Listeners:

Top listeners:

00:00 00:00

Go to album

Paul Verschure on consciousness and distributed adaptive control CSN Podcasts
Edvard Moser on grid cells and entorhinal cortex CSN Podcasts
Giacomo Rizzolatti on mirror neurons and action understanding CSN Podcasts
Robert Axelrod on game theory and prisoner's dilemma CSN Podcasts
Adrian Owen on disorders of consciousness and vegetative state CSN Podcasts
Jonathan Whitlock on markerless motion capture and posterior parietal cortex CSN Podcasts
Luis Puelles on neuroanatomy and prosomeric model CSN Podcasts
Zoltan Molnar on subplate neurons and cortical development CSN Podcasts

Neil Burgess on boundary vector cells and place cells

Episode 8 15.03.2018

PLAY EPISODE

Season 2018

0 Followers 10 Episodes

Description

How does the hippocampus know where you are when all it receives is egocentric sensory input? Computational neuroscientist Neil Burgess explains how boundary vector cells provide the missing link , translating distances to environmental features into the allocentric place code that underpins spatial memory and navigation. Subscribe for more from the Convergent Science Network podcast series. Neil Burgess joins Paul Verschure and Tony Prescott at the BCBT summer school to present his boundary vector cell model of hippocampal place cell firing. The model proposes that place cells receive their spatial tuning from a population of cells, found in subiculum and entorhinal cortex, that each encode the distance and allocentric direction to extended environmental boundaries. A place field emerges as a thresholded sum of these boundary inputs , a simple mechanism that accounts for how place fields stretch, split, or disappear when environments are deformed, and why place cells near walls tend to be more stable than those in open space. The discussion traces the interplay between theory and experiment that has driven Burgess’s career. He explains why the boundary vector cell model uses summation with a threshold rather than multiplication: environment-stretching experiments show place field sub-peaks being pulled apart while maintaining fixed absolute distances from walls, rather than tracking constant ratios , evidence against a Bayesian multiplicative combination. The conversation also addresses the critical role of head direction cells as the compass that orients the entire system, and how retrosplenial cortex likely performs the egocentric-to-allocentric coordinate transformation needed to anchor head direction signals to sensory landmarks. Burgess and the hosts debate whether the hippocampus represents a single best estimate of location or entertains multiple spatial hypotheses simultaneously. While there is limited direct evidence for multiple concurrent hypotheses in place cell firing, running-direction-dependent modulation of split place fields suggests that path integration and sensory inputs are being combined, with different peaks receiving different weights depending on movement direction. The conversation also explores the successor representation idea , that place cells may encode not just current location but the probability of future occupancy, enabling more efficient reward estimation. Key topics include the boundary vector cell model, the relationship between place cells and grid cells, egocentric-to-allocentric transformation in retrosplenial cortex, environment deformation experiments, path integration and its error accumulation, the puzzle of finding boundary vector cells at both input and output stages of the hippocampal loop, and the Bayesian versus competitive interpretations of spatial coding. Part of the Convergent Science Network podcast series from the BCBT Summer School.

Tagged as:
Boundary Vector boundary vector cells Burgess Cell Model place cells Spatial Vector Cell

About the author

CSN Podcasts

Both the triumphs of humanity and its most evil deeds have resulted from collaboration. In a time where humanity is required to aspire to the former and minimize the latter, the question arises of how collaboration arises and why it fails. Surprisingly, this phenomenon, so central to who we are, is not well understood. Hence, a collaborative effort is required to understand collaboration in its full biological, psychological, sociological, cultural, and economic complexity and to translate this understanding into operational impact. This series of podcasts is one step toward achieving these complementary goals. The Collaboration Podcast presents interviews with people who are central orchestrators of collaboration in various domains including business, government, science, art, health, sustainability, and the military. The discussions were conducted by Prof. Dr. Paul F.M.J. Verschure and members of the Program Advisory Committee of the Ernst Strungmann Forum on Collaboration (https://www.esforum.de/forums/ESF32_Collaboration.html) during 2021 and had the goal to sketch a map of opportunities, challenges, and obstacles in human collaboration. The forum took place in May 2022, and now we would like to share this series of interviews with a broader audience. The full report of the Forum will be published in 2023 by MIT Press. The podcast was produced by the Convergent Science Network (https://www.convergentsciencenetwork.org/). Context: The stability of social systems depends critically on realizing sustainable methods of “collaboration,” yet how and by which means collaboration is achieved is not clearly understood; neither are the conditions or processes that lead to its breakdown or failure. Collaboration can be understood as cooperation between agents toward mutually constructed goals. Part of the reason for our lack of understanding is that the phenomenon of collaboration is, by nature, a highly multidisciplinary problem, and effective research into its complexities has been difficult to achieve across the broad range of scientific and technical disciplines involved. The need for a fundamental understanding of collaboration, however, has become increasingly important. Not only does humankind demand answers as it attempts to address critical challenges at multiple scales (e.g., climate change, migration, enhanced automation, social and economic inequality), but ever-increasing technological and economic means of interconnecting people and societies are disrupting long-established, familiar patterns of how we interact. Radical technological changes that are ongoing have the potential to reshape collaboration in ways that are currently hard to predict or influence (e.g., by altering configurations in interaction, information creation, and modes of communication). On one hand, such changes could disrupt hitherto stable forms of collaboration by affecting critical communication channels and traditional roles, as can be observed in the rapidly changing patterns in governance, commerce, and social interaction. Conversely, technology could lead to the emergence of novel, successful forms of collaboration that deviate from traditional “hierarchical” architectures. Evidence of this can be seen in areas as diverse as highly automated manufacturing plants, the open science movement, collaborative software repositories, user-centered services, and the sharing of economy-based modes of organization. Without a fundamental understanding of the mechanisms, processes, and boundary conditions of collaboration, it is not possible to evaluate or predict which of these possible scenarios are sustainable or even plausible. The Forum “How Collaboration Arises and Why it Fails” (May 8–13, 2022, Location: Frankfurt am Main, Germany) Chairs: Andreas Roepstorff and Paul Verschure Program Advisory Committee: Jenna Bednar, Julia R. Lupp, Bhavani R. Rao , Andreas Roepstorff, Ferdinand von Siemens, and Paul Verschure

