your experiment. So, so I think, I think that’s been amazing. But yeah, the game, the fact, the effectively the game generates time series data. So that was when you said video games for research, the really valuable thing about video games is people will do lots of it. If they enjoy it’s a well made game, you’ll have people do it in a naturalistic way. Uh, and we had the capability with the games company to do that. So people intrinsically wanting to do this, unlike paying them, and, um, the data’s really rich.
Hugo 01:18:46 But what I found with the project was that having lots of people play it, you can use the time series of their data, of their trajectory data to look at things in a lot more detail than you ever could on a normal test that’s got a score of one to 25 or whatever. Um, and so we’ve produced quite a lot of papers just looking at the trajectory. Le like lots of distance people are traveling. So the really good navigators take a short distance and they tend to be, um, young people in their twenties, uh, in many countries. They tend to be male. And um, they come the top place in the world they might come from as Finland, for example. So we we’re able to look over the data, uh, and understand how G D P seems to correlate with performance in our task across the nations. Um, but it, it’s been an absolute joy to, to look at big data so you can really screen through it, look at the consistency, and then, you know, be quite robust. You’re no longer thinking our P values, you’re just interested in effect size, how much does this matter? Oh
Paul 01:19:44 Yeah.
Hugo 01:19:46 So for example, we find that left-handers, you know, they navigate maybe differently, but it, or like they sleep longer by one minute <laugh>, you know, and it’s not a real effect, you know? Um, so, so the, the
Paul 01:19:58 Real effect there is one minute longer. Is that, what is that what you said there
Hugo 01:20:01 Somewhere in our data set? I remember think I I I I shouldn’t, that that isn’t a correct figure, but there is a sleep, there will be a sleep difference because you’ve measured, we got 700,000 people telling us whether they were left or right-handed. And there will be, it will be significant if they sleep half a second longer with 700 data samples, it will be significant <laugh>. Uh, but it’s not as meaningless. Yeah. That’s the kind of way of thinking out big data, right?
Paul 01:20:26 One of the, uh, stories that you tell is, um, is about how, uh, people from suburbs, uh, tend to navigate better than people from cities, but not just from like inner cities, but not just inner cities, inner cities that are laid out in a very grid-like manner, what you term grittiness, um mm-hmm. <affirmative>. And, and that, um, if you live in a city that is really gritty, you’re a terrible navigator in, uh, in sea quest, and if you live in a suburb or a city that is less gritty or, um, something more like Pittsburgh, where I’m gonna be soon again, um, that you tend to be better at navigating. So I don’t, you know, flesh that out a little bit more than I just simplistically did. And
Hugo 01:21:13 Yeah, no, I mean that is sort of capturing some of the key, the key result. We, we published that in nature in April last year, where, um, the key, key key question we looked at with the data is this very simple. We asked people, did you grow up in a city or some other situation, suburb, rural mix, um, and the data clusters. So it’s very clear there are people from cities or not cities when you look at the data. Um, and then you, we’d simplistically asking with linear models or linear mix models, depending what’s in the model, whether, um, people who grew up in those two different situations navigate differently. Is there a difference? And like I said, you can get meaningless, like, oh, they are, but they’re only a little bit better. And what we were really shocked when we first looked at the data, I was shocked, was that, um, there is quite a reasonable difference.
Hugo 01:22:00 So, um, across the lifespan from say early twenties through to to late seventies, there’s about a four to five year difference from men and women between people who grew up inside and outside cities where it’s worse growing up inside cities. But if you look at later in life, a woman who grew up, this is on average across all the different nations we looked at, across all the different countries, a woman who grew up outside a city, um, in her sixties is equivalent to a woman who grew up inside a city to 10 year difference in skill. So for some fact, something to moderate your lifespans ability is, is pretty shockingly big effect. And then, like you said, that’s the average in the world. We found this effect was really big in the usa but very, very and and non and very overlapping with Argentina. So we’re like, why are Argentina and the United States coming out where, you know, some other countries are, are much, you know, further down, um, like why is, uh, you know, Ireland far back here?
Hugo 01:22:59 What is it about Ireland? Uh, and at that time we’d seen this geography paper that this analytical paper by Jeff Boeing went round on Twitter showing you could, you could measure how gritty every city in the world was using his, uh, very simple analysis of the entropy of the streets. If they’re, if the streets are all lined up, you can, you can measure that mathematically as a highly organized low entropy system. But if you take somewhere like Sao Paolo, there is almost no, there is no orientation in Sao Paolo, in central Sao Paolo. They all go in different directions. It’s highly entro. Um, and so Antoine Coutre, who’s the first author and spent now five years looking at this data with us initially as a postdoc and then as a pi, he noticed this and said, let’s, let’s look at that. And lo and behold, there was a very significant correlation across 38 countries between the grittiness of the country’s city streets and how badly they were affected by growing up in a city.
