WEBVTT 00:00:00.017 --> 00:00:03.037 All right, so you can wear the headphones and we all look more professional, 00:00:03.137 --> 00:00:04.057 we feel more professional. 00:00:05.957 --> 00:00:09.657 This is the Convergent Science Network. We do the flight of the Concorde's robots. 00:00:10.837 --> 00:00:15.077 Reading researchers in the domain of neuroscience, brain theory and technology 00:00:15.077 --> 00:00:18.697 are interviewed by Paul Boucher and Tony Preston. 00:00:19.317 --> 00:00:22.377 No, it's you that's breathing. No, no, no, no. It is. No, no, 00:00:22.377 --> 00:00:24.877 it's you. No, no, no, no. Honestly, check. 00:00:27.177 --> 00:00:31.257 Okay, glad we settled that. Wear your headphones. This is Paul Vesure with the 00:00:31.257 --> 00:00:35.377 Convergent Science Network podcast together with my colleague Tony Prescott, 00:00:35.617 --> 00:00:40.157 BCBT 2015 in Barcelona. And we're here with Kate Jeffery. 00:00:40.597 --> 00:00:46.497 And Kate, you were presenting your work, which is actually very exceptional 00:00:46.497 --> 00:00:50.577 in the sense that you look at how three-dimensional space is represented in the brain. 00:00:50.697 --> 00:00:55.517 So how did you come to studying three-dimensional space from the perspective of the rodent brain? 00:00:55.777 --> 00:00:59.837 Are there flying rodents somewhere? There aren't flying rodents, 00:00:59.857 --> 00:01:03.597 but I had been studying two-dimensional space for a long time, 00:01:03.617 --> 00:01:05.737 as everybody else had been as well. 00:01:06.577 --> 00:01:10.857 And I guess we just started to get interested in the question of whether this 00:01:10.857 --> 00:01:15.577 map is really just a flat map, or whether it's actually got some three-dimensional 00:01:15.577 --> 00:01:17.157 structure. The world is three-dimensional. 00:01:18.897 --> 00:01:24.217 And we got into this line of work because I was talking with my friend and colleague, 00:01:24.337 --> 00:01:27.997 Andre Fenton, who had been thinking along similar lines in New York and he had 00:01:27.997 --> 00:01:34.117 built this fantastic spiral staircase and the idea had been to see if these 00:01:34.117 --> 00:01:37.917 place cells would produce place fields on the spiral staircase and whether there 00:01:37.917 --> 00:01:41.017 would be a vertical kind of structure to the place fields. 00:01:41.737 --> 00:01:44.317 And he didn't have time to run the experiment so he said, look, 00:01:44.337 --> 00:01:46.017 I'll bring it over to London and you can have a go with it. 00:01:46.677 --> 00:01:50.937 So I sent my student Madeline Veriotis onto recording on this thing. 00:01:50.977 --> 00:01:53.597 It was very, very difficult because the rats go round and round and round and 00:01:53.597 --> 00:01:58.537 round and it's like running up and down a five-floor building all day long, 00:01:58.637 --> 00:02:01.157 so they got quite tired, but she managed to get some really nice data. 00:02:01.997 --> 00:02:07.297 And we found that place fields indeed seemed to extend into the vertical dimension. 00:02:08.197 --> 00:02:11.897 And then around that time grid cells were discovered, and so it became a natural 00:02:11.897 --> 00:02:15.977 question, do grid cells also show some vertical structure? 00:02:16.477 --> 00:02:19.817 So she began recording grid cells on this spiral staircase, case, 00:02:19.877 --> 00:02:24.097 and another student that I had started recording them on this other piece of 00:02:24.097 --> 00:02:25.457 apparatus, a climbing wall. 00:02:26.190 --> 00:02:32.190 And we found these results that were quite surprising and yet consistent with each other, 00:02:32.230 --> 00:02:36.870 which is that the grid cells which fire in this periodic way in the horizontal 00:02:36.870 --> 00:02:41.430 plane didn't do that in the vertical dimension on either of those apparatuses. 00:02:41.790 --> 00:02:46.230 And so then that led us to do some thinking about why that might be and basically 00:02:46.230 --> 00:02:49.590 the whole research program took off from there. Okay. 00:02:49.690 --> 00:02:56.690 So you're in this laboratory, which also includes O'Keefe, who won the Nobel 00:02:56.690 --> 00:02:58.970 Prize for his work on the play cells. 00:02:59.170 --> 00:03:04.510 So the whole environment dedicated to understanding spatial cognition in the rat. 00:03:04.510 --> 00:03:14.190 Um so were you i mean if you're looking at the opportunities you have within that playing field. 00:03:14.730 --> 00:03:17.510 Which in some sense that the structure doesn't give you a 00:03:17.510 --> 00:03:20.170 lot of different cells to look at because you would have your grid cells you should 00:03:20.170 --> 00:03:23.850 play cells and you would have heading direction cells okay so 00:03:23.850 --> 00:03:26.750 um now in 00:03:26.750 --> 00:03:29.590 in the end if you look at the three-dimensional representation or in 00:03:29.590 --> 00:03:32.310 the red brain you focus very much on on the grid 00:03:32.310 --> 00:03:35.530 cells ultimately but but we're not there yet right but because 00:03:35.530 --> 00:03:38.530 that's not necessarily where you started so how did 00:03:38.530 --> 00:03:41.190 you work your way through that system and why did you make them the 00:03:41.190 --> 00:03:46.050 different choices that you did make well it it was partly just chance so we 00:03:46.050 --> 00:03:49.230 started with place cells because at the time we had placed cells and head direction 00:03:49.230 --> 00:03:55.770 cells um and jeff talby who had been working very intensively with head direction 00:03:55.770 --> 00:04:01.030 cells had already started looking at how they they behaved in three dimensions did some really 00:04:01.170 --> 00:04:04.690 beautiful work, which I didn't fully come to appreciate until we started to 00:04:04.690 --> 00:04:07.410 do something similar. I realised it's very, very difficult. 00:04:08.810 --> 00:04:11.110 So play sales are the natural place to start, really. 00:04:11.940 --> 00:04:14.800 But then grid cells came along, and the exciting thing about grid cells is that 00:04:14.800 --> 00:04:20.320 they have this metric component to their activity where they are actually encoding distance. 00:04:21.660 --> 00:04:28.680 So they are plausibly the fabric for this map, if you like, the thing that actually 00:04:28.680 --> 00:04:32.800 gives it the capacity to do navigational calculations and things like that. 00:04:33.260 --> 00:04:36.460 So when grid cells came along, it just became a very, very exciting question. 00:04:36.560 --> 00:04:41.500 Do they measure out distances in three dimensions? And if they do do that, how do they do that? 00:04:41.940 --> 00:04:45.200 And we know they need a compass to be able to do what they do in two dimensions. 00:04:45.320 --> 00:04:47.620 So does that mean they need a three-dimensional compass? 00:04:48.260 --> 00:04:51.880 We know that they form these beautiful hexagonal patterns on a flat surface. 00:04:51.980 --> 00:04:55.180 Does that mean that they form a three-dimensional lattice pattern? 00:04:55.320 --> 00:05:00.900 All sorts of questions came along following on from that discovery. So it was very exciting. 00:05:01.260 --> 00:05:05.640 So you started out in your talk saying that to navigate, you need a map, 00:05:05.660 --> 00:05:09.700 a compass, a way of measuring distance and a way to self-localize. 00:05:09.700 --> 00:05:12.640 But then as your talk progressed it 00:05:12.640 --> 00:05:15.460 seemed that maybe in the rat things aren't quite 00:05:15.460 --> 00:05:18.540 so clear-cut because like you 00:05:18.540 --> 00:05:23.040 say the grid cells when people first found them thought this is the metric this 00:05:23.040 --> 00:05:28.940 is telling me how far i am in space relative to some starting point but then 00:05:28.940 --> 00:05:34.400 as the evidence came in these grid cells also appear to depend upon contextual 00:05:34.400 --> 00:05:38.460 cues and you talked about boundaries reason you talked about odors. 00:05:38.520 --> 00:05:44.740 So, I mean, can you just explain how you think these other conceptual clothes 00:05:44.740 --> 00:05:48.140 are affecting the grid cells and what impact that has on this. 00:05:48.831 --> 00:05:54.951 Idea of how we navigate in space? Well, so contextual cues are things that characterize 00:05:54.951 --> 00:05:58.451 a space but don't have spatial information in and of themselves. 00:05:58.891 --> 00:06:04.151 So in the rat, we use the manipulations of the color of the environment or the 00:06:04.151 --> 00:06:06.591 smell of the environment or things like that. 00:06:06.991 --> 00:06:11.551 For ourselves, we could think of things like the decor in a room or something like that. 00:06:11.611 --> 00:06:17.031 So that helps you know which room you're in, but doesn't really tell you where you are in the room. 00:06:18.011 --> 00:06:22.611 My lab had been interested for quite some time in how those contextual cues, 00:06:22.731 --> 00:06:25.871 those non-spatial cues, modulate the activity of place cells. 00:06:26.891 --> 00:06:30.171 We had come up with a model that suggested that the place cells are getting 00:06:30.171 --> 00:06:35.171 spatial information from somewhere that's fairly raw metric information about 00:06:35.171 --> 00:06:36.771 boundaries and distances and directions. 00:06:37.291 --> 00:06:40.971 They're getting contextual information through another pathway, 00:06:41.171 --> 00:06:47.051 and those two pathways interact, act and the contextual cues act to select which 00:06:47.051 --> 00:06:50.151 of the spatial inputs a given place cell will respond to, 00:06:50.851 --> 00:06:54.311 so that was the model that we had at the time that grid cells were discovered 00:06:54.751 --> 00:06:57.531 and then when they were discovered it seemed like they might be the spatial 00:06:57.531 --> 00:07:01.671 component of this model so these are these things that seem to be spatial but not, 00:07:02.491 --> 00:07:04.151 much more than that as far as we could tell, 00:07:04.971 --> 00:07:08.271 so then we sort of became curious well what do they do when we change the context 00:07:08.271 --> 00:07:11.131 and we actually thought changing the context would have very little effect on 00:07:11.131 --> 00:07:14.411 grid cells because we thought Their job is just to mark out distances. 00:07:14.551 --> 00:07:17.571 Why should they care about the colour or the odour of the environment? 00:07:18.291 --> 00:07:21.831 So it was a little bit surprising when we found that they do actually respond. 00:07:24.511 --> 00:07:29.251 So now, what we're thinking, trying to put together what we've observed together 00:07:29.251 --> 00:07:33.251 with what makes adaptive sense, when you think about the evolutionary function of these things, 00:07:34.691 --> 00:07:40.211 my thinking is that the contextual cues indeed interact with the spatial cues 00:07:40.211 --> 00:07:41.251 to drive the place cells. 00:07:41.251 --> 00:07:45.871 Cells the place cells in turn are helping the grid cells know where to fire 00:07:45.871 --> 00:07:50.531 so we're thinking and it's not our idea it's something it's an idea that many 00:07:50.531 --> 00:07:53.891 people have contributed to but we're thinking that there's this to and fro interaction 00:07:53.891 --> 00:07:57.991 between the place cells and the grid cells where the grid cells help the place 00:07:57.991 --> 00:08:00.751 cells remain oriented in the middle of a big open space, 00:08:01.311 --> 00:08:04.671 and then the place cells help the grid cells to know which environment they're 00:08:04.671 --> 00:08:07.411 in using the context XQs and all the rest of it. 00:08:07.591 --> 00:08:12.591 And that makes sure that the grid cells will fire in the correct place for a given environment. 00:08:12.971 --> 00:08:16.271 So the two cell types are kind of helping each other out. 00:08:16.611 --> 00:08:22.851 And that is a little bit how people go about building simultaneous localization 00:08:22.851 --> 00:08:25.391 and mapping SLAM systems in robots. 00:08:25.971 --> 00:08:32.711 So I know that SLAM was actually informed by research on rodent navigation. 00:08:33.531 --> 00:08:36.731 David Reddish actually was pointing this out to us a few weeks ago. 00:08:37.591 --> 00:08:43.571 To what extent is your research now being influenced by these ideas from modeling 00:08:43.571 --> 00:08:46.611 and maybe even from robotics? David Reddish, 00:08:47.121 --> 00:08:50.261 Well, one of the things that's come along recently, 00:08:50.441 --> 00:08:55.441 which I admit to not knowing very much about, but I'm finding increasingly intriguing, 00:08:55.621 --> 00:09:01.321 are these deep neural networks that roboticists have started to use in some 00:09:01.321 --> 00:09:04.041 of their kind of robot models of navigation. 00:09:04.041 --> 00:09:10.101 And I was always sceptical about neural networks as a general kind of concept 00:09:10.101 --> 00:09:14.321 because a neural network is a very homogeneous thing to my untrained eye. 00:09:15.281 --> 00:09:19.001 Whereas we could see in the brain that it's very modular. There's a module for 00:09:19.001 --> 00:09:22.441 processing compass information, there's a module for processing distance and 00:09:22.441 --> 00:09:23.981 a module for processing this and that. 00:09:24.541 --> 00:09:29.481 It just feels to me like the architecture is very much more intricate than you 00:09:29.481 --> 00:09:30.221 get with a neural network. 