Timestamp

00:00:03 - This is the Convergent Science Network podcast. Leading researchers in the domain
00:00:10 - of neuroscience, brain theory and technology are interviewed by Paul Verschoor and Tony Prescott.
00:00:21 - This is Paul Verschoor with the Convergent Science Network podcast with my colleague
00:00:25 - Tony Prescott. And we're here today with Neil Burgess, who is a speaker in our
00:00:32 - 10th edition BCBT Summer School.
00:00:35 - And Neil was giving an overview of his work on spatial cognition,
00:00:41 - also about how to link neural dynamics with advanced psychological function,
00:00:47 - like our knowledge of space.
00:00:49 - And Neil, what you also presented very much in your description of your work
00:00:55 - is this close coupling between theoretical approaches, modeling work, and experimental work.
00:01:02 - So where did this combination of these two methods, if you want, originate?
00:01:10 - Well, I was originally a computational person. Having done theoretical physics,
00:01:15 - I started to want to do computational models of memory and other forms of cognition at the level of neurons.
00:01:26 - And well, my PhD supervisor, theoretical physicist called Mike Moore,
00:01:31 - said, you know, if you're going in this direction, it's an experimental subject
00:01:35 - and you need to get involved in the experiments, and I'm always grateful for that advice.
00:01:40 - There would be no point sticking to mathematically tractable models if they
00:01:47 - had nothing to do with the biological reality,
00:01:49 - and so it seems very important to have your models directly address experiment
00:01:54 - and equally for your experiments to be theoretically well grounded to answer
00:01:59 - particular hypotheses that make sense.
00:02:01 - And so experiment and theory should always work together.
00:02:07 - But now, how many of the models you have generated in that period have you actually
00:02:12 - completely rejected and changed your mind about and taken a completely different direction?
00:02:21 - That's a good question. I think that typically you're interested in a phenomenon
00:02:29 - and you continue to follow that even when experiments or models don't work.
00:02:34 - And so if you're interested in explaining that phenomenon, your models might
00:02:39 - change and indeed your experiments might. so in terms of all that rejection um.
00:02:51 - I'm i'm struggling so i started off uh doing um kind of hopfield model type
00:02:57 - models of memory and i was working with a psychologist graham hitch who was
00:03:02 - interested in memory for serial order,
00:03:05 - and it seemed that although you can remember sequences of things in hopfield
00:03:09 - models by associating one pattern of activity to the next in in human uh working
00:03:14 - memory for serial order Often you make paired transpositions of items.
00:03:19 - So instead of A, B, C, D, you remember, or you output A, C, B, D.
00:03:24 - And it's very hard to get that kind of error in a straightforward Hopfield model.
00:03:27 - And so I started using competitive queuing type models.
00:03:31 - And I haven't really looked back on the Hopfield type models,
00:03:34 - although I still believe that the attractor network idea is key to how much
00:03:41 - of the brain works, including hippocampal area CA3.
00:03:45 - I haven't really gone back to those models as just a pure attractive model of memory.
00:03:50 - Right. So then you started out by saying,
00:03:54 - by pointing to this Wang and Simmons experiment that makes the point that actually
00:03:58 - there are a number of cues or a number of streams of information that come together
00:04:03 - in our understanding of space.
00:04:06 - So why do you think that that experiment is sort of critical as a starting point of the analysis?
00:04:13 - Well, it's just one experiment. I don't think it's critical,
00:04:17 - but it very nicely presents the fact that you can remember where something is in many different ways.
00:04:27 - And so it's sometimes good to start with that example, so that even though I
00:04:31 - then might drone on about hippocampal place cells for a long time,
00:04:35 - the audience already knows that that's only one kind of strategy for remembering
00:04:40 - where something is and that other parts of the brain would be doing equally
00:04:44 - useful things that can help you remember stuff in different ways.
00:04:49 - So of these different possible ways to deal with space, as a concept,
00:04:55 - how would you then define the way the hippocampus solves the problem of space?
00:05:04 - Well, I mean, all of that part of the model is very much driven by the place
00:05:11 - cell phenomenon and surprisingly single neurons encoding for single places.
00:05:16 - I mean it's a remarkable sort of one-to-one mapping aside from the fact that
00:05:21 - different cells will fire in different environments and provides a very easy to use code.
00:05:26 - You could associate things that happen in places with the activity of the place
00:05:31 - cell that fires in that place very easily or you could,
00:05:36 - having found a place cell that fires in a given place it could
00:05:39 - be that its firing rate is modulated by some other factor
00:05:42 - that happens in that place and that's just a very simple starting point
00:05:44 - for thinking about a memory system
00:05:48 - for location yeah i think
00:05:51 - going back to this question about parallel representations a lot
00:05:54 - of the hippocampal models and i think yours start from this assumption that
00:05:58 - the hippocampus represents one solution to the question of where something is
00:06:04 - and but an alternative Bayesian view might be you'd represent a probability
00:06:09 - distribution across multiple solutions.
00:06:11 - I mean, is that something that you entertain as a possibility?
00:06:17 - Yes. So what's been lacking is direct experimental evidence that you might be
00:06:27 - weighing up the different alternatives and seeing place cells firing for both
00:06:31 - different alternatives.
00:06:32 - But in many situations, you can see, for example, as I showed,
00:06:35 - the single firing field of a play cell becoming bimodal when you stretch the
00:06:40 - environment, for example.
00:06:41 - And it's possible that you could
00:06:43 - think of that as part of a system that's trying to estimate location.
00:06:46 - And because it's been predominantly taking evidence from distances to boundaries,
00:06:51 - it's now got a hypothesis that
00:06:52 - has two peaks in it and that you are indeed representing a distribution.
00:06:57 - And in the experiment that I showed where people had to remember where something was,
00:07:02 - The population vector overlap with the stored place cell representation captures
00:07:09 - well the distribution of responses across subjects in that experiment,
00:07:17 - rather than indicating a single location.
00:07:22 - Yeah, but I think Tony's question is also annoying, right?
00:07:26 - Because in some sense, Tony is saying like, well, sort of this spatial cognition
00:07:31 - can happen in many different ways, in many different areas of the brain,
00:07:35 - it becomes a little bit like wishy-washy, like, okay, it's sort of not such
00:07:40 - a unique feature then of a single structure like the hippocampus.
00:07:45 - That's how I take that question. If it's just Bayesian, we integrate multiple
00:07:48 - factors, and so then why worry about hippocampus?
00:07:52 - But I still believe that hippocampus is making a unique contribution to solving that problem.
00:07:57 - Well, I would agree with that. But the question is that, I mean,
00:08:00 - so I think it's definitely true that we have at least two solutions.
00:08:03 - But I'm here to make Neil's life difficult. Just to clarify my position.
00:08:08 - There's a striatal solution to the navigation problem.
00:08:12 - There's something in the hippocampus. There may well be other ways of navigating,
00:08:17 - certainly towards sort of nearby targets and so on.
00:08:23 - So you can almost say, well, in hippocampus, we might be representing multiple
00:08:29 - solutions, and then the decision gets made in some other part of the brain about which way we would go.
00:08:34 - And there are models of hippocampus that are consistent with the idea of representing multiple hypotheses.
00:08:41 - I want Neil to tell me now which one it is. Yes.
00:08:46 - Well i mean as as i showed um there are
00:08:51 - some things that we do know about the brain and this packard
00:08:54 - and magor experiment was again it was
00:08:57 - only a summary of lots of things that people knew up to that point but it
00:08:59 - was very nice clear expression of the fact that different solutions
00:09:04 - could um seem to be supported by different parts of the brain and you could
00:09:08 - actually inject anesthetic into the two different parts of the brain and see
00:09:13 - the different strategies being expressed in behavior so um in answer to tony's
00:09:18 - question i think well and indeed yours that um.
00:09:23 - If i want to characterize the hippocampal contribution to these
00:09:26 - kind of spatial memory tasks you know it it seems to
00:09:29 - be representing a location relative to
00:09:32 - the environment probably as a conjunction of many kinds of cues some
00:09:35 - of which are distances to boundaries in different directions whereas
00:09:39 - other parts of the brain might be remembering more uh you
00:09:42 - know i turned left at the shoe shop or a sequence of turns on
00:09:45 - a well-known route or in other parts of the brain still