Hugo 01:23:52 And so that was the central result in that research paper that then, of course to, to publish it in a, in a highly rigorous journal, you’d go on a very long journey to do a lot of other things like show it’s not to do with where you live now, test a thousand other people replicate the effect, replicate the effect if you do a driving simulation. So it’s not taking a boat round an aquatic environment that goes. Um, but yeah, as you said, like in the game, we can look at the entropy in the game levels. So across 45 levels, we could see that the, the really, really disorganized levels maximally pull out this disadvantage of growing up in say, Chicago. So if someone from Chicago is gonna find those much harder, but the title of the paper is Not Cities Are Bad For Your Navigation.
Hugo 01:24:36 That’s not what the title of the paper is. It’s, I, I forget the exact title, but it’s like entropy shapes your future spatial ability, because we found that in the really gritty levels in senior request, people from gritty cities actually had a slight advantage. Although there’s a somewhat of an overlap with non-city people. So it’s not that it, it’s, so that goes back to the schemers. So it seems that people who grew up in city environments have an expectation of what they expect to see when they navigate a world and um, it biases them. Um, so, so that’s the real, there’s a lot of, a lot of peripheral and loads of endless supplemental figures in that paper to, to ram home the point and take it apart. And we had five reviewers who asked a lot of questions about that data. Oh wow. It made a much better paper. Yeah, yeah. The editor apologized, which she said, very sorry, but we’ve given you five reviewers to make sure every single one of them is happy before we publish this
Paul 01:25:34 <laugh>. My God. But is there a, so there’s, you were saying that, you know, there’s not a story to be told necessarily of like where you should grow up or something. Um, no, and you were mentioning, you know, with such a large dataset, you’re gonna get, um, good P-value. Well, um, you’re gonna get effects. Just the question is like how large those effects are. But then the question is, you know, even if those effects are kind of large, like do I care what, why do I care about whether yeah, um, I can navigate like for my real world, um, experience moving through the world? Like, just cuz I can’t get around quite as well in a city, does that mean, you know, do I really need to worry about that? How, how much should I worry about about that? About it?
Hugo 01:26:18 No, I agree. I wouldn’t worry about that. I mean, you know, I think if you look at clinical, uh, issues, surgeons doing neurosurgery are not that worried about you losing your spatial skill that they might tamper with your language ability. They’re terrified, right? Mm-hmm. <affirmative>. So that, that is a known fact for neurosurgery. You do not want to damage the language. So, so I think that that’s, that’s here I agree. I, I think it’s fascinating and it’s one of the few studies in the world where we’ve looked at how some experiences in childhood are then shaping your, your future experiences later. It’s very hard to study these on a mass scale, but we kind of, as we would assume that it’s likely in, in areas like language and other, other aspects of how it’s affected. Um, but yeah, we shouldn’t worry about it, I don’t think. I think the only, the only thing there is I’ve was really surprised, like I said, this 10 year difference, and if you think you’re losing this skill, it’s like, well, there is a clear advantage. And as Antoine said to me, once mul over the data, it’s like, well, you know, does it matter growing up in a gritty city? Well, turns out most of the world is not gritty, so you are kind of in disadvantage. Yeah. So it is better probably to grow outside cities and get that experience on average. On average.
Paul 01:27:30 Hmm. So <laugh>, just so in the, uh, in the United States anyway, I guess people who voted for Trump are really good at navigating and people who voted Democrat, uh, maybe, maybe aren’t as good at navigating. I I, I hadn’t thought about this before. Um, and I’m apologize if you are doing this work, and I just o overlooked it already. But are you training re reinforcement learning agents to play see her quest also, and then, um, comparing that like different kinds of reinforcement learning algorithms to the different developmental stages of, um, humans, et cetera, and how they navigate?
Hugo 01:28:07 No, that’s, that’s really, it is really nice you’ve raised that because we, we, we have thought about that and we’ve looked into it. We’ve, we’ve recently got more funding to do things like that. Um, but actually that’s a good example where somebody listen to your podcast might be interested in this and get in touch because the data’s so rich and so large that there’s really good scope for collaborators to, to join and do things. We do have three teams playing with the data. Every single group who comes to do something with the the data will see it differently and have their own questions. So we haven’t tried kind of building reinforcement learning agents to, to do that, to, to play the game. But it is plausible. It’s a very interesting angle. And, um, if anybody is interested, I mean, I, I should also take this point to pitch and hopefully on the website maybe we can put some details where the game is totally free for people to use for research. There’s no, there’s no cost or anything. It’s completely set up to maximum F 60 projects, but also the data’s completely accessible. Everything we’ve done with it, someone can download. Um, and we’d be delighted to kind of facilitate someone doing something with this data. Um, so there, there’s, it’s, it’s, it’s hopefully more people will do things with this data, um, and we’d love to speak to them to help them do it.