00:09:30.221 --> 00:09:33.901 So for a long time I had felt that neural networks 00:09:33.901 --> 00:09:37.721 have their uses but they don't really explain how the brain works but it 00:09:37.721 --> 00:09:41.361 transpires with these deep neural networks that have many many layers that when 00:09:41.361 --> 00:09:44.501 you when you train them up and train them up and train them up they start to 00:09:44.501 --> 00:09:48.461 acquire some internal structure actually and when you probe elements of these 00:09:48.461 --> 00:09:52.641 things you do see things that look like they have subcomponents of the of the 00:09:52.641 --> 00:09:55.981 cognitive computation so I guess what I'm 00:09:56.141 --> 00:10:00.461 starting to take back from the robotic field is that maybe you can get something 00:10:00.461 --> 00:10:03.601 that looks like a modular system out of something that started out fairly homogeneous 00:10:03.601 --> 00:10:07.301 with some simple rules and a lot of parallel processing capability. 00:10:07.981 --> 00:10:09.981 And I think that's kind of an intriguing idea. 00:10:12.472 --> 00:10:15.852 I'm not sure how it would inform our experiments, but it certainly informed 00:10:15.852 --> 00:10:19.552 how I think the brain might come to do what it does. 00:10:19.732 --> 00:10:24.872 Right, so maybe it would inform experiments on the development of this system? 00:10:25.032 --> 00:10:27.452 I don't know if there's a big literature on that yet, is there? 00:10:27.752 --> 00:10:36.752 A literature is starting to develop. So recently, people from UCL and also from 00:10:36.752 --> 00:10:40.952 Trondheim in parallel have been looking at development of the glugurud cell 00:10:40.952 --> 00:10:42.792 and place cell in the head direction. system. 00:10:43.372 --> 00:10:46.912 And these cells come on stream very early. 00:10:47.532 --> 00:10:50.272 Interestingly, the grid cells seem to come on latest of all. 00:10:50.372 --> 00:10:52.492 And the first thing to come along is the head direction cells. 00:10:53.852 --> 00:10:57.032 Which makes a certain amount of sense, actually, because you can imagine that 00:10:57.032 --> 00:10:58.332 compass direction is primary. 00:10:59.392 --> 00:11:04.912 But we were starting to think that place cells emerged from the activity of grid cells. 00:11:05.152 --> 00:11:08.132 And then when these developmental findings came along, we started to think, actually maybe 00:11:08.132 --> 00:11:11.192 it's the other way around and maybe um maybe it's 00:11:11.192 --> 00:11:15.612 far more interactive than we had realized um i 00:11:15.612 --> 00:11:19.472 think the development development story 00:11:19.472 --> 00:11:24.092 assumes a certain amount of hardwired modularity so i don't think it's completely 00:11:24.092 --> 00:11:29.772 analogous to the deep deep learning networks quite yet but i think we will probably 00:11:29.772 --> 00:11:32.272 be moving towards some kind of hybrid system where the deep learning networks 00:11:32.272 --> 00:11:36.632 have a modularity to start with and then they take off and you probably know 00:11:36.632 --> 00:11:37.832 far more about this food than I do. 00:11:37.832 --> 00:11:42.252 Well, I think, yeah, we've had some interesting discussions about deep learning. 00:11:42.372 --> 00:11:45.612 We've managed to avoid this as a topic for this summer school until now, 00:11:45.772 --> 00:11:47.452 so I'm pleased it's come up. 00:11:47.592 --> 00:11:50.952 I know Paul has some strong views on deep learning. 00:11:51.132 --> 00:11:53.912 I thought I detected some raised eyebrows there, so it would be interesting 00:11:53.912 --> 00:11:56.392 to hear the pros and cons, actually. 00:11:56.512 --> 00:12:01.792 I think they're deeply overrated and overhyped in computational neuroscience, 00:12:02.012 --> 00:12:07.032 these ideas years of, let's say, uniform computational principles giving rise 00:12:07.032 --> 00:12:12.012 to cortical-like filter hierarchies are around since the late 80s, 00:12:12.012 --> 00:12:15.632 where people have been calling them objective functions. 00:12:15.792 --> 00:12:19.052 So we can think about building classifier hierarchies using. 00:12:19.532 --> 00:12:25.052 Let's say, the reduction of redundancy in the course of sparse coding, 00:12:25.292 --> 00:12:27.592 which Olshausen and Fields proposed first. 00:12:27.892 --> 00:12:34.312 Other people like Peter Koenig, myself, have been emphasizing issues like smoothness, 00:12:34.472 --> 00:12:40.692 like you can acquire a whole cortical-like hierarchy from V1 to hippocampus 00:12:40.692 --> 00:12:46.212 by hippocampal-like place cells by just optimizing the slowly varying features 00:12:46.212 --> 00:12:48.632 in your input and decorrelating them within layers. 00:12:48.732 --> 00:12:51.012 So this stuff we know for at least 20 years. 00:12:51.212 --> 00:12:53.792 And deep learning is a variation on that theme. 00:12:54.792 --> 00:13:02.592 It's just done by people who tell a somewhat different story and who are a little 00:13:02.592 --> 00:13:06.092 bit less informed about the neuroscience, so they have a tendency to overgeneralize. 00:13:06.432 --> 00:13:09.312 Because the point is, what does it mean to explain the brain? 00:13:10.168 --> 00:13:13.248 You can take inspiration from whatever you want. You can take inspiration from 00:13:13.248 --> 00:13:17.108 this cup of water. You can take your inspiration from a deep learning network. 00:13:17.348 --> 00:13:20.508 But explaining the brain also means that you have to account for the constraints 00:13:20.508 --> 00:13:21.728 that you find in that system. 00:13:21.988 --> 00:13:27.888 And if you look at the grid cell system as an example, just the anatomical structure 00:13:27.888 --> 00:13:32.008 has very specific properties that you just don't get for free. 00:13:32.328 --> 00:13:35.368 They have a very interesting dorsal ventral organization. 00:13:35.368 --> 00:13:38.308 They are organized also in sort 00:13:38.308 --> 00:13:41.048 of in the laminar sense in a very curious kind of way 00:13:41.048 --> 00:13:44.928 these are the kinds of things you must explain right. 00:13:44.928 --> 00:13:49.148 And that you can have a model that learns these grid-like properties 00:13:49.148 --> 00:13:51.948 in some way okay but that's that's not really explaining 00:13:51.948 --> 00:13:55.088 anything that tells you okay i can get a grid-like 00:13:55.088 --> 00:13:57.788 response in something but i think you have to 00:13:57.788 --> 00:14:00.548 also show that that you do it with networks that are 00:14:00.548 --> 00:14:03.388 informed and constrained by the known physiology and anatomy 00:14:03.388 --> 00:14:06.248 to me and that's usually not done for instance there are 00:14:06.248 --> 00:14:09.068 a number of so-called attractor models of grid cells where people 00:14:09.068 --> 00:14:12.008 show i can get grid-like responses if i 00:14:12.008 --> 00:14:14.788 have an attractor which basically means i have the liberty 00:14:14.788 --> 00:14:17.528 in the world to wire up a bunch of cells and now i have a 00:14:17.528 --> 00:14:21.268 grid-like response but that's not the grid-like response you 00:14:21.268 --> 00:14:24.248 have a grid-like response if you can show if you drive your map 00:14:24.248 --> 00:14:26.988 with the velocity vector that is continuous and relates 00:14:26.988 --> 00:14:29.808 to the movement of your agent that then you have a repetitive pattern 00:14:29.808 --> 00:14:32.988 of firing that follows the a good cell structure and that's 00:14:32.988 --> 00:14:36.248 a very different kind of of challenge than replicating 00:14:36.248 --> 00:14:39.408 a picture you took from a journal of neuroscience or from 00:14:39.408 --> 00:14:42.148 nature neuroscience in med lab it's not building a 00:14:42.148 --> 00:14:44.988 model yeah right so in this sense i'm we have done it 00:14:44.988 --> 00:14:47.808 we've seen it i don't think we are explaining much with it 00:14:47.808 --> 00:14:50.708 yeah no i agree that i think 00:14:50.708 --> 00:14:55.348 these one-size-fits-all models don't work because you know i i just think the 00:14:55.348 --> 00:14:59.008 brain is far too complex but i think you can learn you can get insights from 00:14:59.008 --> 00:15:04.328 some of these models that make you think about the system in a way that maybe 00:15:04.328 --> 00:15:09.108 you weren't attending enough to certain kinds of things. 00:15:09.388 --> 00:15:14.828 For example, you may have been postulating more complexity in the assumed wiring 00:15:14.828 --> 00:15:17.368 than is necessary to get the complex wiring. 00:15:18.585 --> 00:15:21.565 Kind of behaviors that you see like cell types with extremely 00:15:21.565 --> 00:15:24.965 specific properties for example and um and 00:15:24.965 --> 00:15:27.685 and so i think it just makes you think about things in a different 00:15:27.685 --> 00:15:33.005 way as a physiologist you for example i've started to think maybe maybe our 00:15:33.005 --> 00:15:39.845 parcellation of cells into these categorical um you know categories like like 00:15:39.845 --> 00:15:45.225 grid cells and play cells maybe we're being constrained by our predetermined modular model. 00:15:45.485 --> 00:15:49.445 And in fact, if we really approached it from a slightly different perspective, 00:15:49.565 --> 00:15:52.845 we'd see that there's a continuum of response types in this area. 00:15:53.005 --> 00:15:56.425 And when you look in entorhinal cortex, indeed, it's true there are grid cells 00:15:56.425 --> 00:15:58.765 and there are head direction cells and there are border cells, 00:15:58.885 --> 00:16:00.345 but there's everything in between as well. 00:16:00.565 --> 00:16:06.345 And so we may have been over-egging the modular side of things. 00:16:06.545 --> 00:16:09.905 Yeah, I think the point is that the interaction goes both ways. 00:16:09.905 --> 00:16:19.105 So the biologists learn from the people who are developing models and also AI 00:16:19.105 --> 00:16:21.765 systems about the power of learning. 00:16:21.925 --> 00:16:27.045 But also self-organization to restructure networks. 00:16:27.045 --> 00:16:30.245 Networks um and then actually there's 00:16:30.245 --> 00:16:33.525 a lot coming back the other way so that the um without 00:16:33.525 --> 00:16:36.305 necessarily acknowledging where it came from people are 00:16:36.305 --> 00:16:39.685 using ideas from the brain to develop these these new 00:16:39.685 --> 00:16:43.045 intelligent self-organizing systems so um i 00:16:43.045 --> 00:16:46.525 think we can all agree that that when 00:16:46.525 --> 00:16:49.305 we look at deep learning systems as they are now they're not very 00:16:49.305 --> 00:16:52.565 brain-like but what i think is has 00:16:52.565 --> 00:16:55.345 been shown in the last decade say is 00:16:55.345 --> 00:16:58.105 that with sufficiently powerful computers and the 00:16:58.105 --> 00:17:01.145 brain is a very powerful computer some simple principles 00:17:01.145 --> 00:17:04.305 can give you really powerful performance and 00:17:04.305 --> 00:17:07.945 that's now coming through in the world of technology so 00:17:07.945 --> 00:17:10.845 i'm just saying tony that we already knew that 00:17:10.845 --> 00:17:13.425 well you know i'll start at old house in the 00:17:13.425 --> 00:17:16.625 fields as a huge tradition look this 00:17:16.625 --> 00:17:20.645 is the the whole problem right we're always riding this wave of amnesia like oh 00:17:20.645 --> 00:17:23.445 i rediscovered something people i wrote up five years ago but 00:17:23.445 --> 00:17:26.425 but the time constant of collective memory in the field is just so short 00:17:26.425 --> 00:17:29.505 people all forgot about it but it's not just that i think that's deeply annoying 00:17:29.505 --> 00:17:33.665 people dismissed it because they said well these things are limited in terms 00:17:33.665 --> 00:17:37.365 of what they can do no they're not giving us the solutions we want let's go 00:17:37.365 --> 00:17:41.845 and look elsewhere and then you see 20 years later the computers just get more 00:17:41.845 --> 00:17:46.325 powerful and these things are delivering some of the things that Well, 00:17:46.365 --> 00:17:48.485 there's a dirty trick, though, that people don't talk about. 00:17:48.725 --> 00:17:50.305 The dirty trick is it's all supervised. 00:17:51.245 --> 00:17:53.245 It's all supervised learning. The brain doesn't have this luxury. 00:17:53.445 --> 00:17:56.065 There is no supervisor in the brain telling, oh, wait, no, no, 00:17:56.125 --> 00:17:59.585 this grid cell response is actually wrong. No. 00:18:00.645 --> 00:18:06.665 So, sure, great if you have the superpower of a supervisor who knows everything, 00:18:06.805 --> 00:18:09.085 like God is training you to be a world champion. 00:18:09.905 --> 00:18:13.365 Fabulous. But that's just not the luxury that the red brain has. 00:18:13.545 --> 00:18:17.905 So, this is my whole point about constraints. So, we must satisfy the pertinent 00:18:17.905 --> 00:18:20.945 constraints. And that's the discussion we have to have between the theatricians 00:18:20.945 --> 00:18:24.745 and the biologists is not like, oh, let me overwhelm you with my mathematosis. 00:18:24.945 --> 00:18:29.405 No, it's about what are the specific constraints that I'm satisfying that you 00:18:29.405 --> 00:18:33.165 have identified as experimentalist and what are the testable predictions I'm giving back to you? 00:18:33.425 --> 00:18:36.105 This is a dialogue that we need to establish. And that's not happening. 