00:09:48 - you know a visual visual snapshot matching uh
00:09:51 - this is this looks like where i was before and all
00:09:55 - of these things are happening in parallel and it is an interesting question
00:09:59 - what how they get combined who chooses what and indeed on the input there's
00:10:04 - different kinds of evidence coming in and they are probably waiting in a bayesian
00:10:07 - way with more things that seem to to be more stable over time or more certain,
00:10:12 - being given a stronger weight than things that seem more variable.
00:10:15 - But would it be fair then to say that the hippocampus really is critical in
00:10:20 - generating this allocentric representation of space?
00:10:25 - Yes, I would think so, yeah. But in many of these experiments,
00:10:29 - they're a memory experiment.
00:10:31 - You have to remember a location and go there.
00:10:33 - And we know that the hippocampus is critical for many forms of memory.
00:10:36 - And so in a spatial memory experiment, yes, if you inactivate it,
00:10:40 - then often the person or animal will not perform very well.
00:10:44 - So there's good evidence that the hippocampus is responsible. responsible and
00:10:47 - in that pakada magore experiment you could see that
00:10:50 - it was particularly responsible for one of the two available
00:10:52 - strategies for that task yeah the general hypothesis
00:10:56 - that you're following then is that the there's a population response in hippocampus
00:11:00 - of of play cells which in some way represent the brain's best guess as to where
00:11:06 - it is in the world in an alec-centric coordinate frame in those tasks yes yeah
00:11:10 - okay we're done so now we we have an hypothesis
00:11:15 - about what the hippocampus is doing.
00:11:18 - And it gives you this, an ellicentric representation of space.
00:11:21 - So in world coordinates, which is, which is a fantastic accomplishment.
00:11:24 - And now we have to try to understand how this is done.
00:11:27 - Right. And, and then you build very much your, your theory or model on that
00:11:32 - around, at least in the, in your presentation now around this notion of the boundary vector cells.
00:11:37 - This is really the starting point of your whole analysis, right?
00:11:39 - So, so why, why do you feel that this notion of boundary vector cells is then
00:11:45 - so decisive to understand this ability to generate allocentric representation of space?
00:11:50 - Well, first of all, the head direction cells, I think, are fundamental because
00:11:54 - this overall orientation, sense of orientation, is important for everything
00:11:58 - that comes subsequent to that.
00:12:01 - But taking that as granted, then you need some distances from some things to know where you are.
00:12:10 - And really, I think that if you were to imagine an animal with a sort of range
00:12:16 - finder, looking at the distances to stuff around it in full 360 degrees,
00:12:21 - you know, obviously, the biggest contribution will be to either things that
00:12:24 - are very nearby or things that are quite extended.
00:12:27 - And so topographical features of the environment that are extended are going
00:12:32 - to dominate that kind of representation.
00:12:34 - And I think that that's really the boundary vector cells.
00:12:38 - Cells we heard also and and
00:12:41 - i talked about object vector cells and we heard from them from edvard
00:12:44 - moser you know the
00:12:47 - boundary vector cells would do actually respond to uh
00:12:50 - small objects just they have a very small firing
00:12:54 - field is provoked by a very small object and so
00:12:57 - they're sort of responding like a a range
00:13:00 - finder tuned in a particular direction allocentric direction
00:13:04 - presumably governed by the head direction cells and and
00:13:07 - you know with once you're oriented the next
00:13:10 - thing you need is distance and and these are range finders
00:13:14 - tuned to directions and they seem to
00:13:16 - provide the simplest explanation
00:13:20 - for the sort of broad qualities of the firing fields of place cells when you
00:13:25 - put an animal in different environments environments that differ in shape and
00:13:29 - size but not differ in in texture and odor and and so on enough to make different
00:13:33 - play cells fire in which case you wouldn't be able to try and analyze these different kinds of.
00:13:38 - So then you're saying, look, I have a heading vector. This gives me,
00:13:42 - if you want, that's my compass. That's my directional system.
00:13:47 - And now I glue to that some sensory feature.
00:13:52 - Might be somatosensory. It might be auditory. It might be visual.
00:13:56 - And that now gives me, let's say, a reference in space of how that compass coordinate
00:14:02 - maps onto that specific environment.
00:14:05 - Environment well i should say that in most normal environments
00:14:08 - there'll be a you know a broad range of different
00:14:11 - local cues also which would help you in a very direct way
00:14:14 - to localize yourself and they are probably also important to
00:14:17 - play cells but in most experiments most people
00:14:20 - do they try to control for those cues so that they're only
00:14:24 - the cues that that they're aware of remaining like
00:14:27 - the boundary of the environment these distant orientation cues for the
00:14:30 - head iteration cells and so um given
00:14:34 - that you've tried to remove reliable local queues
00:14:37 - also usually the floor is rotated between trials then
00:14:41 - what you have left you have distances to to
00:14:44 - extended things right but now the head direction system depends on path integration
00:14:50 - path integration is noisy because it's an integrator um so isn't that a little
00:14:56 - bit of an obstacle for for the such which is to keep on working reliably over
00:15:00 - extended periods of time. Yes, indeed.
00:15:02 - So that's an interesting thing that I didn't talk about, but exactly the same
00:15:07 - concerns about how you translate between egocentric and allocentric in terms
00:15:11 - of place cells and sensory inputs, which have to be egocentric.
00:15:16 - Applies to head direction cells.
00:15:18 - So a head direction cell will fire whenever the animal's facing,
00:15:22 - let's say, north. I don't actually mean compass north.
00:15:25 - Wherever it is in the environment. And so actually in all experimental situations
00:15:30 - when they've been recorded.
00:15:33 - That tuning is parallel across the environment, but if they were simply responding
00:15:37 - to sensory cues, you would see parallax, which you do not see.
00:15:40 - It is an allocentric response, the head direction cells. And so again,
00:15:43 - you need to have this translation mechanism between egocentric sensory input
00:15:47 - and this allocentric response.
00:15:48 - And so we think probably retrosplenial cortex, as in the model that I explained,
00:15:53 - has to do that same job for the head direction cells so that they can be reliably
00:15:56 - anchored to sensory input to prevent the accumulation of error that you would
00:16:01 - inevitably get if you're just integrating angular acceleration. Right.
00:16:07 - But that is the sensory cue that you link the heading vector to.
00:16:13 - That in itself might already be an invariant representation of some sensory state.
00:16:18 - It might not be necessarily egocentric. right if
00:16:21 - we go through some some perceptual hierarchy and if
00:16:24 - i climb up that hierarchy i might get to more invariant representations
00:16:27 - of space maybe i have a very abstract representation say
00:16:30 - just okay this is a wall i'm not
00:16:34 - interested anymore in its orientation or its color or
00:16:36 - its size a wall so so how would that notion
00:16:39 - of invariant representations of the sensory feature then actually
00:16:43 - be a challenge to that to that model well i
00:16:47 - think it's um the invariance we're
00:16:50 - talking about uh is exactly you know the head direction
00:16:52 - cells show that very nicely they're at
00:16:55 - the top of the hierarchy for direction they're invariant to all of the parallax
00:17:01 - and sensory stuff that happens as you move around and they extract what would
00:17:05 - be a compass you know and and uh the rest of the system is probably built on
00:17:11 - that and now if it would replace the heading vector with a movement vector?
00:17:15 - Would that change your approach very much?
00:17:24 - I'm not sure what you mean. As opposed to integrating, let's say, my rotations in space,
00:17:29 - I could also say I just interpolate between subsequent movements,
00:17:34 - and this gives me a vector that I might then use to get to some sort of LSM
00:17:40 - representation of space.
00:17:42 - Well, so there's two separate things here. One is how do the head direction
00:17:45 - cells fire to encode the orientation of the head. If you then move on to path
00:17:49 - integration and perhaps what drives grid cells to fire, then you have to integrate
00:17:53 - movement rather than head direction.
00:17:55 - So you need to know which direction you've moved in, not which direction your head is facing.
00:18:00 - Exactly. So would that make a big difference to your proposal?
00:18:05 - Not for the boundary vector cells, I don't think. Because usually the head of
00:18:10 - the rat is... Because the boundary vector cell model of place cell firing is
00:18:14 - not path integration. It's just feed-forward sensory input.
00:18:17 - What direction from the head is the boundary feed forward.