Paul 01:29:19 Do you have like a ton of other, like, I, I know that one of the things that you were interested in was sleep. Why were you interested in sleep? And then of course the finding, the robust finding is that seven hours is the perfect amount of sleep for navigation, right, <laugh>.
Hugo 01:29:32 That’s right. So this is an nature communications article end of last year. Yeah. So we, so when we, when we set up, I should say when we, the data looks really rich, right? And it is rich spatially. We have wonderful time series and 4 million people, um, but maybe 70% or less fill out all these demographic questions and working on the project, we had a lot of fight with the developers and the team, but they said that you really can’t ask more than nine questions. People are not gonna answer <laugh>. I think they might have, but their, their criteria as well was not only, we wouldn’t ask loads of questions, but it was, these have to translate into 17 languages and be iconic. Oh, you can just click and understand. Um, and so sleep was an easy one to say, how many hours of sleep do you typically get?
Hugo 01:30:15 Um, but it does also related to sleep disruption in Alzheimer’s. So a lot of our questions were kind of else would this help in refining the data? Um, but it’s a great question. Um, and yeah, like you said, what we found was that later in life, right, so after the age, you, after your, you know, early fifties, um, through to to 70, getting seven people who report seven hours are better navigators than people report eight, nine, or any other range doesn’t recommend you should sleep seven hours. It’s just the people who tell you that they sleep seven hours tend to perform better. Uh, so you can make that causal relationship and say, it is good to get seven hours, but the data on by itself doesn’t, doesn’t provide the direct <laugh> direct evidence for that. Um, but yeah, it, it was really struck. It, it wasn’t.
Hugo 01:31:03 And the key, the key, the key exciting thing I saw in that data set, um, with the sleep data was observing a pattern no one had seen before. So it’s not too surprising that getting optimal sleep might help your cognitive skill, but what we hadn’t known until that paper was that if you look across the lifespan for a reported sleep, it’s not smooth. There’s something happening to humans that between the age of early life, through the end of your teenage years to 33, it declines. You get less and less probably when you’ve got children, it’s really low in plateaus. And then after 53 mm-hmm <affirmative>, it goes higher. So it’s like, it’s not a smooth u shape. It’s, and we don’t know why we say in the article, we don’t know why those particular age bands are fixed like this. But the amazing thing for us was when you separate the data set in half and just say, well, look at men and we’ll look at women, you think these numbers are gonna change and they didn’t, 33 53 for both men and women, huh?
Hugo 01:32:04 And you say, well, all the western, you know, people didn’t, who’ve got low. The high GDP countries who are all living in like rich scenarios with loads of facilities and lay video games and the, these group, maybe they all have different numbers to the people in low middle income countries. No, same numbers again, 33 and 53. So there’s something going on with the human lifespan and sleep around 33 and 53 on average that we don’t know why. So maybe, maybe at the end of my career, someone will answer me and say, ah, you know what it is, I don’t know
Paul 01:32:37 The moment. By then you’ll be sleeping more.
Hugo 01:32:39 I hope so, <laugh>, that’s what the data suggests.
Paul 01:32:42 Yeah. Yeah. <laugh>, you know, I I know it’s 10:00 PM in your world and you, your children have slept through this whole thing. Yeah. By way. Thank you for that. We, we got lucky. But, but you gotta get some sleep. Even though you’re at that stage of your life where you don’t get any sleep. Yeah. You still need a real sleep. So I appreciate you staying up so late to talk to me and, um, finally doing this with me. We’ve had a registry reschedule a few times, and so thanks.
Hugo 01:33:04 Oh, no problem. It’s been a pleasure. I’ve been really enjoying your podcast for, for years now, and, uh, yeah, it’s a real delight to, to finally make it after scheduling onto this. So I really appreciate you taking the time to, to talk to me and, um, yeah, I’ll continue to enjoy your podcast going forward. I’m sure it’s fantastic. Ranger guests, you’ve had so great and you, you’ve asked me fantastic questions that other podcasters don’t ask. And, um, it’s, it’s great. Really appreciate it.
Paul 01:33:44 I alone produce Brainin inspired. If you value this podcast, consider supporting it through Patreon to access full versions of all the episodes and to join our Discord community. Or if you wanna learn more about the intersection of neuroscience and ai, consider signing up for my online course, neuro Ai, the quest to explain intelligence. Go to brand inspired.co to learn more, to get in touch with me, email paul brenna inspired.co. You’re hearing music by the new year. Find firstname.lastname@example.org. Thank you. Thank you for your support. See you next time.