00:18:36.505 --> 00:18:40.445 People make a lot of noise about how fantastic this all is. 00:18:40.585 --> 00:18:44.965 And they can deliver even more cat videos to your doorstep with deep learning. 00:18:45.145 --> 00:18:49.065 And it's deeply uninteresting. I think that's a good point to go back into your 00:18:49.065 --> 00:18:53.465 talk because actually you were identifying some constraints because this whole 00:18:53.465 --> 00:19:00.465 question around how does the rat brain actually generate these different elements 00:19:00.465 --> 00:19:02.425 of a system that can map space. 00:19:02.425 --> 00:19:05.865 And there's this interesting question as to whether, for instance, 00:19:06.085 --> 00:19:09.545 they really have a full 3D map, even if they live in a 3D world. 00:19:09.665 --> 00:19:14.605 And you began by talking about the problem of just encoding head direction when 00:19:14.605 --> 00:19:17.605 you're moving on something that's not just a flat plane. 00:19:17.845 --> 00:19:22.985 And there I think you were also arguing that this isn't a system that can represent 00:19:22.985 --> 00:19:26.025 orientation sort of in a universal way. 00:19:26.125 --> 00:19:30.745 It's very much grounded in the ecology of the rat and the kind of life that it has. 00:19:31.245 --> 00:19:34.705 Yes, well we're still collecting data on this, it's early days, but certainly. 00:19:36.293 --> 00:19:39.293 What we're seeing in the grid cells and the head direction cells are sort of 00:19:39.293 --> 00:19:44.773 hints that the system is preferring to just make a flat map of local space. 00:19:45.353 --> 00:19:49.753 And that makes a lot of sense from an evolutionary perspective, 00:19:49.853 --> 00:19:54.573 because supporting a brain is enormously expensive in terms of energy. 00:19:55.033 --> 00:19:59.813 And to make a 3D model of the world, to make a full 3D model of the world, 00:19:59.853 --> 00:20:03.613 you need vastly more neural resources than to just make a flat model of the world. 00:20:04.353 --> 00:20:08.293 Of course, a flat model has a lot of limitations, and I did talk about some 00:20:08.293 --> 00:20:12.133 of those. You can only get so far with a flat map when you're on a hilly surface. 00:20:12.293 --> 00:20:17.573 It would be not very good for planning optimal routes across hilly terrain, for example. 00:20:18.893 --> 00:20:22.713 But if you take a flat map and then inject a little bit of three-dimensional 00:20:22.713 --> 00:20:27.433 information into it, then you might have something that works pretty well for 00:20:27.433 --> 00:20:28.753 all practical purposes. 00:20:28.753 --> 00:20:32.053 And that's really what the problem that RAT has to solve is how to practically 00:20:32.053 --> 00:20:35.433 get by in the world with the minimal expenditure of energy. 00:20:35.893 --> 00:20:41.033 Yeah, I think you distinguish different ways in which you could think about three dimensions. 00:20:41.193 --> 00:20:44.893 You could say, well, I'm interested in surface structure, so if it's undulations, 00:20:45.033 --> 00:20:47.213 I care about whether I'm going up and down. 00:20:47.213 --> 00:20:51.193 And then you talked about if I'm in the ocean or in the air, 00:20:51.253 --> 00:20:54.093 I'm interested maybe where I'm in a 3D volume, 00:20:54.233 --> 00:21:02.353 which would then, as you say, require more space to encode all of the information about that volume. 00:21:02.493 --> 00:21:06.333 And then you talked about some mixture models. And I think where you're going 00:21:06.333 --> 00:21:09.013 in your work is towards more of a mixture model. Is that right? 00:21:09.293 --> 00:21:12.953 Yes, I think so. So we call it a multi-planar model. 00:21:12.953 --> 00:21:17.753 I don't know if that's the best word for it, but the intuition is that the map 00:21:17.753 --> 00:21:21.393 of three-dimensional space, at least, I should qualify it and say that we're 00:21:21.393 --> 00:21:25.373 studying surface-dwelling animals like rats and mice and probably humans. 00:21:26.553 --> 00:21:34.733 So the map of a complex three-dimensionally, topologically undulating world is a lot of, 00:21:35.933 --> 00:21:40.633 we think of it as like mosaic fragments, each local one of which is two-dimensional, 00:21:40.773 --> 00:21:43.533 but which are related to each other with some three-dimensional structure. 00:21:45.093 --> 00:21:49.513 And we are in the process of collecting data about how that might work and for 00:21:49.513 --> 00:21:54.433 example how the head direction system could cope with a system like that and 00:21:54.433 --> 00:21:56.633 how it would avoid accumulating errors and so on. 00:21:57.313 --> 00:22:02.253 The acid test is really going to be what happens when we get rats to move through a volumetric space. 00:22:02.533 --> 00:22:06.073 And we've started these experiments in my laboratory at the moment where we're 00:22:06.073 --> 00:22:10.873 training rats to move through a lattice that's conceptually like moving through 00:22:10.873 --> 00:22:13.373 the branches of trees in a forest or something like that. 00:22:13.433 --> 00:22:18.653 So where the rat really can move in all three dimensions. and then we'll have 00:22:18.653 --> 00:22:23.773 to see what the grid cells particularly do in that situation right so but then um, 00:22:24.590 --> 00:22:28.290 If we now look at this situation from also this perspective of modularity, 00:22:28.410 --> 00:22:33.030 in some sense you can also say that head direction cells and grid cells have also great redundancy. 00:22:33.470 --> 00:22:37.470 Because in some sense, the grid cell gives you a spatial representation of the 00:22:37.470 --> 00:22:39.370 temporal signal you get from your head direction cells. 00:22:39.770 --> 00:22:43.210 It's the velocity vector that is driving the grid cells, right? 00:22:43.330 --> 00:22:45.030 So there's really redundancy there. 00:22:45.430 --> 00:22:49.830 So would you find cells that are, let's say, partially head direction cell and partially grid cell? 00:22:49.890 --> 00:22:54.510 Would you have mixtures? Well, it's certainly the case that there are so-called 00:22:54.510 --> 00:22:59.730 conjunctive cells, like grid cells that only produce their grids when the rat's 00:22:59.730 --> 00:23:01.450 facing in a particular direction. 00:23:02.730 --> 00:23:09.450 So I don't know if that was the kind of cell you meant, but they certainly have been well reported. 00:23:09.450 --> 00:23:15.630 But the other interesting thing is that if you deprive ordinary multidirectional 00:23:15.630 --> 00:23:22.190 grid cells of their place cell input, then what's left seems to be a head direction signal. 00:23:22.650 --> 00:23:26.970 So this is work from the MOSA lab where they inactivated the hippocampus. 00:23:27.010 --> 00:23:30.470 And so it looks like the head direction signal may be a sort of a fundamental 00:23:30.470 --> 00:23:34.290 input to grid cells, and that helps the grids to become oriented. 00:23:34.290 --> 00:23:39.990 But I'm not sure that you could function only with well-oriented grid cells 00:23:39.990 --> 00:23:43.870 in the absence of head direction cells, because of course there's the six-fold 00:23:43.870 --> 00:23:48.390 redundancy or replication ambiguity, I suppose is the word. 00:23:48.990 --> 00:23:51.850 You wouldn't know which of the six directions you were moving in necessarily 00:23:51.850 --> 00:23:54.110 if you didn't have a proper head direction system. 00:23:54.250 --> 00:23:58.210 But now your reference model for the head direction cells to now jump to three-dimensional 00:23:58.210 --> 00:24:04.150 space is the ring attractor, right? So basically you have a bunch of neurons 00:24:04.150 --> 00:24:05.710 that are coupled together in a ring. 00:24:06.010 --> 00:24:10.230 Every neuron is an encoding, a heading direction, and activity is just moving 00:24:10.230 --> 00:24:13.810 through this ring, exploiting the fact that heading directions also continuously change. 00:24:15.090 --> 00:24:19.010 So and then the question is, okay, how could I exploit such a ring attractor 00:24:19.010 --> 00:24:22.790 that works great in a planar space for a three-dimensional space? 00:24:23.130 --> 00:24:28.770 So your idea then is that the ring becomes a sphere, so I'm covering now also 00:24:28.770 --> 00:24:30.390 all elevation values, values? 00:24:30.990 --> 00:24:38.030 Or do you have another model in mind? Well, if the brain tries to form a fully 00:24:38.030 --> 00:24:41.450 volumetric 3D map, then I think you would want something, 00:24:41.790 --> 00:24:45.850 by extension from what we know in two dimensions, I think you would want a fully 00:24:45.850 --> 00:24:48.830 three-dimensional compass, which is to say a spherical attractor. 00:24:49.536 --> 00:24:55.876 Which also, unlike in two dimensions, also has to take into account the orientation 00:24:55.876 --> 00:25:00.116 of the body of the animal, because of course in three dimensions it can roll 00:25:00.116 --> 00:25:03.116 around the long axis of its body. 00:25:04.636 --> 00:25:07.736 So, theoretically speaking, I think that's what one would need. 00:25:07.736 --> 00:25:11.076 Now, we've not tried to model a spherical attractor, 00:25:11.096 --> 00:25:17.036 but it seems to me that it would be vastly more complicated than a ring attractor 00:25:17.036 --> 00:25:20.876 because of the problems that I talked about with things like the non-commutativity 00:25:20.876 --> 00:25:23.336 of rotations and all of these things. 00:25:23.416 --> 00:25:29.196 The problem of extracting azimuth from your rotations when the animal is not 00:25:29.196 --> 00:25:32.236 actually in a horizontal plane and all of these things that make it very complicated. 00:25:32.436 --> 00:25:35.816 So if we're thinking that this is a system that's trying to be economical, 00:25:36.056 --> 00:25:39.136 a more economical solution would be to just stick with your ring attractor 00:25:39.136 --> 00:25:41.796 and just to have it work on whatever plane you happen to be 00:25:41.796 --> 00:25:45.356 on regardless of its orientation and then figure 00:25:45.356 --> 00:25:51.276 out a way to map that signal back onto the horizontal and i i suspect that's 00:25:51.276 --> 00:25:56.076 what's happening with the so but an alternative might be that i just take my 00:25:56.076 --> 00:26:00.216 cardinal axis of movement and have single rings for those and maybe i take two 00:26:00.216 --> 00:26:04.496 populations that are tuned to this cardinal axis of movement and I have a little offset between them. 00:26:04.696 --> 00:26:08.816 And then via interpolation, I could actually extract all my possible heading 00:26:08.816 --> 00:26:10.016 directions in three dimensions. 00:26:11.096 --> 00:26:15.916 So why did you not consider this more, let's say, simplified version? 00:26:16.336 --> 00:26:20.156 Yeah, it's a possibility too. You mean like to have three orthogonal ring attractors? 00:26:20.196 --> 00:26:21.296 Yeah, exactly. Yeah, that would be a possibility. 00:26:21.516 --> 00:26:23.616 And Cynthia Moss has shown that 00:26:23.616 --> 00:26:29.216 that would account quite well to make a three-dimensional compass signal. 00:26:29.356 --> 00:26:33.996 I think that's entirely possible. It doesn't fully solve the problem because 00:26:33.996 --> 00:26:41.956 if you're on a plane that's not orthogonal to any of those three ring attractors, 00:26:42.036 --> 00:26:45.356 then you need to kind of map the yaw rotation that you're making onto one of those. 00:26:45.436 --> 00:26:49.436 So you've still got a transformation that's modulated by the angle between the 00:26:49.436 --> 00:26:53.616 surface that you're on and the three now ring attractors. 00:26:53.656 --> 00:26:55.796 So it's still not an entirely simple problem. 00:26:55.936 --> 00:26:59.616 On the other hand, the brain is quite good at solving non-simple problems. 00:26:59.636 --> 00:27:01.136 So I wouldn't rule it out at all. 00:27:01.456 --> 00:27:06.896 But now the intuition would be that we, so let's say, whether it's just a set of rings or a sphere, 00:27:07.096 --> 00:27:11.236 let's say we have now a 3D heading direction system, and then the intuition 00:27:11.236 --> 00:27:14.676 would be, okay, if these guys now drive my grid cells, I would have 3D grid 00:27:14.676 --> 00:27:16.656 cells, right? This is roughly the idea. 00:27:18.456 --> 00:27:20.776 But this is not exactly what you found. Right. 00:27:22.034 --> 00:27:26.474 Brad, you didn't literally find grid cells that are tuned in 3D. 00:27:27.014 --> 00:27:31.034 It was a bit more complicated than that. It's a bit more complicated than that, 00:27:31.054 --> 00:27:32.894 but we haven't done the acid test. 00:27:33.194 --> 00:27:36.794 So that's the experiment I mentioned a moment ago where the animal really can 00:27:36.794 --> 00:27:38.574 move freely in all three dimensions. 00:27:39.094 --> 00:27:42.274 So the experiments that we've done with grid cells in three dimensions have 00:27:42.274 --> 00:27:46.054 been with the animal on a surface that extends into the vertical dimension. 00:27:46.954 --> 00:27:52.594 Mentioned. And we don't know how that surface is constraining the signal and 00:27:52.594 --> 00:27:58.774 how it's producing a signal that might not be there if the rat was really able to move freely. 00:27:59.314 --> 00:28:05.114 So what we see is on a vertical surface, the pattern depends on the orientation of the animal's body. 00:28:05.254 --> 00:28:10.354 So if the animal is oriented horizontally, as it is on the pegboard where it's 00:28:10.354 --> 00:28:12.874 standing on pegs that stick out of a wall. 00:28:13.134 --> 00:28:19.714 So it's oriented horizontally, but it's moving up and down as well as in horizontal dimension. 