00:18:21 - So there's not an issue there. When you come on to how you do path integration,
00:18:25 - how you integrate your own bodily movement to estimate where you are,
00:18:28 - which may well happen through grid cells,
00:18:31 - then indeed you need to know your movement direction, your movement vector,
00:18:34 - not your heading vector. That's right. Perfect, yeah.
00:18:37 - And so there's various hypotheses as to where this signal comes from because
00:18:41 - it can't be just the head direction. That's right.
00:18:45 - Multiplied by running speed. has to be movement direction. Would it be your grid cells then?
00:18:51 - No, what we're talking about are what are the inputs that would go to something
00:18:55 - like grid cells that is the movement signal that is integrated.
00:19:00 - If I would just look at subsequent grid cell responses, I could infer a movement vector.
00:19:07 - Yes, you could do. So that's what I was after. Yes,
00:19:09 - so I think maybe this question came up
00:19:12 - with Edvard Moser's talk that whether you could
00:19:16 - look at it either way if you think grid cells know uh
00:19:19 - how to fire because they're doing path integration you need
00:19:22 - an input which is a movement vector or you could say well
00:19:25 - grid cells fire as they do in which case a movement
00:19:28 - vector could be an output yes that's it yes exactly indeed it could be but then
00:19:32 - you're still left with a question of how the grid cells fired in that way in
00:19:37 - the first place right the um so the the activation of the play cells you have
00:19:43 - of being driven by the boundary vector cells.
00:19:46 - And so it's basically a thresholded sum of the inputs from the boundary cells.
00:19:54 - Clearly, you could choose the weighted sum of the inputs as just the first thing to try.
00:20:00 - Are there other motivations for doing it that way, for example,
00:20:04 - from environment warping experiments that say that it's the weighted average,
00:20:09 - it's the best thing to use?
00:20:11 - Well, it was the simplest thing, the first thing to try.
00:20:14 - But in some experiments as you um expand the
00:20:17 - box you have a place field which slowly reduces in
00:20:20 - firing rate and disappears right which implies a
00:20:24 - sum with a threshold you know as the
00:20:26 - over initially overlapping inputs from opposing
00:20:29 - walls uh move apart then the the bit you've got left slowly falls below threshold
00:20:35 - and the cell stops firing so there's there's good reason to use a thresholded
00:20:40 - sum there's a reason for a threshold and um we really did see um in some cases individual,
00:20:48 - sub peaks being moved apart which means you have
00:20:51 - to sum the two rather multiple if you multiply them then
00:20:54 - you get for example a bayesian you know average position
00:20:57 - in between the two peaks as they well this is what i was
00:21:00 - getting at is there support for multiple hypotheses in some
00:21:02 - of those well um just to
00:21:05 - finish answering question there's there's clearly adding these
00:21:08 - inputs from the different walls and having a threshold was the
00:21:12 - most obvious thing to do given all the things we've said um so if you were trying
00:21:18 - to combine um hypotheses probabilistically you would probably multiply and that's
00:21:24 - not what we saw right um it's interesting because of these experiments by um,
00:21:30 - galasell and cheng and also by uh cartwright and collett um john o'keefe in
00:21:36 - this original stretchy box experiment was expecting to see.
00:21:40 - Cells, place cells, firing.
00:21:43 - In the constant ratio of the distances across the box, rather than the sub-peaks
00:21:50 - being pulled apart and sticking to the fixed absolute distance from the walls
00:21:55 - as they were pulled apart. And so...
00:21:58 - If what he'd been expecting had happened, that would have been an argument for
00:22:01 - multiplying and having the combined estimate of two probabilities distributions,
00:22:06 - but that isn't what we saw.
00:22:07 - But it sounds like what the experiment says is a bit more sophisticated than the averaging.
00:22:15 - So there might be something to look at there in
00:22:18 - the future in terms of how in this inconsistent
00:22:21 - consistent environment that's you know where uh the
00:22:25 - boundary effect cells don't all point to
00:22:28 - the same place in space then you might start to represent
00:22:30 - multiple hypotheses in terms of place cell activity
00:22:33 - well i don't i i don't know if there's
00:22:36 - evidence for that i think there is definitely evidence for um
00:22:40 - hypotheses about what happens in
00:22:43 - a place you know place cell a place cell
00:22:46 - might fire in the same place trial after trial but its
00:22:49 - firing rate might vary quite strongly according to
00:22:52 - other hypotheses to do with sensory stimuli
00:22:56 - that are present whether objects are nearby smells how
00:22:59 - fast you're running so I think there's a lot of
00:23:02 - orthogonal hypotheses to place which also modulate the firing rate but it more
00:23:10 - like a gain field representation okay so the actual firing field itself uh isn't
00:23:16 - obviously doing something nice and bayesian about estimating location.
00:23:21 - Uh when you do these weird experiments when you deform the environment which
00:23:25 - are rather unnatural but they probably are combining the cues to them such as
00:23:30 - we haven't talked about the path integration input and the sensory input are
00:23:34 - probably being combined in a more bayesian way yeah and also as you say maybe
00:23:38 - there's something happening in time you know that,
00:23:41 - you can swap between hypotheses over time and explore different interpretations.
00:23:45 - Well, actually, on that point, if you do this stretching experiment and you
00:23:50 - get a single place field stretching into two place fields as you've expanded the box,
00:23:56 - then you notice that each place field fires a bit more according to running direction. So the...
00:24:04 - The place field that appears to be attached to the wall behind the animal fires
00:24:08 - a bit more, so that when you're running in one direction, one peak will be higher,
00:24:12 - and when you're running in the other direction, the other peak will be higher,
00:24:14 - which we interpreted as being this path integration input,
00:24:18 - adding to this environmental sensory input, and being stronger if you just run from a wall.
00:24:25 - Obviously, you have a good estimate of how far away it is from path integration,
00:24:28 - so that peak gets a bit higher because it's got a stronger path integration
00:24:32 - input but down the peak that's attached to the opposite wall that you're running towards.
00:24:36 - Yeah, yeah. But I find it very interesting.
00:24:39 - You guys seem so, let's say, motivated to interpret this in Bayesian terms.
00:24:45 - Well, if you look at the dynamics of CA3, CA1, where the real action in the
00:24:49 - end happens, it's also highly competitive, and in the end, it's relatively sparse.
00:24:54 - You can also think about it much more as, let's say.
00:24:57 - The endpoint selector that sits on top of some Bayesian integrator as opposed
00:25:02 - to being a Bayesian integrator itself because you just don't have the dynamics
00:25:06 - for it because it's rather sort of selective,
00:25:10 - rather sparse, and these are not the features you want in your Bayesian integrator.
00:25:14 - So is it not fair to place that bit outside of the hippocampus, gentlemen?
00:25:19 - No, I think that's fine. As I said initially to Tony, there isn't good evidence
00:25:25 - of multiple hypotheses being entertained at the same time.
00:25:29 - It would be a good way for the system to work but it may well not do that.
00:25:35 - And then can't we not explain the elongation of the place field also in terms
00:25:40 - of a perceptual learning process because the animal starts also carve out specific
00:25:47 - trajectories through that space.
00:25:49 - It's just not moving randomly, right? And the cells you
00:25:52 - measure from is probably also on trajectories the animal
00:25:55 - will visit more frequently than other trajectories so it
00:25:58 - will actually be moving more often from that specific
00:26:01 - initial position to a final position and therefore expose itself more to the
00:26:08 - specific sensory features that are already associated with that place field
00:26:12 - and therefore strengthen the specific intra-place field associations and then,
00:26:18 - if you want, elongating its response.
00:26:20 - Well, there are interesting trajectory effects on place cell firing,
00:26:24 - but they're rather different than these effects in this unusual situation where
00:26:28 - we change the shape and size of the environment.
00:26:30 - I mean, I think that's an unusual experimental manipulation that shows us something about the inputs.
00:26:35 - However, the trajectory effects are interesting so that after... So if the...
00:26:41 - The animal's always running in the same direction. You see that firing fields
00:26:44 - tend to elongate backwards a little bit along the track. So my anchor, Mater, noticed this.
00:26:51 - But that's been interpreted as perhaps the play cell beginning to fire in expectation
00:26:56 - of getting to its, quote, true place, original place.
00:27:00 - And there's some interesting theoretical work by Matt Botvinick and some co-authors
00:27:06 - suggesting that these play cells might encode a successor representation for