00:28:20.134 --> 00:28:24.814 There we see that the grid cells don't seem to map out distances in the vertical dimension. 00:28:25.534 --> 00:28:29.694 So they're not, in that case, mapping out distances in the direction orthogonal 00:28:29.694 --> 00:28:32.334 to the plane of the body of the animal. 00:28:33.074 --> 00:28:36.814 If, on the other hand, the plane of the body of the animal is parallel to the 00:28:36.814 --> 00:28:41.514 wall, so the rat's actually walking around on the wall and climbing around on chicken wire. 00:28:42.094 --> 00:28:46.174 Now we see something that looks more like a grid. So it looks like there's some 00:28:46.174 --> 00:28:48.254 attempt by the system to perform odometry. 00:28:49.154 --> 00:28:54.274 So the relationship of the animal to the surface is really important, 00:28:54.534 --> 00:28:56.594 at least in these particular environments. 00:28:56.754 --> 00:29:00.634 Now whether that generalizes to a fully volumetric environment, we don't know. 00:29:01.174 --> 00:29:04.834 There was something interesting about the pack board result. 00:29:05.154 --> 00:29:08.354 So you have this wall with these little sticks sticking out, 00:29:08.354 --> 00:29:12.214 and the animal can make horizontal trajectories essentially across this wall. 00:29:14.527 --> 00:29:18.647 Essentially, what you would sort of see there is sort of a strip-like organization 00:29:18.647 --> 00:29:21.347 of the grid cell response, right? 00:29:21.767 --> 00:29:25.367 So does it imply that in that case, 00:29:25.567 --> 00:29:32.527 the medial entorhinal cortex is cutting through the three-dimensional plane 00:29:32.527 --> 00:29:36.967 a bunch of horizontal planes and saying, okay, actually, this is like a number 00:29:36.967 --> 00:29:38.487 of alleys that I'm running through, 00:29:38.607 --> 00:29:42.207 and I just ignore then the third-dimension part of it. 00:29:42.207 --> 00:29:45.367 I just map out every local horizontal stretch. 00:29:45.707 --> 00:29:50.747 Yes, I think that's one interpretation. That seems to be the likeliest interpretation. 00:29:52.387 --> 00:29:57.747 As to what it is about the dimension that the grid cell is not mapping out, 00:29:57.807 --> 00:29:58.707 so in the vertical dimension, 00:29:59.687 --> 00:30:03.227 I don't know, because we haven't done all the various control experiments, 00:30:03.487 --> 00:30:08.047 I don't know whether the system just doesn't like to do odometry that's not 00:30:08.047 --> 00:30:10.687 in the direction that the rat's running, which is a possibility, 00:30:10.687 --> 00:30:16.347 or if it's something specific to the direction that's orthogonal to the plane of the animal. 00:30:16.447 --> 00:30:20.347 So there are various experiments we need to do to distinguish between those possibilities. 00:30:20.507 --> 00:30:24.587 It may turn out the grid cells, they really only are interested in counting 00:30:24.587 --> 00:30:27.447 footsteps in the direction that the animal's running or something quite simple. 00:30:28.507 --> 00:30:30.187 Modulated, of course, by a running direction. 00:30:30.887 --> 00:30:35.587 But if the rat were to run sideways, we might also see that grid cell odometry fails. 00:30:35.687 --> 00:30:38.647 So we haven't tried that yet, nor backwards. 00:30:38.647 --> 00:30:44.407 I think going back to your suggestion that the grid cells are maybe being driven 00:30:44.407 --> 00:30:45.667 by the head direction cells, 00:30:46.027 --> 00:30:52.127 these results maybe make sense in the light of the study that you told us about 00:30:52.127 --> 00:30:57.947 where a rat is climbing, I think, on chicken wire around a sort of square pillar. 00:30:58.487 --> 00:31:04.547 And you're looking at the head direction cells when the rat is on either side 00:31:04.547 --> 00:31:09.607 of the pillar and how those changes it moves around the pillar. And you are saying. 00:31:10.638 --> 00:31:18.578 That it wasn't changing in a way consistent with having a full 3D compass, 00:31:19.478 --> 00:31:23.718 and it wasn't changing in a way consistent with having an entirely local compass. 00:31:23.878 --> 00:31:28.518 It was something that you called, I think, it was either locally global or globally local. 00:31:29.138 --> 00:31:33.158 Can you explain what you mean by that? Well, so what we found, 00:31:33.358 --> 00:31:38.998 which is very similar to results from Jeff Tauby a few years ago, show. 00:31:39.358 --> 00:31:45.598 But what we've done is show that this is an active modulation process rather than a passive one. 00:31:45.838 --> 00:31:50.478 So what Taobi's group showed is that if a rat walks from a horizontal surface 00:31:50.478 --> 00:31:55.898 to a vertical surface, then the head direction cells essentially remain unchanged. 00:31:56.598 --> 00:32:00.098 So as the animal walks towards the wall, the head direction cell that's active 00:32:00.098 --> 00:32:04.398 while it's on the floor will continue to be active as the rat walks up onto 00:32:04.398 --> 00:32:05.678 the wall and is now facing upwards. 00:32:06.398 --> 00:32:10.498 And then when the rat does yaw rotations on the wall, then the activity moves 00:32:10.498 --> 00:32:13.458 around the ring of head direction cells in the usual way. 00:32:14.078 --> 00:32:19.238 So what we have done is take that experiment a step further and shown that when 00:32:19.238 --> 00:32:24.098 the animal moves from one vertical surface to another differently oriented vertical surface, 00:32:24.378 --> 00:32:30.298 then the head direction cells actively rotate their signal by 90 degrees as 00:32:30.298 --> 00:32:32.418 the rat goes around a 90-degree corner. 00:32:34.178 --> 00:32:35.458 So that the... 00:32:38.131 --> 00:32:44.191 The representation is still essentially related to the representation that would 00:32:44.191 --> 00:32:49.631 be on the horizontal surface, but it's been updated by a movement that wasn't a yaw rotation. 00:32:50.811 --> 00:32:54.891 So in other words, non-yaw rotations around the vertical axis can update the 00:32:54.891 --> 00:32:57.311 head direction signal when the rat's not on a horizontal surface. 00:32:57.311 --> 00:33:00.031 Surface and the the consequence of that 00:33:00.031 --> 00:33:02.851 is that when the rat goes back down 00:33:02.851 --> 00:33:05.731 onto a horizontal surface then the signal has been appropriately 00:33:05.731 --> 00:33:09.131 updated such that it's consistent and still pointing 00:33:09.131 --> 00:33:12.011 the correct way so so what 00:33:12.011 --> 00:33:14.831 we have is a very simple rule basically for updating the head 00:33:14.831 --> 00:33:18.751 direction signal as the rat goes around vertical corners so what 00:33:18.751 --> 00:33:22.311 the rat really cares about is uh knowing 00:33:22.311 --> 00:33:25.331 where it is in the horizontal plane is that right 00:33:25.331 --> 00:33:28.111 i i would say that's a fair interpretation that's certainly 00:33:28.111 --> 00:33:30.951 our working hypothesis and that's the 00:33:30.951 --> 00:33:35.071 main thing if that then is the signal driving the grid cells then does 00:33:35.071 --> 00:33:37.791 that perhaps explain some of the grid cell results because now the head 00:33:37.791 --> 00:33:40.851 direction cells don't care so much about where you 00:33:40.851 --> 00:33:44.511 are vertically but they care a lot about where you are horizontally so inevitably 00:33:44.511 --> 00:33:49.931 the grid cells are going to be coding much more strongly for the horizontal 00:33:49.931 --> 00:33:55.091 uh dimensions well so the the head direction cells I'm not sure we could say 00:33:55.091 --> 00:33:57.831 that they don't care about the vertical encoding because, 00:33:57.931 --> 00:34:01.671 in fact, the specificity of the signal on the vertical wall is just as good 00:34:01.671 --> 00:34:02.811 as it is on the horizontal. 00:34:05.171 --> 00:34:09.811 Maybe it's not vertical versus horizontal, but it's something about the surface 00:34:09.811 --> 00:34:14.771 that you're on, but then how that surface is relative to the true horizontal, 00:34:14.811 --> 00:34:15.831 which you know through gravity. 00:34:16.131 --> 00:34:19.531 So those are the things that the head cells care about, is that right? 00:34:19.891 --> 00:34:23.691 Well, yes. Yes, I think that ultimately the consistency that the system is trying 00:34:23.691 --> 00:34:28.051 to maintain is consistency in how they encode azimuth. 00:34:28.191 --> 00:34:31.711 So in other words, horizontal direction, compass direction, essentially. 00:34:31.911 --> 00:34:38.571 So I think that these rules about updating the signal for rotations around the 00:34:38.571 --> 00:34:44.511 vertical axis, I think the function of those rules is to maintain horizontal consistency. 00:34:44.831 --> 00:34:48.851 Now how that maps to what the grid cells are doing, it's not quite so clear, 00:34:48.851 --> 00:34:53.751 because the head direction signal is very much the same on the vertical wall 00:34:53.751 --> 00:34:55.751 as it is on the horizontal floor. 00:34:55.991 --> 00:34:59.291 But the grid cell signal seems to be quite different. So the scale is expanded. 00:34:59.991 --> 00:35:04.871 The space between fields is relatively much larger than it should be. 00:35:05.731 --> 00:35:12.311 And something has changed about the oscillatory activity, the theta rhythm and so on. 00:35:12.571 --> 00:35:15.631 Something is different about how the grid cells are computing distances on the 00:35:15.631 --> 00:35:18.291 wall. and yet the head direction cells are just doing what they, 00:35:19.064 --> 00:35:22.384 should do. So there's a slight dissociation there, which we haven't yet understood. 00:35:22.384 --> 00:35:23.864 But isn't that Tony's point? 00:35:23.964 --> 00:35:27.724 Because I think the point you're making is that if I'm on this vertical wall, 00:35:27.904 --> 00:35:32.404 I'm still mapping my head direction cell back to a horizontal plane. 00:35:32.824 --> 00:35:37.844 So now I get also sort of rounding errors and I get imprecisions in that mapping 00:35:37.844 --> 00:35:41.824 because my head direction cell is still believing we're moving around on this 00:35:41.824 --> 00:35:44.124 horizontal plane, but actually it's a vertical plane. 00:35:44.404 --> 00:35:48.204 So that means if this is now the key driving input of my grid cells, 00:35:48.444 --> 00:35:49.864 this also will get distorted. 00:35:50.404 --> 00:35:56.644 And that might lead to sort of a collapse or a remapping of the grid cell response. 00:35:57.444 --> 00:35:59.624 I think this is what you had in mind, Tony, roughly. 00:36:01.104 --> 00:36:06.204 That wasn't quite my idea, but my thought was that if you had a perfect 3D compass, 00:36:06.504 --> 00:36:10.284 the grid cells might automatically develop a nice volumetric mapping. 00:36:10.744 --> 00:36:14.604 But I don't know how we would test that. Okay, it's the other extreme then, okay. 00:36:15.804 --> 00:36:19.544 Because for instance, Kate, you showed this really beautiful experiment where 00:36:19.544 --> 00:36:24.004 you had these animals crawling around this cube on the chicken wire. 00:36:24.244 --> 00:36:28.304 And then what you showed is that as the animal turns the corner in these two 00:36:28.304 --> 00:36:30.984 vertical planes, goes from one vertical plane to the other vertical plane, 00:36:31.124 --> 00:36:34.944 there's a very rapid shift of the heading direction response with 90 degrees. 00:36:35.584 --> 00:36:38.964 So you would expect that if you then look in that condition to the grid cells, 00:36:39.244 --> 00:36:41.884 that there should be some massive change in the response of the grid cells. 00:36:41.884 --> 00:36:50.144 Well, we might predict that all other things being equal, that the grid pattern on the. 00:36:51.689 --> 00:36:55.569 The two vertical walls would be 90-degree rotations of each other. 00:36:56.929 --> 00:37:00.309 But there are all sorts of qualifications to that. One is that, 00:37:00.409 --> 00:37:05.729 as I've mentioned, the pattern of the grid cells on the vertical wall is so 00:37:05.729 --> 00:37:09.209 different that it's not even clear that it's a hexagonal close-packed array. 00:37:09.369 --> 00:37:14.409 So we don't even know that we could determine what the orientation was of the 00:37:14.409 --> 00:37:19.909 grids. The other thing is that we've shown that the grid cells are somewhat 00:37:19.909 --> 00:37:21.509 sensitive to context information. 00:37:21.749 --> 00:37:26.329 And of course, the east wall and the south wall could, to the system, 00:37:26.489 --> 00:37:27.669 seem like different contexts. 00:37:27.749 --> 00:37:30.449 So maybe the grid cells would just do something entirely different. 00:37:31.929 --> 00:37:36.009 So it's a little difficult to predict what we would see. And I'm not sure how 00:37:36.009 --> 00:37:37.849 easy it would be to interpret what we saw. 00:37:39.469 --> 00:37:44.089 So right now, we're really analyzing this three-dimensional representation of space. 00:37:44.609 --> 00:37:50.449 Or the representation of three-dimensional space from the heading direction system perspective. 00:37:51.369 --> 00:37:55.069 So the heading direction response is, again, very much predicated on what your 00:37:55.069 --> 00:37:56.409 vestibular system will tell you. 00:37:56.789 --> 00:38:01.649 So it's maybe this inability to really map out three-dimensional space accurately 00:38:01.649 --> 00:38:07.609 in the rodent dependent on just getting signals that are less reliable or more 00:38:07.609 --> 00:38:11.669 noisy or less precise in the vertical plane as opposed to the horizontal plane. 00:38:11.669 --> 00:38:15.469 So it's just a matter of the sensory front end, the sensory apparatus, 00:38:15.769 --> 00:38:20.189 not providing you with the information to actually have an accurate head direction 00:38:20.189 --> 00:38:21.429 response in the third dimension. 00:38:22.749 --> 00:38:26.969 Well, the vestibular system is pretty good at providing information. 00:38:27.809 --> 00:38:34.629 So, you know, it's sensitive to angular and linear information in the various different directions. 00:38:34.629 --> 00:38:43.649 But the hypothesis that we're toying with at the moment is that the vestibular 00:38:43.649 --> 00:38:49.609 signals that are feeding into the grid cell system that normally work on the horizontal plane, 00:38:49.769 --> 00:38:52.569 they comprise all of the semicircular canal information, 00:38:52.789 --> 00:38:54.789 so all of the rotational information, 00:38:55.109 --> 00:38:58.169 together with linear information from the otolith organ. 00:38:59.349 --> 00:39:02.609 So our kind of working hypothesis is that on the vertical surface, 00:39:02.889 --> 00:39:07.169 now the otolith organ, which is sensitive to acceleration in the horizontal 00:39:07.169 --> 00:39:11.149 plane and normally has the gravity vector orthogonal to that, 00:39:11.649 --> 00:39:18.569 is now in a different state of alignment and that perhaps the system copes with that. 00:39:18.769 --> 00:39:23.289 And rather than developing a whole new way of processing the signal, 00:39:23.329 --> 00:39:25.129 just says, let's just do without the otolith signal. 00:39:25.329 --> 00:39:29.749 So we'll just work with the semi-circular canals and let's forget the whole 00:39:29.749 --> 00:39:31.069 linear acceleration thing. 00:39:31.189 --> 00:39:35.469 So that may be why the grids are expanded on the wall because they're missing 00:39:35.469 --> 00:39:37.549 one of their vestibular inputs. Right, that would give quite a bias, 00:39:37.649 --> 00:39:42.309 right, because if we now go to these annoying bats that show actually a three-dimensional 00:39:42.309 --> 00:39:45.849 representation in their place cell system and they're. 00:39:47.602 --> 00:39:55.402 Do they have an oscillate type canal also running in the orthogonal axis, in the vertical axis? 00:39:55.762 --> 00:39:58.302 Is that what helps them to develop a three-dimensional representation? 00:39:58.882 --> 00:40:03.462 I think, as far as I know, the vestibular system is pretty similar. Okay. 00:40:04.102 --> 00:40:09.022 And we don't fully know the details about the three-dimensional nature of their representation. 00:40:09.462 --> 00:40:14.002 So Nakamulanovsky and his group have done some really beautiful work showing 00:40:14.002 --> 00:40:18.042 that place cells seem to form place fields that pack of volume, 00:40:18.282 --> 00:40:20.922 as you would predict for a three-dimensional map. 00:40:21.722 --> 00:40:28.322 The head direction cells are sensitive to all three of the directions, but not equally. 00:40:28.542 --> 00:40:33.082 So there are many more azimuth-sensitive cells than there are cells sensitive 00:40:33.082 --> 00:40:37.742 to pitch, and there are very few cells sensitive to roll. 00:40:37.902 --> 00:40:42.802 There are a small handful that are sensitive to combinations of all three of 00:40:42.802 --> 00:40:47.582 those angles, so they are true 3D compass cells but there are very few of them. 00:40:48.042 --> 00:40:52.502 So I don't think that even the bat which moves a lot through three-dimensional 00:40:52.502 --> 00:40:58.362 space really has a true volumetric compass that works evenly in all three of the dimensions. 00:40:58.542 --> 00:41:02.202 I think it's still biased towards encoding the horizontal plane. 00:41:02.742 --> 00:41:06.202 The story for grid cells and bats we're waiting with bated breath to see what 00:41:06.202 --> 00:41:10.562 happens there and I think that'll raise some very interesting questions. 00:41:10.802 --> 00:41:16.442 I'm looking forward to those data coming. But then if we take a rat and we sort 00:41:16.442 --> 00:41:21.922 of glue it to a drone and we have the rat fly around in the lab to get its food. 00:41:22.822 --> 00:41:29.042 Would you predict that it would develop three-dimensionally tuned play cells as the bat? 00:41:30.182 --> 00:41:34.642 So you're asking a question about experience and is experience enough to create 00:41:34.642 --> 00:41:38.822 a 3D bat? Right, because apparently the periphery is, as far as we know, rather comparable. 00:41:39.862 --> 00:41:45.042 Yes, yes. So I guess the answer would have to be, I don't know. 00:41:45.662 --> 00:41:51.482 No, we want the prediction. Well, so we have been raising rats in a fairly three-dimensional 00:41:51.482 --> 00:41:56.802 environment so as to have subjects that are as 3D competent as we possibly can. 00:41:56.922 --> 00:42:01.082 So they spend all of their time climbing through climbing frames and up and 00:42:01.082 --> 00:42:05.422 down things, and they're pretty competent. They don't fly, and we've not put 00:42:05.422 --> 00:42:09.922 them on a drone, but in all other respects, they're pretty good at this volumetric. 00:42:10.774 --> 00:42:15.774 Navigation. But we don't see any difference in the encoding of their neurons. 00:42:16.014 --> 00:42:20.134 So if I had to guess, I would say that I don't think experience is going to 00:42:20.134 --> 00:42:22.074 create a 3D map if there isn't one there already. 00:42:22.074 --> 00:42:29.534 But now we know that maybe in the rats that you're allowing to develop in a 00:42:29.534 --> 00:42:33.014 3D spatial space in the lab, they could still get away with just slicing it 00:42:33.014 --> 00:42:36.254 up in many horizontal planes in which they operate, no? 00:42:36.434 --> 00:42:41.114 Right, because they move around on surfaces. Exactly. So if I glued a rat to 00:42:41.114 --> 00:42:45.394 another bat or to a drone, I have linear motion, but then really in three-dimensional 00:42:45.394 --> 00:42:47.914 space. Yeah, I don't know. 00:42:47.974 --> 00:42:51.954 I think we'd have to do it. I mean, there are hand-waving arguments for why 00:42:51.954 --> 00:42:52.994 they might be able to do that. 00:42:53.054 --> 00:42:57.674 One of them being that we all evolved from fully 3D-competent animals, 00:42:57.914 --> 00:43:01.334 so fish, which move around in a volumetric space. 00:43:01.334 --> 00:43:07.354 And it's quite possible that all of this evolved eons ago and that it's become 00:43:07.354 --> 00:43:09.694 somewhat vestigial, or at least we don't use it. 00:43:09.874 --> 00:43:15.074 Surface-dwelling animals like us and rats and mice and things don't use it, 00:43:15.094 --> 00:43:17.274 but it's still there. So that's possible. 00:43:18.734 --> 00:43:24.294 In support of that, there's research on astronauts, or observations on astronauts, 00:43:24.454 --> 00:43:26.434 not formal studies as far as I know. 00:43:26.434 --> 00:43:30.314 No, but observations of astronauts find that in the first few days of weightlessness, 00:43:30.614 --> 00:43:37.054 they tend to try and form a flat map, if you like, where they reference their 00:43:37.054 --> 00:43:40.994 knowledge of the layout of the environment to a notional floor. 00:43:42.102 --> 00:43:45.822 So they define a nearby surface as the floor and then the surface opposite it as the ceiling. 00:43:46.022 --> 00:43:50.102 And then if they drift across that space in their weightlessness and find themselves 00:43:50.102 --> 00:43:53.002 too close to the ceiling, then they suddenly reorient their sense of direction 00:43:53.002 --> 00:43:54.542 and now the ceiling becomes the floor. 00:43:55.482 --> 00:43:59.982 So that suggests to me that they didn't start out with a three-dimensional map 00:43:59.982 --> 00:44:04.342 because why would you have this reorientation if you were really quite happy 00:44:04.342 --> 00:44:09.282 floating around in a space and just able to encode X, Y and Z with equal facility? 00:44:10.062 --> 00:44:13.862 On the other hand, this disorientation abates over time, as far as I know. 00:44:14.302 --> 00:44:16.882 So it's possible that they do develop some competency. 00:44:17.722 --> 00:44:22.242 Whether it's a kind of a kludge, or it's just a kind of a hack, 00:44:22.342 --> 00:44:26.702 not really a fully three-dimensional map, but one that functions like one, 00:44:26.782 --> 00:44:30.402 I think we'd have to do probably some electrophysiology on them, 00:44:30.482 --> 00:44:31.702 and I'm not sure we'd be allowed to. 00:44:32.242 --> 00:44:34.802 But yeah, it's an open question. It's an interesting question, 00:44:34.942 --> 00:44:38.942 I think, whether we have the capacity to do that, even if we don't use it. 00:44:39.282 --> 00:44:42.922 One thing I think is quite potentially interesting is when we move to virtual 00:44:42.922 --> 00:44:47.462 reality and we become more competent in navigating virtual environments where 00:44:47.462 --> 00:44:49.362 we don't have the same constraints, 00:44:49.662 --> 00:44:54.562 could we engage a fully three-dimensional map and could we even create a four-dimensional 00:44:54.562 --> 00:44:57.242 one or more higher dimensional representations? 00:44:57.842 --> 00:45:02.542 And that I think would be interesting to play around with. Are we being taken 00:45:02.542 --> 00:45:05.602 in too much by this concept of a map? 00:45:06.062 --> 00:45:11.602 Because yeah, if I have a good map, I have a compass, I can count steps, 00:45:11.822 --> 00:45:13.882 I can know where I am in space, but now... 00:45:14.861 --> 00:45:17.461 Uh if i if i'm in 00:45:17.461 --> 00:45:20.581 the wrong place if i'm not where i'm on my map i can get lost but usually 00:45:20.581 --> 00:45:24.141 there are ways of recovering from that for instance uh 00:45:24.141 --> 00:45:27.281 i can look and see where some tall building is 00:45:27.281 --> 00:45:30.021 or if i'm you know a bird i can 00:45:30.021 --> 00:45:32.761 look at the constellation of the stars or where 00:45:32.761 --> 00:45:35.561 the sun is in the sky or i can tell from the direction 00:45:35.561 --> 00:45:38.381 the wind is blowing you know if i'm a seal maybe 00:45:38.381 --> 00:45:42.301 it's water currents you know thermal gradients uh gravity 00:45:42.301 --> 00:45:45.461 of course is always there as a queue so these systems 00:45:45.461 --> 00:45:49.241 give us queues that will help us localize 00:45:49.241 --> 00:45:52.881 ourselves or at least get something like a homing vector even when 00:45:52.881 --> 00:45:56.041 the uh the hippocampal map fails so 00:45:56.041 --> 00:46:01.441 maybe is it possible the hippocampal map is is just part of this bigger navigation 00:46:01.441 --> 00:46:05.801 system and it's one of the things that are constraining our choices about going 00:46:05.801 --> 00:46:10.021 the world but we're not tied to it in such a strong way yeah i think that's 00:46:10.021 --> 00:46:13.001 exactly right i think there are you know there's quite a lot of research to 00:46:13.081 --> 00:46:19.601 suggest that there are multiple spatial systems and what we've been calling map-based navigation, 00:46:19.841 --> 00:46:24.961 which really refers to using constellations of cues to extract distance and 00:46:24.961 --> 00:46:27.101 direction metrics for navigation. 00:46:27.381 --> 00:46:31.321 I think that's only one of several different strategies, and you've mentioned some. 00:46:31.401 --> 00:46:34.761 So beacon navigation, for example, where you just head towards the nearest tall 00:46:34.761 --> 00:46:36.061 building or whatever it was. 00:46:36.181 --> 00:46:39.661 You may not necessarily know where you are, but you can see where you need to 00:46:39.661 --> 00:46:41.481 get to, and so you just head towards it. 00:46:43.221 --> 00:46:47.141 Or remembering sequences of left-right turns that you have to make. 00:46:48.421 --> 00:46:54.561 I think a lot of our navigation is this stimulus-based, route-based navigation, 00:46:54.801 --> 00:46:57.441 where we're not really thinking about where we are within the global scheme 00:46:57.441 --> 00:47:02.621 of things and not really computing directions, but remembering patterns of behavior anchored to landmarks. 00:47:02.761 --> 00:47:05.781 When I go to work in the morning, I drive to the corner and I turn left. 00:47:05.781 --> 00:47:11.081 I don't think about the overall direction I'm going. I just know my route, you know. 00:47:11.741 --> 00:47:16.441 So, yeah, I think the brain has multiple systems, and I think it's possible 00:47:16.441 --> 00:47:18.821 to exploit that for various purposes. 00:47:18.921 --> 00:47:21.441 For example, I think it's something that robotics could usefully do, 00:47:21.541 --> 00:47:26.521 is to have these parallel multiple systems that interact. 00:47:26.681 --> 00:47:31.421 And one of the things that's emerged from animal studies is that certainly the 00:47:31.421 --> 00:47:34.741 root-based and the map-based systems seem to operate almost, 00:47:35.001 --> 00:47:37.861 I mean, not in opposition, but in an either-or fashion. 