00:27:14 - reinforcement learning,
00:27:15 - which is an idea from Peter Diane in the 90s, that if you were to tweak your
00:27:20 - state representation to actually include the likelihood of eventually ending
00:27:25 - up at that state, even when you're at other states,
00:27:28 - then that enables you to estimate future reward at a given location much more efficiently.
00:27:37 - And so it's possible that if you have repeated trajectory, trajectories,
00:27:40 - then you can build a little bit of the probability that you now know that if you're here,
00:27:46 - you're likely to end up somewhere else into the state representation so that
00:27:51 - you can then rather more easily estimate likely future reward from a given location
00:27:57 - because you've built in the transition probabilities,
00:27:59 - which are now not uniform because I always run in this direction when I'm in this place.
00:28:03 - So it's the kind of model that reduces the whole brain into the hippocampus.
00:28:07 - Well, no. I mean, it may be that that kind of representation,
00:28:10 - you know, perceptual learning and so on, affects, you know, it could explain
00:28:13 - what happens in lots of parts of the brain, not just hippocampus. Right, okay.
00:28:17 - But now, so in your physiology, so then with that model, you went to look at the physiology,
00:28:21 - and then you show that these sort of vector border cells are to be found both
00:28:29 - in the subiculum and the entorhinal cortex, which is interesting, right?
00:28:33 - Because if you look at the overall loop, we start in entorhinal,
00:28:37 - we go to the dentate, gyrus, CA3, CA2, CA1, then subiculum, and then we go out to entorhinal cortex.
00:28:43 - So isn't it really strange that we find these cells actually at the input stage
00:28:47 - and the output states of that whole loop?
00:28:49 - It's like we're wasting all this circuitry in between to do nothing,
00:28:52 - and we just have our border vector cells.
00:28:55 - It is. It is strange. I mean, it is a loop, and we may be interpreting how it
00:29:00 - works incorrectly when we think of entorhinal as the input and subiculum as the output.
00:29:05 - I mean, there's plenty of connections from subiculum back to entorhinal,
00:29:07 - and presubiculum has very strong connections to entorhinal.
00:29:13 - My colleague Colin Lever is always keen to point out that something that Pat
00:29:18 - Sharp originally noticed, that when you get a complete remapping of place cells
00:29:22 - in CA1 between two environments,
00:29:24 - there's actually spatial responses in subiculum that are nice and stable that
00:29:29 - you can record, and they show no remapping at all.
00:29:32 - And yet the standard model would be that the major input to subiculum is the
00:29:36 - output of CA1, and yet you've got this complete change in CA1 representation, no change in subiculum.
00:29:42 - Them so yes it is a puzzle you know maybe
00:29:45 - subiculum is the input and it goes around to enter
00:29:48 - and then into you know who knows but it is a puzzle
00:29:50 - it's puzzling from the anatomy you wouldn't expect it no but
00:29:53 - it's quite a challenge there's something there that but it
00:29:56 - is a challenge to your model right because you want to say these border vector
00:29:59 - cells are if you want a representational primitive on which i built everything
00:30:03 - else but now we see whatever wherever you want to put the input and the output
00:30:06 - they still sit there well and and what is also true if you you know you You
00:30:11 - could also ask me the sort of functional difficult question would be the place cells are a basis set.
00:30:18 - You know, it could be that boundary vector cell firing responses are made out
00:30:23 - of weighted sums of place cell firing. That could be possible.
00:30:28 - But now, if you look at subiculum and entorhinal, how big a proportion of cells
00:30:33 - in those areas would actually show these properties?
00:30:37 - So I think it's around 20% or so is Colin Lever's best estimate in subiculum,
00:30:43 - and it's a bit less, maybe more like 10%, but still quite common,
00:30:46 - these border cells that Edvard Moser described, i.e.
00:30:50 - The boundary vector cells that fire quite close to the boundary. boundary
00:30:53 - they're quite common there have been a few reported
00:30:56 - that fire at a distance but it's a handful and so that's
00:31:00 - another another strange yeah we that remains to be um the exact relationship
00:31:06 - between the border cells and the boundary vector cells and the object vector
00:31:09 - cells is still not quite clear okay so but then could we argue that in entorhinal cortex there's if you
00:31:17 - want a rank order of cells and in terms of how foundational they are.
00:31:22 - You could argue, well, grid cells might be very foundational.
00:31:27 - Single modality cells in lateral entorhinal cortex might be foundational.
00:31:31 - And now I start to merge them. I start to merge them in your border vector cells.
00:31:35 - So because now I have within entorhinal cortex, I have access to both movement
00:31:40 - vectors, grid cells, and have access to sensory features of the world,
00:31:43 - lateral entorhinal cortex. and I just have a subpopulation of cells that is
00:31:47 - associating the two, and now I have my border vector cells.
00:31:52 - Yes, the grid cells are also a basis set, and you could form any responses out of them.
00:32:00 - So it's a good question, although it's not quite clear what foundational means in your question.
00:32:05 - And so we have looked developmentally. So Francesca Cucci in John O'Keefe's
00:32:10 - lab was looking at this and also Edvard Moser's lab at the same time.
00:32:17 - And what you see during development as pups learn to start to crawl around and
00:32:22 - open their eyes and so on is that the head direction cells are foundational
00:32:25 - in the sense that they seem to be there as soon as you can record them.
00:32:28 - There's also theta rhythmicity as early as you can record them in this situation.
00:32:33 - Then you also see place cell responses as they start to crawl around,
00:32:37 - the rat pups, and they get better with experience over several weeks.
00:32:43 - So there's a slow improvement in the spatial tuning of the play cells.
00:32:48 - Grid cells are not present until four or five developmental days after you see these other cells.
00:32:55 - And so they're not foundational in that sense, it seems.
00:32:59 - And this has led to people suggesting that maybe
00:33:02 - the grid cells require some stable spatial input maybe from
00:33:05 - play cells to wire themselves up to fire
00:33:08 - properly i mean uh the
00:33:11 - head direction cell seems to be a very strong attractor network
00:33:14 - so that the the head direction cell responses are
00:33:18 - mutually coherent with each other even if they are drifting a little
00:33:21 - bit early on in development it may be the same with the grid cells that they're
00:33:24 - mutually connected well so that it's an attractor it's just drifting so much
00:33:28 - relative to the world that you can't ever record it that's a possibility and
00:33:33 - then you only see it when it gets attached to stable sensory inputs because
00:33:36 - then you can average over movements across the environment.
00:33:39 - But yeah they're all interesting questions and you can try the good
00:33:42 - thing about this field is that you can try to answer them with experiments absolutely
00:33:46 - and the developmental trajectory you're talking about there is based on the
00:33:50 - chemical properties of the cells not not their functional properties no no no
00:33:54 - sorry the function this is recording um uh the the the spiking activity as
00:34:00 - pups first start to move around as they start to explore around the nest.
00:34:05 - Francesco will give a talk about this tomorrow, I think. Right.
00:34:09 - But now the object vector cells, this is the next complexification that also
00:34:14 - Edward talked about yesterday.
00:34:16 - You see them as a further variation on that same theme, or should we now be
00:34:21 - really shocked and say, oh, this is again completely different?
00:34:25 - And oh well they're obviously closely related to
00:34:28 - boundary vector cells in the sense of firing at a allocentric displacement
00:34:32 - vector from something and the only thing that's different is that they
00:34:35 - seem to be specifically tuned to small objects right
00:34:39 - so a boundary vector cell as i said will fire to an extended object but will
00:34:41 - also fire to a smaller object but just with
00:34:44 - a very small firing field because uh obviously it
00:34:47 - only has to move a little way and then it's not in in the receptive
00:34:50 - field for that cell whereas an extended boundary will produce a long stripify
00:34:54 - the object vector cells seem to be different in that they uh don't seem to respond
00:35:01 - to the extended boundaries of the environment they just respond to objects put
00:35:05 - within it now it could be that over time uh.
00:35:12 - There's something to do with the object being novel and the boundaries being
00:35:15 - familiar means that it discriminates because the same object vector cell will
00:35:19 - respond similarly to different objects.
00:35:21 - It's not specific, it seems, to specific objects, but will fire to any small
00:35:27 - object, but not to an extended boundary.
00:35:29 - But so far, the experiments, I don't think, have really been done to see whether