00:47:38.121 --> 00:47:40.801 So you tend to use one or the other. You switch between them. 00:47:41.281 --> 00:47:45.721 But wouldn't there be some of these other mechanisms could be involved in stitching 00:47:45.721 --> 00:47:49.461 together these sort of globally local hippocampal maps? 00:47:49.701 --> 00:47:55.121 Because what I got from your talk is that if I'm exploring on a piece of flat 00:47:55.121 --> 00:47:58.161 ground, I have a map for this. 00:47:58.201 --> 00:48:01.341 But then if I decide to run up this tree, I'm into another map. 00:48:02.061 --> 00:48:06.861 But I need then to know how these maps fit together in sort of bigger space and. 00:48:08.281 --> 00:48:12.321 It seems less likely that I have a global map, maybe at a coarser scale, 00:48:12.461 --> 00:48:17.461 but maybe I'm using other sets of cues to try and integrate between these local patches. 00:48:18.101 --> 00:48:25.741 Yeah, I think there's probably more than one way of creating a larger scale patchwork map as well. 00:48:25.781 --> 00:48:31.001 I think you've identified one, so it is quite possible that there could be these 00:48:31.001 --> 00:48:36.421 root-based ways of stitching together your behavior, if you like. 00:48:36.421 --> 00:48:39.661 As for stitching together 00:48:39.661 --> 00:48:43.021 the actual fragments of the map to make a larger map i think there 00:48:43.021 --> 00:48:46.801 you would want a brain system that had some notion 00:48:46.801 --> 00:48:53.901 about about direction because ultimately even if you're using a coarse more 00:48:53.901 --> 00:48:58.221 topological map where you've not got fine-grained distance information you still 00:48:58.221 --> 00:49:03.441 need some general idea of which direction to go and so i think it's more 00:49:03.481 --> 00:49:07.301 likely that it would be some brain system that's interacting with the head direction system. 00:49:07.481 --> 00:49:10.641 So we think that route-based navigation depends on the striatum, 00:49:10.721 --> 00:49:16.021 which is involved in controlling behavior in response to stimuli in the environment. 00:49:16.441 --> 00:49:22.021 But I think a more spatial, a more globally spatial system, more connected with 00:49:22.021 --> 00:49:25.661 the compass system, would be likely for the larger scale map. 00:49:25.901 --> 00:49:30.821 And we've been looking quite a lot at retrospinal cortex and other cortical 00:49:30.821 --> 00:49:36.181 regions, which talk to the hippocampal system, but they also talk to the head 00:49:36.181 --> 00:49:40.221 direction system and also quite a lot to other sensory systems. 00:49:40.461 --> 00:49:44.541 It seems to be quite a waypoint for many converging information streams. 00:49:45.398 --> 00:49:49.178 What about humans? Because I know some that have no sense of direction. 00:49:49.598 --> 00:49:53.238 I mean, I don't know about myself, but yeah. Members of my family, 00:49:53.298 --> 00:49:56.438 for instance, have no idea which direction they should go in. 00:49:56.438 --> 00:49:58.938 There's a lot of interesting research on humans. 00:49:59.218 --> 00:50:00.858 It's quite an enormous literature. 00:50:01.518 --> 00:50:05.578 And people vary a lot in their spatial capabilities. 00:50:06.178 --> 00:50:09.998 And one person who's been doing some interesting work on this is Eleanor McGuire, 00:50:10.158 --> 00:50:13.978 who has been looking recently. 00:50:14.158 --> 00:50:18.158 She's done a lot of work over many years looking at different aspects of navigation. 00:50:18.158 --> 00:50:22.458 But recently she's been looking at how people encode landmarks and use them in navigation. 00:50:23.098 --> 00:50:29.178 And she's found that people vary in their ability to decide how permanent a 00:50:29.178 --> 00:50:33.218 landmark is, which is an odd thing to have doubts about. 00:50:33.378 --> 00:50:35.858 But apparently people vary along this continuum. 00:50:36.138 --> 00:50:39.898 And she finds that people who are not very good at specifying how permanent 00:50:39.898 --> 00:50:41.838 landmarks are also not very good at navigating. 00:50:42.038 --> 00:50:46.078 They perform quite poorly in tests. and interestingly the brain structure that 00:50:46.078 --> 00:50:51.298 lights up when people are deciding about permanence of landmarks is retrospinal cortex. 00:50:52.138 --> 00:50:55.398 Now that's a brain structure that my lab has been interested in recently because 00:50:55.398 --> 00:51:00.898 it's very interested in landmarks and it has a lot of head duration cells and 00:51:00.898 --> 00:51:06.358 we think that it may be doing the job of processing landmarks and deciding to 00:51:06.358 --> 00:51:09.898 what extent they're useful to the head duration system and then attaching them 00:51:09.898 --> 00:51:11.698 or not attaching them to the the head direction signal. 00:51:12.378 --> 00:51:16.398 So this is work that's just beginning, but I think we're slowly starting to 00:51:16.398 --> 00:51:19.618 build up this picture that it's not really all about the hippocampus. 00:51:19.618 --> 00:51:23.338 The hippocampus is the core of a much bigger system. 00:51:24.098 --> 00:51:27.478 That's also what you argued towards the end, right? 00:51:27.558 --> 00:51:33.918 Because, okay, so in some sense, I guess you were hoping to find a clean three-dimensional 00:51:33.918 --> 00:51:38.258 tuning of, let's say, the grid cells, although we didn't really talk too much 00:51:38.258 --> 00:51:39.498 about the place cells in this context. 00:51:39.498 --> 00:51:42.618 But in some it doesn't come out so clean right and 00:51:42.618 --> 00:51:45.658 then you propose that maybe these are like the grid 00:51:45.658 --> 00:51:48.678 cells could be could be thought of as cylinders that cut through 00:51:48.678 --> 00:51:52.598 it of a uniform way the third dimension but 00:51:52.598 --> 00:51:55.798 the physiology was maybe not always that clean 00:51:55.798 --> 00:52:00.498 with respect to then still a grid-like structuring of the response because i 00:52:00.498 --> 00:52:03.698 remember on some of the walls you get a huge clustering of the response in one 00:52:03.698 --> 00:52:08.198 corner or the other corner right right and then And the point you made is that 00:52:08.198 --> 00:52:13.238 the modulation of this response is strongly dependent on gravity and the orientation of the body. 00:52:14.418 --> 00:52:20.718 So how does this now come in to the modulation of this grid cell response that you recorded? 00:52:22.118 --> 00:52:27.158 Well, the hypothesis, and it's really only a hypothesis, we don't have any data 00:52:27.158 --> 00:52:32.758 on the role of gravity, but the hypothesis is that the grid cell system The 00:52:32.758 --> 00:52:35.918 system wants to create a grid, 00:52:36.198 --> 00:52:42.958 and the grid is essentially a flat thing, and it needs to decide what surface 00:52:42.958 --> 00:52:44.358 is it going to lay its grid on. 00:52:45.362 --> 00:52:47.382 And normally, in a normal environment, 00:52:47.462 --> 00:52:49.762 we were just walking around on the floor, then it uses the floor. 00:52:49.902 --> 00:52:52.982 And of course, when we record rats in the laboratory, normally they're walking 00:52:52.982 --> 00:52:54.682 around on the floor and we see grids on the floor. 00:52:55.062 --> 00:52:58.102 But it's when we start to have rats walking around on things that aren't the 00:52:58.102 --> 00:53:03.662 floor, that we start to see this slight modulation of what the grid cells are doing. 00:53:03.662 --> 00:53:06.842 And the simplest explanation for 00:53:06.842 --> 00:53:10.962 the patterns we see is that the grid cell system sometimes chooses the wall 00:53:10.962 --> 00:53:15.542 as its reference plane and it's trying to produce a grid on the wall and sometimes 00:53:15.542 --> 00:53:20.742 it uses the floor even if the rat's walking around on a wall if the rat is actually 00:53:20.742 --> 00:53:24.182 oriented horizontally and the floor is beneath it and it can see it. 00:53:24.302 --> 00:53:27.742 So the pattern that we see on the pegboard with the stripes, 00:53:28.002 --> 00:53:31.282 we think that's happening because the grid cell system has decided to use the 00:53:31.282 --> 00:53:33.502 floor as its reference plane and not the wall. 00:53:33.662 --> 00:53:36.182 Even though the rat's climbing on the wall. 00:53:38.262 --> 00:53:42.502 That kind of simple model has to be qualified because of this finding that when 00:53:42.502 --> 00:53:45.802 the rat is walking around on the wall, we don't see these neat grids. 00:53:46.622 --> 00:53:50.962 Possibly they're neat grids if we could have a huge wall and have the rat walk around on it. 00:53:51.422 --> 00:53:55.902 I'm trying to persuade my PhD student, Julia, to do that experiment. 00:53:56.002 --> 00:53:57.862 It would be a very difficult one. Okay. 00:53:58.682 --> 00:54:02.902 So then the grid cells, the story 00:54:02.902 --> 00:54:05.882 is not finished. In which perspective is tuning in the third dimension. 00:54:06.622 --> 00:54:11.122 However, I think there's this strong conviction by everybody in the field that 00:54:11.122 --> 00:54:13.342 they do represent a metric, right? 00:54:13.402 --> 00:54:16.942 They really contribute to having a metric representation in this case of distance. 00:54:17.742 --> 00:54:21.042 But then in some of the heuristic is often like, well, you know, 00:54:21.042 --> 00:54:23.482 grid cells give you this great representation of space. 00:54:23.762 --> 00:54:27.282 And with those, we can build place cells. So that's fantastic. 00:54:27.282 --> 00:54:32.002 This was the original intuition, and then it came out that also if you look 00:54:32.002 --> 00:54:34.522 at development, it actually not occurs in that order. 00:54:34.802 --> 00:54:38.002 It's much more that the play cells help you to structure grid cells. 00:54:38.542 --> 00:54:40.942 So that raises a question about 00:54:40.942 --> 00:54:44.982 also the directionality of the information processing in this system. 00:54:45.202 --> 00:54:48.182 So you could also argue, look, if we take the cortical sheet, 00:54:48.442 --> 00:54:54.282 then at one end we will have this entorhinal cortex running across it, 00:54:54.362 --> 00:54:56.322 and then from there hangs our hippocampus. 00:54:56.322 --> 00:55:00.562 So we have now this interface between hippocampus and cortex through entorhinal 00:55:00.562 --> 00:55:02.742 cortex where we have this metric. 00:55:05.482 --> 00:55:08.802 It could also actually be a metric that helps Cortex to read out what the hell 00:55:08.802 --> 00:55:09.962 is going on in hippocampus. 00:55:10.282 --> 00:55:16.062 Is that an option you would consider? Yes. Yeah, I think the system is unlikely 00:55:16.062 --> 00:55:18.202 to have a directionality. 00:55:18.342 --> 00:55:24.522 I think the system is very bidirectional, very highly interconnected, in fact. 00:55:25.622 --> 00:55:30.402 The only exception to that really is this relatively one-way flow of information 00:55:30.402 --> 00:55:32.322 through the hippocampus itself. 00:55:33.002 --> 00:55:36.742 But even then, there are multiple shortcuts. So information coming from the 00:55:36.742 --> 00:55:41.522 entorhinal cortex to the dentate gyrus also takes a shortcut to CA3 and information 00:55:41.522 --> 00:55:44.342 going to CA3 also takes a shortcut to CA1 and so on. 00:55:45.482 --> 00:55:49.882 But then the output goes back to the entorhinal cortex and then it goes back out to cortex. 00:55:50.262 --> 00:55:55.922 So I agree that I think that the information flow goes both ways. 00:55:56.022 --> 00:55:58.202 I think all the structures depend on each other to some extent. 00:55:58.482 --> 00:56:05.362 So I think that the play cells, they're getting lots of information other than the grid cell input. 00:56:05.522 --> 00:56:08.142 In fact, you could knock out the grid cell signal and the place cells still 00:56:08.142 --> 00:56:11.822 produce quite nice fields, so they're quite capable of forming place fields. 00:56:12.982 --> 00:56:18.342 But I think what we will find if we do the relevant experiments is that in that 00:56:18.342 --> 00:56:20.942 situation, there's not really metric information. 00:56:21.162 --> 00:56:24.042 So the animal, for example, can't path integrate. 00:56:24.302 --> 00:56:30.762 And in fact, there's some work from Marseille that shows that if If you lesion 00:56:30.762 --> 00:56:33.682 into a rhino cortex, then animals lose the ability to calculate distances properly. 00:56:34.342 --> 00:56:39.342 So I think that right now, 00:56:39.442 --> 00:56:46.862 I would probably favor a model in which the place cells form a sort of a, 00:56:46.902 --> 00:56:50.622 what I think of as almost like a pixel map of the environment. 00:56:50.782 --> 00:56:55.502 So they sort of respond to the constellation of sensory cues that are present 00:56:55.502 --> 00:56:57.122 at each particular point in space. 00:56:58.282 --> 00:57:01.222 Together with a grid cell input, but they don't have to have the grid cell input. 00:57:02.522 --> 00:57:06.602 And the grid cells, in turn, use the place cells to know how to attach their 00:57:06.602 --> 00:57:07.882 grids to a given environment. 00:57:08.422 --> 00:57:11.182 And they do need the place cells. So if you knock out the place cells, 00:57:11.202 --> 00:57:12.502 the grids become very unhappy. 00:57:13.922 --> 00:57:19.142 And so the function of the grid cells is to help the place cells appropriately 00:57:19.142 --> 00:57:22.722 position their fields, for example, in the middle of a large open field where 00:57:22.722 --> 00:57:25.162 you're not near any boundaries and you've not got a lot of other information 00:57:25.162 --> 00:57:28.502 and so on. or if you close your eyes and walk around in the dark or something. 00:57:28.622 --> 00:57:33.222 So the grid cells are basically providing metric information that can substitute 00:57:33.222 --> 00:57:38.162 for the sensory cues to the place cells if the sensory cues drop out for some reason. 00:57:38.502 --> 00:57:42.122 So yes, I think it's very bidirectional. I think there's a lot of mutual dependency 00:57:42.122 --> 00:57:49.002 and the brain is highly interconnected and systems are all helping each other all the time, I think. 00:57:49.242 --> 00:57:54.102 But then the grid cell is the system that now can bidirectly interact between the two. 00:57:54.322 --> 00:57:58.142 It's driven by a velocity signal. This velocity signal also comes through quite 00:57:58.142 --> 00:58:01.322 a cascade of processing stages, including the thalamus. 00:58:02.080 --> 00:58:06.760 So this might suggest that there are other systems than your vestibular system 00:58:06.760 --> 00:58:09.880 that might grab hold of driving this velocity signal. 00:58:09.960 --> 00:58:14.020 So it means I could actually start to distort this metric or I could even impose 00:58:14.020 --> 00:58:15.520 a completely different kind of metric. 00:58:15.720 --> 00:58:18.300 Yes. Do you consider that option? 00:58:18.740 --> 00:58:26.400 Yes. So a lot of very elegant work has been done using virtual reality to independently 00:58:26.400 --> 00:58:33.440 manipulate various aspects of the signals that the brain could be using to extract velocity. 00:58:33.980 --> 00:58:39.920 So that includes things like motor cues. So how many footsteps and how quickly 00:58:39.920 --> 00:58:40.660 are they being produced? 00:58:41.520 --> 00:58:46.420 How quickly is the optic flow signal moving past the eyes and so on? 00:58:46.420 --> 00:58:49.240 So there are other cues than just the vestibular signal. 00:58:49.740 --> 00:58:55.380 And so people have played around with independently varying these to see what happens. 00:58:55.980 --> 00:58:59.180 And I think the story that's emerging is a little bit complex. 00:58:59.280 --> 00:59:04.760 One of the things that is quite striking to me is that nobody has yet shown 00:59:04.760 --> 00:59:08.500 a head direction signal in virtual reality. 00:59:09.280 --> 00:59:15.600 So I think the head direction signal is possibly quite dependent on the vestibular signal. 00:59:15.660 --> 00:59:19.180 Of course, the vestibular signal is the thing that's missing in virtual reality. Exactly. Usually. 00:59:19.720 --> 00:59:23.660 At least that's true for animals where the head is fixed. If there are some 00:59:23.660 --> 00:59:27.940 variants of virtual reality, for example, David Tank's group has a virtual reality 00:59:27.940 --> 00:59:33.060 setup where the animals can rotate, but they can't move in linear space. 00:59:33.060 --> 00:59:36.080 So they're running on a ball and they can freely rotate so they can get the 00:59:36.080 --> 00:59:40.680 angular component of the vestibular signal and there there are head direction 00:59:40.680 --> 00:59:46.140 cells and indeed they see grid cells and place cells but the grids are quite expanded, 00:59:47.300 --> 00:59:49.180 which I think is interesting because. 00:59:51.727 --> 00:59:55.967 And slightly analogously to the findings that we have on the wall, 00:59:56.207 --> 00:59:57.747 the thing that's missing in 00:59:57.747 --> 01:00:01.167 that apparatus is a linear acceleration signal from the vestibular system. 01:00:01.567 --> 01:00:04.207 And that was one of the things that made us start thinking maybe that's what's 01:00:04.207 --> 01:00:06.787 wrong with our grid cells on the wall is the absence of the signal. 01:00:06.987 --> 01:00:11.747 So I think the vestibular system is not the only velocity signal. 01:00:11.927 --> 01:00:15.187 It's not necessarily even the most important velocity signal, 01:00:15.287 --> 01:00:20.847 but I think it normally is there and it normally is quite supportive to the system. 01:00:20.847 --> 01:00:23.987 It's it's a little bit worrying that you're 01:00:23.987 --> 01:00:27.307 not getting head direction cells given what we've talked about uh in 01:00:27.307 --> 01:00:30.347 these virtual reality environments where you know 01:00:30.347 --> 01:00:35.507 animals are running on balls and things because that's uh one way in which the 01:00:35.507 --> 01:00:39.567 field has moved quite a lot just in the last decade you have an animal head 01:00:39.567 --> 01:00:45.087 fixed so you can do a lot of detailed recording but you've thrown away um perhaps 01:00:45.087 --> 01:00:47.467 quite a significant amount of the ethological relevance, 01:00:47.947 --> 01:00:51.207 to what those actual recordings might mean yeah and it's 01:00:51.207 --> 01:00:55.067 it's it struck me as well listening to your talk today that you're actually 01:00:55.067 --> 01:01:00.707 using uh very enriched environments but uh and the system you're you're looking 01:01:00.707 --> 01:01:06.027 at is then perhaps a bit closer to the you know the natural animal and its free 01:01:06.027 --> 01:01:09.707 state than some of these other environments that um. 01:01:10.623 --> 01:01:14.163 People have been using to look at the hippocampal system before. 01:01:14.323 --> 01:01:17.863 And you also, you know, you noted that when we went from testing animals in 01:01:17.863 --> 01:01:21.523 small boxes to testing animals in slightly bigger boxes, two meters, 01:01:21.663 --> 01:01:24.083 which isn't huge, then we discovered grid cells. 01:01:24.583 --> 01:01:30.343 So I'm just wondering, you know, what else are we going to discover when we 01:01:30.343 --> 01:01:33.863 really take seriously the lifestyle of the animal? 01:01:33.983 --> 01:01:37.303 You know, sort of these animals live in tunnels a lot. So, you know, 01:01:37.303 --> 01:01:42.623 they move around nocturnally most of their time, and we're testing them under lighted conditions. 01:01:42.903 --> 01:01:44.823 Are we missing other important things? 01:01:45.203 --> 01:01:50.863 I think that's a really, really important point. And I totally agree that the 01:01:50.863 --> 01:01:55.583 conditions that we're recording in at the moment are not very ethologically valid. 01:01:55.823 --> 01:01:59.263 And there are pluses and minuses to that. So one of the things that you do, 01:01:59.303 --> 01:02:03.643 of course, is to reduce the complexity, and that enables you to look at factors in isolation. 01:02:03.903 --> 01:02:09.003 But, of course, what you lose is the real-world relevance. So ultimately, 01:02:09.163 --> 01:02:11.903 you have to start putting all of the stuff back together again. 01:02:12.003 --> 01:02:16.523 And I think one of the big outstanding questions for me is what do grid-in-place 01:02:16.523 --> 01:02:22.963 cells do in a normal cluttered environment like a burrow system or a field with 01:02:22.963 --> 01:02:24.203 trees and rocks and things like that? it. 01:02:24.783 --> 01:02:32.803 And my belief, which we've not tested yet, but it'd be relatively easy to test, is that. 01:02:34.043 --> 01:02:42.403 A day in the life of a grid cell wouldn't result in a nice hexagonal close-packed array of firing fields. 01:02:42.523 --> 01:02:45.183 I think if you just recorded a grid cell for a week or two in a rat, 01:02:45.303 --> 01:02:48.883 you'd find it almost never produced a hexagonal close-packed array of firing fields. 01:02:49.143 --> 01:02:53.183 Most of the time it would produce blobs scattered at what look like random places 01:02:53.183 --> 01:02:55.783 around the environment because of course the rat is, you know, walking. 01:02:56.223 --> 01:03:00.723 Rats, they have these very stereotyped kind of behaviours in familiar territory 01:03:00.723 --> 01:03:04.163 where they have these little rat runs, you know, little paths that they like to follow and so on. 01:03:04.543 --> 01:03:08.523 They don't tend to forage in an even way across the surface of the environment. 01:03:09.642 --> 01:03:14.842 So I think when we start to think about what are these cells actually doing 01:03:14.842 --> 01:03:21.302 for the animal, we need to bear that in mind, that we're very captivated by 01:03:21.302 --> 01:03:23.642 the regularity of the pattern that we're able to elicit. 01:03:23.682 --> 01:03:26.842 But the regularity of the pattern may not be the thing that the brain really 01:03:26.842 --> 01:03:29.902 cares about with grid cells. It may be something else. 01:03:30.662 --> 01:03:34.402 It may be, for example, the function of a grid cell is just to separate out 01:03:34.402 --> 01:03:39.202 pieces of the environment so that the representations of them don't kind of 01:03:39.202 --> 01:03:42.962 bleed into each other or something like that that's kind of a bit different 01:03:42.962 --> 01:03:43.942 from how we've been thinking. 01:03:44.062 --> 01:03:46.702 So we really don't know what they're for yet. 01:03:46.942 --> 01:03:49.962 And so I think your ethological point is extremely valid. 01:03:50.582 --> 01:03:55.182 So they might as well be de-correlating spatial representations given that they 01:03:55.182 --> 01:03:59.202 have only, let's say, interleaved responses to space. 01:03:59.782 --> 01:04:02.642 Possibly, possibly. I think it's a hypothesis worth considering. 01:04:04.702 --> 01:04:12.562 So, Kate, I mean, you have attacked these really complicated problems of, you know, rats in space. 01:04:12.942 --> 01:04:17.162 And you're going to be busy with that for a little while. You also have been 01:04:17.162 --> 01:04:19.842 very close, if you want, to the discovery of grid cells. 01:04:20.222 --> 01:04:24.382 You've been exposed to all this work around spatial cognition, place cells. 01:04:25.742 --> 01:04:29.962 So, in that sense, you really represent a very specific tradition in neuroscience, 01:04:30.182 --> 01:04:33.822 a systems neuroscience. science, trying to link behavior to the neural substrate. 01:04:34.102 --> 01:04:38.642 So if we now would like to follow in your tradition, what would be Kate's Law 01:04:38.642 --> 01:04:41.842 that we have to write on the wall and read every morning when we wake up? 01:04:43.722 --> 01:04:46.402 Kate's Law. Ooh. Um. 01:04:48.451 --> 01:04:51.031 Okay, you've caught me on the back foot there. I'd have to go away. 01:04:51.671 --> 01:04:56.471 I mean, I can, I've, you mean, you mean a law as in a law about the functioning of the brain? 01:04:56.571 --> 01:05:00.351 No, a law that we have to adhere to in terms of studying and understanding the brain. 01:05:00.531 --> 01:05:03.231 It's kind of your advice to your new grad student. 01:05:03.951 --> 01:05:06.571 I think. Look at me as a new grad student. 01:05:08.811 --> 01:05:14.231 Well, one law is the brain is very complicated. The other, the other thing is 01:05:14.231 --> 01:05:15.191 the brain is very simple. 01:05:15.191 --> 01:05:27.711 I think the brain is made of these very slowly computing blobs of jelly. 01:05:28.091 --> 01:05:30.991 And I think wherever possible, it's trying to optimize the problems it's trying 01:05:30.991 --> 01:05:33.351 to solve so as to save itself as much work as possible. 01:05:33.471 --> 01:05:36.911 And I think we have to remember that, particularly when we're designing artificial 01:05:36.911 --> 01:05:40.871 intelligent machines and things like that, we have to think to what extent are 01:05:40.871 --> 01:05:44.011 the systems biologically realistic and to what extent should they be? 01:05:44.011 --> 01:05:47.551 Because biology really is just trying to keep the animal alive and it may not 01:05:47.551 --> 01:05:51.491 be trying to do it in the most elegant way so much as the way that works for 01:05:51.491 --> 01:05:55.191 the environment of the animal at the time. So I guess that would be it. 01:05:55.411 --> 01:06:00.891 Okay. So then Tony actually likes trains a lot and he also likes to take the 01:06:00.891 --> 01:06:04.651 train often from Sheffield to London and five years from now I'll buy him a 01:06:04.651 --> 01:06:06.651 train ticket to go to London, visit you, 01:06:06.771 --> 01:06:10.511 and he will visit your lab and he will come with a piece of paper that says, okay, Kate, 01:06:10.591 --> 01:06:15.571 five years ago you made this prediction and today we want to know whether it 01:06:15.571 --> 01:06:20.911 was verified or not or rejected so what's the most important prediction you 01:06:20.911 --> 01:06:25.451 would like to make today in this time window of five years that you really want 01:06:25.451 --> 01:06:28.911 to see tested in that time frame um, 01:06:31.176 --> 01:06:33.816 Well, several predictions. I mean, five years is not very long, 01:06:33.856 --> 01:06:35.016 so I should hedge my bets. 01:06:37.576 --> 01:06:41.896 Don't, you can't wait that long, Kate, sorry. Sorry. Important prediction. 01:06:42.056 --> 01:06:48.976 So one thing I think is I really like this idea that the vestibular cerebellum 01:06:48.976 --> 01:06:56.016 provides a signal that modulates the updating of the head direction signal in 01:06:56.016 --> 01:07:01.016 three dimensions and allows us to essentially to relate all of our frames of reference. 01:07:01.016 --> 01:07:06.576 So I think that the function of the cerebellum may well be thought of as a way 01:07:06.576 --> 01:07:11.056 of transforming between reference frames, and that it uses the gravity signal 01:07:11.056 --> 01:07:12.516 and the other vestibular signals to do that. 01:07:12.616 --> 01:07:17.876 So that's something that I would like to have at least started on in five years, 01:07:17.976 --> 01:07:19.816 and we're just thinking about how we might do that. 01:07:20.856 --> 01:07:26.156 Second thing, grid cells. I would like to have found out by then whether grid 01:07:26.156 --> 01:07:29.296 cells form a hexagonal close-packed lattice in three dimensions. 