00:35:33 - if you put a novel extended object in that looks a bit like a boundary,
00:35:37 - whether those cells will fire or not. Exactly, that's the question, right?
00:35:39 - When does an object become a boundary? The reverse experiment has sort of been
00:35:43 - done in that Bruno Poussey's group had showed that some wine bottles put inside
00:35:50 - one of these boxes, when you're calling place cells,
00:35:53 - typically don't cause much place cell firing.
00:35:56 - Although we know now from Jim Nerium's work that a small number of place cells
00:36:00 - will fire relative to these, indeed, looking like object-based cells.
00:36:04 - But when they placed three wine bottles in a row together, they started to have more influence.
00:36:09 - Probably because of an extended thing, they would be driving more boundary vector
00:36:12 - cells. But the same kind of experiment hasn't been done with these object vector cells yet.
00:36:16 - But the interesting thing about
00:36:18 - now these object vector cells is that they appear in CA1, apparently.
00:36:24 - Well, a very small proportion of CA1 cells, yes.
00:36:27 - Yes, but Edvard was reporting a medial entorhinal and some interesting sort
00:36:33 - of memory object, well, memory times object cells had been reported in lateral entorhinal.
00:36:40 - But can't we, but these are highly plastic circuits and they are exposed to
00:36:46 - multiple, let's say streams of modalities and submodalities.
00:36:50 - So if we just combine these two considerations, is it then not sort of predictable
00:36:56 - that these kinds of combinatorial or conjunctive encodings emerge rather naturally.
00:37:04 - Well, yes, but before we proposed them from our analysis of play cell firing, nobody knew that.
00:37:16 - An offset vector and essentially offset vector was one of
00:37:19 - the ingredients that would get combined with boundary detectors or
00:37:22 - whatever and until the moses recorded grid cells
00:37:25 - nobody thought there's this sort of furrier like representation of space would
00:37:30 - would exist or needed to exist and so yes in retrospect but before the fact
00:37:35 - no no i didn't want to trivialize it i just want to say this might mean that
00:37:40 - if you now go for some ambiguous border object object and for sure
00:37:45 - you'll find a cell that'll respond to it because we look at this combinatorics.
00:37:49 - Maybe, but I mean, the interesting thing, I think, as Edvard pointed out about
00:37:53 - the hippocampus and entorhinal cortex is that coming from the point of view
00:37:58 - of a Hopfield model for memory,
00:38:00 - you might expect a random distributed binary code that was uninterpretable.
00:38:05 - And in fact, you see these incredibly clear, discrete responses,
00:38:09 - say to compass direction or to location, or a Fourier-like like grid or,
00:38:17 - you know, a boundary vector.
00:38:19 - Now there are, as Edvard said, you know, you do get conjunctive directional
00:38:22 - grid cells, for example, and there may be other conjunctions,
00:38:25 - but the overall impression is that, and there are plenty of spatially modulated
00:38:30 - cells that haven't been characterized also in entorhinal cortex.
00:38:35 - They're probably now less than, you know, they're probably a minority of all
00:38:40 - the cells because so many cells have been characterized. But still,
00:38:44 - the overriding impression is of these incredibly discrete types of encoding being present.
00:38:50 - And sure, they are present in some combinations also, but overall,
00:38:54 - you wouldn't say this is just a mush of everything.
00:38:56 - You would say, Jesus Christ, why are those such specific responses there?
00:39:00 - But as a mathematician, it must appeal to you to perhaps to find some theory
00:39:05 - that explains the emergence of these different kinds of cell types,
00:39:09 - which are particularly useful for navigation.
00:39:12 - And then to be able to perhaps generalize that and say what further kinds of
00:39:17 - cell types we might be able to predict that we should see, perhaps because they're
00:39:22 - adding more orthogonal information to the mix,
00:39:26 - perhaps sensory cells, for example.
00:39:29 - And I think one of the things you alluded to, and also Edvard in his talk,
00:39:32 - was how the grid cells have this sort of change in scale, which follows a sort
00:39:38 - of mathematical law, which is consistent with efficient coarse coding.
00:39:42 - So it looks as though there is, again, something which doesn't look at all accidental,
00:39:48 - but must be the result of this system responding in some very effective way
00:39:55 - to choose the right inputs for mapping space.
00:39:59 - Yes, I think when you come to the grid cells and the discrete jumps in scale.
00:40:05 - That is clearly well arranged for large-scale navigation.
00:40:11 - You get a big combinatorial power for the range over which you can encode locations
00:40:18 - with that kind of representation.
00:40:20 - So what is your prediction for the next kind of cell type?
00:40:25 - Um...
00:40:29 - And it's funny, I don't think the field at the moment is working that way.
00:40:35 - It's true that we predicted boundary vector cells from looking at how place cells responded.
00:40:39 - And going back, that argument, that way of thinking was certainly used by O'Keefe and Nadel.
00:40:45 - And John O'Keefe, having found place cells, predicted there should be directional
00:40:49 - cells and something like path integration or distance cells.
00:40:53 - And so early on, having found place cells, he did predict the head direction
00:40:58 - cells and then something to do with distance.
00:41:00 - And they have been borne out. And something to do with vectors,
00:41:04 - it seems like, is necessary. And we have boundary vector cells and object vector cells now.
00:41:09 - I think the interesting thing, I mean, so there's already a lot of ingredients
00:41:12 - there for making quite a sophisticated system for space.
00:41:15 - And at the moment, I think people are thinking more, how could all this stuff
00:41:19 - we've learned about space tell us about more general memory properties of the hippocampal system?
00:41:24 - That you know how does it store other information which
00:41:28 - which we know that the human hippocampus is required for all
00:41:30 - sorts of forms of memory and i think that's interesting
00:41:34 - so i haven't predicted any more kinds of
00:41:37 - spatial cells i must say but that's particularly what i was getting at you know
00:41:40 - sort of human episodic memory which is more than about a lot of things more
00:41:45 - than space um and where it's it's less easy sort of coming at it from the geometry
00:41:52 - or whatever to say what kind of basis functions you would want to have.
00:41:56 - So we might look at some mathematical way of finding appropriate basis functions
00:42:02 - for learning about episodes in time.
00:42:04 - Yes. Well, so I think it's very interesting that the cells in the human hippocampus,
00:42:10 - most famously Jennifer Aniston type cells, are.
00:42:15 - They're a local tuning curve to a concept, in this case of Jennifer Aniston,
00:42:20 - among all famous people or people you might have heard of.
00:42:24 - And a cell will fire to that concept, whether it's a written name or however you bring it to mind.
00:42:30 - And other cells will fire for other concepts, be they spiders or semantic well-known
00:42:37 - landmarks or famous actors.
00:42:39 - And so you might indeed see
00:42:43 - the sort of locally tuned place cell response as very
00:42:46 - similar to these locally tuned responses in semantic space and
00:42:50 - this very interesting um stretchy bird experiment by
00:42:53 - tim behrens and his group showing what looked like grid-like responses in in
00:42:58 - humans to semantics two-dimensional spaces and uh in rats to one-dimensional
00:43:05 - continuously varying tones showing play cell-like responses from David Tank's group,
00:43:11 - do imply that these principles that have been easiest to spot in space and freely navigating animals.
00:43:21 - Might well apply to all sorts of other kinds of information and how it's organized
00:43:24 - and how it's retrieved and so on. And then that would obviously be a...
00:43:29 - Starting with these very well-characterized spatial responses,
00:43:31 - I think, is a good place to try and understand everything else,
00:43:35 - which is obviously much harder to get a grip on.
00:43:39 - So, I mean, that would suggest that one thing we might look for is sort of compression-type
00:43:43 - algorithms, which will give us a low-dimensional.
00:43:47 - Or relatively low-dimensional description of the current context,
00:43:52 - which we can encode then in entron and cortex and use to run our sort of location
00:43:59 - finder or memory finder algorithm.
00:44:01 - Is that a good metaphor? Well, the grid cells certainly appear to be a compressed
00:44:08 - code like a Fourier code or something that would be very efficient compared to the place cells,
00:44:13 - which in their own way are very useful for tagging specific information to specific places,
00:44:18 - but are less efficient for covering large areas in a metric way where you can
00:44:25 - infer the vector between any two parts.
00:44:28 - If I have a place cell that fires in one place and another one that fires in
00:44:31 - another place with no overlap, it's hard to know how they're related to each other.
00:44:35 - Whereas with grid cells, you can infer the vector between two grid cell codes
00:44:42 - for two different locations. And so that's very powerful. It's a compressed representation.
00:44:46 - You know, it looks a bit like a Fourier pattern. But equally,
00:44:49 - Dory Durdickman's group has pointed out that if you do PCA on the place cell
00:44:54 - input that you get as your animal is running around, then you get things that look like grid cells.
00:45:01 - So indeed, the grid cells might be doing a PCA of this place representation.
00:45:06 - And there's a similar argument to be made for the successor representation that
00:45:12 - I mentioned for place cells in that the eigenvectors of the successor representation
00:45:17 - also look like grid cells.
00:45:19 - And so they're clearly related, the.
00:45:22 - Eigenvectors of the covariance matrix being the same as the
00:45:25 - pca it could well be that the grid cells are a
00:45:28 - compressed way of arbitrarily representing
00:45:31 - any of this location coded or you know single peaked tuning curve uh grandmother
00:45:37 - type cell response uh representation of arbitrary information which is very
00:45:42 - useful for then associating things to that information in the hippocampus yeah
00:45:47 - so i think if i maybe put words into your mouth but you're agreeing that
00:45:51 - the entorhinal is a compressed representation of the context,
00:45:55 - and then that's expanded out in CA1, CA3,
00:46:00 - perhaps after being sparsified in dentate gyrus, as often as that people have.
00:46:04 - Well, you know, David Marr's model of hippocampus remains the best model of hippocampus in memory.
00:46:14 - You know, it's probably surviving that and his cerebellum model and surviving
00:46:18 - really well, maybe even better than his later work in vision.
00:46:21 - Um and yes those ideas hold a lot of currency although it's important to remember
00:46:27 - entorhinal cortex is not just grid cells so even within medial entorhinal cortex
00:46:32 - it may be that head direction cells are more numerous than grid cells and there
00:46:36 - are also these border cells.
00:46:38 - Conjunctive cells and spatially modulated but not grid cell cells so you know
00:46:45 - yes for the for the grid cells they look like compressed code uh there's other
00:46:48 - things going on also uh sensory inputs to place cells,
00:46:51 - we think, not just the path integration type inputs.
00:46:54 - But aren't you, first you guys are all swept away by Bayesian,
00:46:57 - and now you're all swept away by, let's say, compression, which in some sense
00:47:02 - is already happening to a large extent outside of Hippocampus.
00:47:05 - And earlier, I thought we had agreed that Hippocampus is contributing to constructing
00:47:09 - allocentric representations.
00:47:11 - So, let's not worry about compression then, but now you guys are all worried about compression.
00:47:16 - So, what is it? Compression, allocentric, or is
00:47:19 - it both what do you want uh well you're
00:47:23 - comparing apples and oranges i i don't i think you
00:47:27 - can have an allocentric or an egocentric representation and you
00:47:29 - might want it to be compressed or not so we i don't
00:47:32 - know we can compare allocentric egocentric no no you have no no what i'm saying
00:47:36 - compression or allocentric is it both is it combined is one building on the
00:47:40 - other are these independent orthogonal interpretations well so so the those
00:47:45 - codes the head direction cells and the place cells and grid cells appear to
00:47:49 - be allocentric in the sense that...
00:47:53 - The reference to the world, it's which way you're facing in the world,
00:47:56 - where you are in the world.
00:47:57 - But they are where the animal currently is, which you might say is egocentric.
00:48:05 - It's referring to the animal's own location or orientation.
00:48:07 - Having said that, I think then the grid cells look like a compressed version of the play cells.
00:48:13 - So it may be that an expanded representation is useful for attaching things
00:48:17 - to, and that might be the play cells.
00:48:19 - And a compressed representation might be useful for generating large-scale metrics
00:48:26 - within which to compare or know the relative positions of lots of information.
00:48:32 - About 10 years ago, we showed that grid cells actually carry more information
00:48:36 - about space than place cells.
00:48:37 - So if you do position reconstruction from grid cells versus place cells,
00:48:40 - you have more accuracy from the grid cells.
00:48:42 - So that would be then inconsistent with the idea that they compress the place
00:48:45 - cells because then the place cells would have more information for position
00:48:48 - reconstruction. Are you comparing the same number of place cells and grid cells?
00:48:52 - Yeah, this was all control.
00:48:53 - Yeah, no, well, exactly. So if you have a sparse representation,
00:48:56 - which is easy to attach things to, then obviously it's more efficient. It is compressed.
00:49:01 - You can have a coverage of a larger environment with the same number of grid
00:49:06 - cells. That's why it's a compressed representation.
00:49:10 - Okay, look. But now the other... So this is a controversy we saw,
00:49:18 - which we will solve later,
00:49:19 - but now you presented a model where you wanted to explain how now you could
00:49:24 - use allocentric representations to also reconstruct, if you want, imagined locations,
00:49:31 - how you could use imagery to sort of make now predictions that would be relevant for navigations.
00:49:38 - What's the contribution of that model now to our understanding of the hippocampus?
00:49:44 - Well, that model was trying to take these sort of spatial things that we recorded
00:49:49 - in and around the campus and say, how could they apply to human episodic memory?
00:49:53 - And, you know, the experience of human episodic memory is reliving the event.
00:50:00 - And in terms of the visual content, it's imagining the scene.
00:50:05 - And so the proposal was that this spatial system is a way of pulling out the
00:50:10 - stored allocentric or abstract information that you have in your temporal lobes
00:50:14 - to create a coherent spatial scene from a single location with a single direction,
00:50:21 - you know, provided by the head direction cells and the place cells respectively.
00:50:26 - And then how you could project that information into an egocentric frame so that you can imagine it.
00:50:30 - So that's a necessary thing that has to happen if we want
00:50:33 - to make contact with with human imagery and so
00:50:36 - that's why i put it in the model um in principle you could solve vector navigation
00:50:43 - just using the the grid and place cell model and maybe do that you need this
00:50:50 - extra stuff to be able to sort of visualize uh but of course that gives you
00:50:54 - know extra functionality you might Right.
00:50:56 - In principle, if you've got a lot of stored abstract information,
00:51:01 - you could pull it together to give you the viewpoint, if you like,
00:51:09 - on all of that information that's consistent with being in a single location.
00:51:14 - And that's a very powerful way of reducing the enormous amount of stored information.
00:51:20 - You're just going to pull out all those bits that are consistent with me being
00:51:23 - here, possibly facing this direction.
00:51:25 - Now what what is around me and if
00:51:28 - you think conceptually you've got
00:51:31 - a whole lot of stored information a whole sort of life's worth of experience
00:51:34 - if you like uh if i ask you about jennifer
00:51:37 - aniston you probably want to pull out just the information relative to
00:51:40 - to that location in in the conceptual space
00:51:43 - of all actors or people that you've ever seen and so
00:51:47 - uh you know it's more that these are
00:51:52 - ways in which memory could function you have to consider the retrieval
00:51:55 - process you could store an awful lot of information but
00:51:58 - how do you retrieve it and episodic memory seems to a very specific way of retrieving
00:52:02 - information you impose a particular location and perhaps a particular direction
00:52:08 - to interrogate your stored data
00:52:12 - with but now in the model itself gives you like a lookup table, right?
00:52:17 - You can sort of project back from the egocentric views to the allocentric views.
00:52:21 - And to what extent is the model then able to capture also the physiological
00:52:25 - and anatomical characteristics of this hippocampal system? Yeah.
00:52:31 - Well, it's quite a high-level model because it's trying to deal with cognition,
00:52:35 - and we know so little that it would be inappropriate to try and build too much
00:52:42 - physiological and anatomical detail in from the start.
00:52:45 - But this translation circuit, as shown by Alex Puget and various other people,
00:52:51 - corresponds to the observed parietal gain field neurons that have been recorded
00:52:56 - and should function as a translation circuit irrespective of what it is you're retrieving,
00:53:05 - from allocentric or egocentric coordinates and translating into the other.
00:53:09 - I'm not sure if that answers your question, actually. Well, because it's also
00:53:13 - a bit a step towards the experiments you did with human subjects where you then
00:53:17 - start to match the response of the model to fMRI experiments with human subjects
00:53:23 - that were in some sort of navigational test.
00:53:26 - So how well was the match then between the fMRI results and the predictions of your model?
00:53:31 - Well, in terms of the fMRI, it was more that we could have, to most of the areas