01:07:29.596 --> 01:07:33.596 And I'm honestly agnostic about that. I've been selling this multi-planar idea, 01:07:33.796 --> 01:07:37.656 but I have an open mind about whether that's really true. 01:07:38.576 --> 01:07:43.756 And then the third thing is that we're very interested in the retrosplenial cortex. 01:07:44.236 --> 01:07:47.976 I think it's a really interesting and barely understood structure. 01:07:48.316 --> 01:07:53.116 And I'm I'm hoping that we will have made a big step towards understanding what 01:07:53.116 --> 01:07:55.096 it's doing with all that information that it's getting. 01:07:56.216 --> 01:07:59.176 Great. Okay, Jeffrey, thank you so much for this conversation. 01:07:59.336 --> 01:08:00.216 Thank you. Thank you. Thank you. 01:08:02.927 --> 01:08:08.887 The CSN podcast was produced by the Convergent Science Network of Biometrics 01:08:08.887 --> 01:08:15.007 and Biohybrid Systems, a project funded by the European Sevens Research Framework 01:08:15.007 --> 01:08:15.847 Programme. Oh, that was fun. 01:08:16.707 --> 01:08:19.667 Yeah. You did so well. For more interviews, recorded lectures, 01:08:20.127 --> 01:08:28.427 or upcoming conferences in the field of biometrics and biohybrid systems, go to csnnetwork.eu. 01:08:28.707 --> 01:08:30.427 Thank you for listening. 01:08:33.327 --> 01:08:38.227 I'm trying to think of all the people I've insulted along the way write all 01:08:38.227 --> 01:08:44.827 these apologetics oh I didn't mention so and so oh I attributed this to them no don't worry, 01:08:45.927 --> 01:08:52.887 now it's interesting that indeed the system is indeed not as clean as we would 01:08:52.887 --> 01:08:56.347 like it to be and how we want also how it is represented in literature, 01:08:57.047 --> 01:09:03.227 right this is I think we have to be so careful with that Because we built this 01:09:03.227 --> 01:09:04.387 caricature of what this is. 01:09:04.467 --> 01:09:08.447 I mean, of course, as soon as you get closer to more realistic task conditions, 01:09:08.767 --> 01:09:12.207 it will not look at all like this. So maybe even take the wolf situation. 01:09:13.207 --> 01:09:17.447 Maybe this big blob you see in the corner is much closer to an ecologically 01:09:17.447 --> 01:09:21.027 valid response of your grid cells than the beautifully, nicely, right? 01:09:21.107 --> 01:09:23.787 It's definitely possible. We cannot exclude that right now. 01:09:23.927 --> 01:09:26.367 Yeah, yeah. And it's sort of a scary thought. 01:09:27.227 --> 01:09:35.367 Yes. Well, scary, but also interesting. I mean, I think the field tends to attract 01:09:35.367 --> 01:09:41.067 a lot of engineers and physicists and people who have a certain way of thinking about things. 01:09:41.927 --> 01:09:45.587 And I think the sort of niche that I inhabit is a slightly unusual one because 01:09:45.587 --> 01:09:49.087 I'm more of a psychologist slash ethologist by inclination. 01:09:50.867 --> 01:09:56.607 I look at these beautiful elaborate models of interacting oscillations and this, 01:09:56.647 --> 01:10:00.427 that and the other. And I think there's a spectacular intellectual achievement. 01:10:01.227 --> 01:10:05.367 But is that really what it's for? You know, in the messy real world when an 01:10:05.367 --> 01:10:07.087 animal is climbing up and down over rocks? 01:10:07.447 --> 01:10:11.087 Have we really got this, you know, these fine, great, maybe we have. 01:10:11.947 --> 01:10:16.967 But I do think we need some people who are slightly more grounded in the psychology 01:10:16.967 --> 01:10:19.027 and the behavior and stuff. 01:10:19.987 --> 01:10:23.987 Well, this is also the point that I think also the point Tony made about the 01:10:23.987 --> 01:10:27.627 robots, for instance. So to just try to convince people, look, 01:10:27.667 --> 01:10:30.687 if you want to build models of these kinds of things, link it to a robot because 01:10:30.687 --> 01:10:32.427 it gets you a little bit closer to real behavior. 01:10:32.627 --> 01:10:36.567 Because Leash, this is also a bit my beef earlier about these grid cell models, 01:10:36.707 --> 01:10:37.627 these attractive grid cell models. 01:10:38.987 --> 01:10:43.247 They all came out in 2006. We also produced one. Yeah. 01:10:44.472 --> 01:10:46.892 The only one that actually was linked up to a robot and showed, 01:10:47.032 --> 01:10:52.312 look, this gives you a grid-like response over time, doesn't want to be produced 01:10:52.312 --> 01:10:53.672 because it was linked to a robot. Right. 01:10:53.872 --> 01:10:57.592 All the others are more conceptual models. Like, okay, yes, you can imagine 01:10:57.592 --> 01:11:02.472 that, but the stability conditions are dramatically different. Right, right. 01:11:02.792 --> 01:11:06.552 So I think this is really a contribution that this more robot-oriented thing can make. 01:11:06.572 --> 01:11:10.372 You can say, look, consider the computational principles in the context of this 01:11:10.372 --> 01:11:14.452 embodied real-world system. Yeah, yeah. It has to satisfy all these constraints. 01:11:14.672 --> 01:11:18.392 If the robot doesn't give you behavior that looks plausible, forget the model. 01:11:18.652 --> 01:11:20.712 Yeah, because I mean, I think it's a really important point. 01:11:20.812 --> 01:11:23.632 Because, you know, when you're building a computational model, 01:11:23.692 --> 01:11:24.952 you can have perfect senses. 01:11:25.252 --> 01:11:27.692 You know, you can have an absolutely veridical velocity signal, 01:11:27.752 --> 01:11:30.052 but the real world isn't like that. Sensors are very imperfect. 01:11:30.572 --> 01:11:35.332 They accumulate error quickly. There's a lot of, you know, uncertainty and conflict 01:11:35.332 --> 01:11:36.432 and this, that, and the other. 01:11:36.892 --> 01:11:39.352 I mean, so you asked me what, you know, robotics has taught me. 01:11:39.352 --> 01:11:42.952 And one of the things I learned early on, so my husband, 01:11:43.052 --> 01:11:50.152 Jim, who is a roboticist, who was trying to get robots to integrate sensory 01:11:50.152 --> 01:11:53.352 inputs from different sensory modalities. 01:11:53.372 --> 01:11:55.712 And it seemed like a really simple problem. 01:11:55.832 --> 01:11:58.672 It just turned out to be incredibly difficult because when you look at what 01:11:58.672 --> 01:12:01.632 raw sensory data look like, it's just a mess. 01:12:02.432 --> 01:12:07.332 And how do you extract spatial signals from all of that? And so I think you're absolutely right. 01:12:07.332 --> 01:12:11.012 Right, I think if you want your computational model to have any kind of credibility, 01:12:11.172 --> 01:12:14.132 you need to show that it could work in the messy real world, 01:12:14.252 --> 01:12:17.992 given the constraints of noisy sensors and this, that, and the other. 01:12:18.272 --> 01:12:21.552 Right, and I think that's why you need the robustness of multiple systems. 01:12:21.972 --> 01:12:27.692 Yeah. So I think the grid cell data is beautiful, that you have this multi-module 01:12:27.692 --> 01:12:32.732 system that the Moses have shown that gives you an actual location in space. 01:12:33.572 --> 01:12:36.392 But that's still only one way of understanding, 01:12:37.100 --> 01:12:39.980 of of working out where you are and it could be wrong and then 01:12:39.980 --> 01:12:42.900 what's your backups and i think the beautiful thing 01:12:42.900 --> 01:12:46.160 about animals is they have lots of different backups 01:12:46.160 --> 01:12:49.340 and they're potentially independent because 01:12:49.340 --> 01:12:52.140 you might want them to be you don't want you don't want everything to depend 01:12:52.140 --> 01:12:56.940 on your head direction system and people without a good sense of direction get 01:12:56.940 --> 01:13:00.380 around perfectly well in everyday life so you know how they do well some do 01:13:00.380 --> 01:13:07.520 some do some well they make mistakes But they generally have strategies for compensating. 01:13:07.600 --> 01:13:12.620 And that's what I think biological things are good at, is compensating for the 01:13:12.620 --> 01:13:17.040 lack of perfect information, compensating when some parts of the system fail. 01:13:17.660 --> 01:13:21.380 And that's where robots fall down right now. We usually have one way of doing stuff. 01:13:21.980 --> 01:13:25.040 For instance, we're trying to build driverless cars now. 01:13:25.460 --> 01:13:29.660 And some people, like Elon Musk, say five years away we'll have it. 01:13:29.660 --> 01:13:34.940 But what he's got is a system that works when you've got nice white lines down 01:13:34.940 --> 01:13:36.560 the edge of the road, then you can drive. 01:13:36.700 --> 01:13:41.600 But he doesn't have anything like the multiple systems that actually give you 01:13:41.600 --> 01:13:46.200 a robust, fail-proof, fail-safe system that's never going to crash. 01:13:47.500 --> 01:13:51.700 And other people will say, look, we're 50 years away from having that for driverless cars. 01:13:51.700 --> 01:13:55.040 Another example of that is the importance 01:13:55.040 --> 01:13:58.060 of this more embodied ecologically valid approach take 01:13:58.060 --> 01:14:01.020 all this noise we had about uh what's 01:14:01.020 --> 01:14:04.200 called deep q learning like now we have a deep learning network that can 01:14:04.200 --> 01:14:10.960 also act and then it can learn these atari games um okay that's cute but actually 01:14:10.960 --> 01:14:15.400 what you see is that in order to make that work they have to randomly sample 01:14:15.400 --> 01:14:19.340 the input space now think about that right that's great when you write an algorithm 01:14:19.340 --> 01:14:22.460 when i'm a behaving system My input stream is continuous. 01:14:22.700 --> 01:14:26.800 I cannot jump around like a frog and make sure I have an even sampling. 01:14:27.360 --> 01:14:31.360 This has been documented since the early 90s in these more embodied models. 01:14:32.600 --> 01:14:36.720 This has also been shown to be a major weakness in these sort of hierarchical 01:14:36.720 --> 01:14:39.620 classifier systems that people also exploit in deep learning. 01:14:41.484 --> 01:14:44.784 That means as soon as he thinks, and this is not considered a problem right 01:14:44.784 --> 01:14:47.584 now, because no one is thinking it through in behavioral terms. 01:14:47.724 --> 01:14:48.964 And I think that's a real problem. 01:14:49.144 --> 01:14:52.324 Because now, in some sense, we're getting a lot of noise in the literature, 01:14:52.524 --> 01:14:53.624 a lot of people being distracted. 01:14:54.484 --> 01:15:00.504 And also, I think, misinformed because we're not imposing the right constraints. 01:15:00.684 --> 01:15:05.344 And in my opinion, in these models, we must insist on bringing together behavior, 01:15:05.764 --> 01:15:06.484 anatomy, and physiology. 01:15:06.884 --> 01:15:10.144 Any model must answer that. And if you're not able to do that, 01:15:10.264 --> 01:15:13.364 it's under constraint and for me, it's noise. 01:15:13.744 --> 01:15:19.344 It's not necessarily helping, but it's not, of course, it makes life more difficult. 01:15:20.464 --> 01:15:23.224 Replicating colorful pictures in MATLAB is a lot easier. I mean, 01:15:23.224 --> 01:15:24.964 it sort of depends what you're trying to do though. 01:15:25.044 --> 01:15:28.284 I mean, some people are trying to understand the brain, but some people are 01:15:28.284 --> 01:15:33.124 trying to build neural networks that do things and looking at the brain, 01:15:33.224 --> 01:15:35.904 it may give you some ideas, years but replicating the brain 01:15:35.904 --> 01:15:38.864 is not necessarily a good thing to do because it is that's 01:15:38.864 --> 01:15:41.784 fine but they shouldn't claim that they explain anything yeah 01:15:41.784 --> 01:15:45.464 so but as the people you know i can i can see you know i could see that the 01:15:45.464 --> 01:15:49.764 discipline is full of all sorts of people doing stuff that's just totally useless 01:15:49.764 --> 01:15:54.604 but i do think that there are um there are kind of insights that you can get 01:15:54.604 --> 01:15:59.804 from stuff like this as well but it's also a matter of being being upfront about 01:15:59.944 --> 01:16:01.664 what your constraints are. 01:16:01.764 --> 01:16:05.384 Of course, anyone is free to play around to get an idea about some competition. 01:16:05.704 --> 01:16:09.704 Great. Yeah. Right, but be clear about it. Don't pretend that suddenly now we 01:16:09.704 --> 01:16:11.964 have an explanation for how a brain might do something. 01:16:12.304 --> 01:16:15.964 Yeah, yeah. People are overconfident. Yes, absolutely. And there's too much 01:16:15.964 --> 01:16:17.444 methamphetosis going on. Yeah. 01:16:17.604 --> 01:16:20.844 In my opinion. Yeah. But anyway, behavior is the way forward. 01:16:21.504 --> 01:16:25.164 Well, you know, yeah. You need all of these different approaches. 01:16:25.164 --> 01:16:26.344 Would you do any modeling yourself? 01:16:27.384 --> 01:16:31.404 Um, not, not much. I've done a little bit. Okay. 01:16:31.544 --> 01:16:38.804 Um, and I've just hired a modeling person actually, who's, who's worked on the ring tractor network. 01:16:38.864 --> 01:16:43.464 And so we're going to get him to try and model a fair color tractor and just play around with it.