00:53:39 - of activity, the areas in the brain that showed increased metabolic activity
00:53:43 - during this kind of task,
00:53:45 - we could ascribe some kind of putative function.
00:53:48 - It's very speculative, but the model says what that activity should be actually
00:53:52 - doing. you know in the hippocampus the play style should be doing this and in
00:53:55 - retrospinal cortex it should be a translation circuit or whatever and so it's
00:54:00 - not really that there was a um.
00:54:03 - Particular the model wasn't good enough
00:54:06 - to constrain um what activity
00:54:11 - you should see it's more that my choices of
00:54:13 - where these bits of the model should be in the brain were broadly
00:54:17 - borne out and they give us some kind of way of trying to understand what that
00:54:20 - activity might represent in terms of mechanism so the model would be as heuristic
00:54:24 - to help you not to further think it helped to interpret those results it wasn't
00:54:28 - that we predicted results and you know We could measure the error between the
00:54:33 - predicted results and the actual results.
00:54:35 - But now, in terms of the performance, so you had humans in a virtual reality
00:54:41 - environment that they navigated around.
00:54:45 - There was a target that they had to sort of memorize.
00:54:47 - So we did do a behavioral experiment where we did match the behavioral predictions
00:54:53 - of the model and the behavior of the people.
00:54:55 - And in that case, yes, of the models we tried,
00:54:59 - the boundary vector cell place cell model was the best match to the distribution
00:55:04 - of responses across participants of where they thought this thing was when you
00:55:09 - change the shape and size of the virtual room.
00:55:11 - Right. But we discussed this also during your talk which is I think an interesting
00:55:17 - challenge now because you basically took the response of all the subjects.
00:55:22 - The subjects had to estimate where they were relative to this target.
00:55:26 - So this gives you distribution over all estimates over all subjects, right?
00:55:31 - And then you show that the model,
00:55:32 - your model could capture well the distribution over that whole group.
00:55:38 - But now if we split it out to the individual level, and also you showed that
00:55:41 - data, you actually see that the individual distributions can have a rather different shape.
00:55:47 - Actually, it's clustered in, let's say, three different groups.
00:55:50 - Much less variance within subject than between subject.
00:55:54 - So then in the end, what are we modeling, right? Are we modeling statistically
00:55:58 - over populations, or do we have to take into account, must we be able to also
00:56:06 - account for these individual differences?
00:56:08 - Because you saw that the subjects that were accurate had a pretty nicely clustered set of responses.
00:56:14 - And indeed, the ones who were inaccurate. Well, to be fair, actually,
00:56:18 - in that experiment, because you've changed the environment, there's no right or wrong, in fact.
00:56:23 - You know, you can say... They had a task. They had a task to put their marker
00:56:28 - where they thought the flag was.
00:56:30 - That's it. But the encoding environment had changed.
00:56:32 - Sure. So it's a bit of an arbitrary question, and they did their best.
00:56:35 - And some people, they made different guesses.
00:56:38 - But to answer your question about what the model was showing,
00:56:40 - I think the model was showing that as a population,
00:56:44 - people had a vague idea where things were that related to the walls of the environment,
00:56:49 - and the model captured some aspect of that but didn't capture the individual
00:56:55 - differences by any means at all.
00:56:57 - But it was interesting that as a group, they kept within this sort of very broad framework.
00:57:06 - Distribution that the the model could show i mean we could have
00:57:08 - cranked up the model and said just sliced off the very top uh
00:57:11 - likelihood match for location and
00:57:15 - that would have been you know a very narrow uh distribution
00:57:19 - and some subjects put their responses there and others put them miles away and
00:57:23 - some subjects were also confused right did unacceptable things but looking at
00:57:27 - the at the distribution across the subjects we see three groups right there's
00:57:30 - one group that say let's call them accurate and they're they're at the northeast
00:57:34 - corner of the environment close to the wall.
00:57:36 - The second cluster is sort of southeast, so close to the wall,
00:57:40 - but they're like reversed, right? They've mirrored, if you want that space.
00:57:44 - And the third group is completely off in the opposite corner.
00:57:47 - So is there a set of parameters in your model that would be able to then tune
00:57:52 - the model to capture each of those clusters of responses accurately?
00:57:55 - Well, I think that there was one participant that was off in the corner.
00:58:01 - I think we have to just ignore, who knows what
00:58:04 - they were thinking about then there's two other groups of
00:58:07 - responses and indeed they're not uh one is within
00:58:10 - to see the blue one there is actually some they've responded in
00:58:13 - both um groups um but
00:58:16 - it seems like of these subjects more of them
00:58:18 - because the original location was nearer to
00:58:21 - the north wall and the south wall more of them were if you
00:58:24 - like trying to replicate that distance to the south wall although
00:58:27 - they don't replicate that actual distance it's stretched down somewhat and
00:58:31 - then fewer of them are perhaps if you like paying more
00:58:34 - attention during encoding to the distance to the south wall
00:58:37 - and they're replicating that to a greater extent and so
00:58:40 - who knows as I showed their viewing direction it's not that the ones paying
00:58:46 - attention to the north wall were all looking at the north wall some of them
00:58:48 - were but there's a few that were looking at the north wall that matched the
00:58:52 - distance from the south sure but it's interesting because then the model replicated
00:58:56 - something that was not present in any of your subjects Dudes.
00:59:00 - Well, we did discuss this during the talk. I think some of these subjects who
00:59:03 - are what you call quite accurate. The blue one, okay.
00:59:06 - Yeah, well, and maybe this sort of reddish one that's in there as well.
00:59:09 - A few of them may have actually fitted.
00:59:11 - But yes, the variance across subjects is not captured in the model,
00:59:14 - and it seems like you'd have to have different weights for the north-boundary
00:59:19 - vector cell and the south-boundary vector cell and tweak that between subjects
00:59:22 - to get this sort of response in a way which wouldn't be very satisfactory.
00:59:25 - No, exactly. But you could do many, many trials and then see if you're still
00:59:31 - predicting their responses, if you could find some way of knowing which wall
00:59:36 - they're going to be paying a bigger weight to.
00:59:38 - We didn't try and do that, but you might be able to. But do you see this as
00:59:44 - a critical challenge to the model, or you think that this is a detail that you can handle?
00:59:50 - Well, so that model only
00:59:53 - requires the boundary only vector cells and the place cells and so now
00:59:56 - we have this more elaborated model which includes the the visual visual
00:59:59 - imagery and would allow us to also incorporate what some of them seem to be
01:00:03 - trying to do by lining up their viewpoint at encoding and retrieval which would
01:00:07 - be to match the current visual scene with the imagined visual scene from encoding
01:00:12 - we could try to add that that would be you know these three different frameworks that i mentioned at the
01:00:18 - beginning i think the spatial
01:00:22 - updating or path integration probably can't be
01:00:25 - used here because there's a random teleportation to
01:00:28 - a different spot uh before they make their response but the other two the visual
01:00:32 - matching and the environmental location surely are relevant so we could try
01:00:38 - adding the visual matching to this and see if that improves and maybe if something
01:00:42 - about the the orientation at encoding gives some of this inter-subject variation,
01:00:47 - because the encoding positions are all different.
01:00:49 - That might be a way to go. We haven't tried to do that so far.
01:00:53 - But that would mean that models on navigation would have to start to include
01:00:57 - a notion of individual style.
01:01:00 - Only if you want. I mean, yes, if you're interested in those inter-individual
01:01:04 - differences, or you might say like we did in this original experiment, who cares?
01:01:08 - Let's see if we can get any kind of match to anything averaged over whatever.
01:01:11 - Yeah, but now you have matched to an average person that doesn't exist. Yeah.
01:01:18 - It's a typical problem in modeling, right? Well, and other good points could
01:01:22 - be made, like all models are wrong. Yeah, sure.
01:01:24 - But actually, a model is only replaced by a better model also.
01:01:28 - Yeah, no, no, absolutely right. So, look…,
01:01:35 - Let's get to the – you then started to map these ideas, also reasoning from these models.
01:01:43 - You looked at the response in the human brain using fMRI.
01:01:47 - But now the second problem that you could face there is, of course,
01:01:51 - also the matching in time, right?
01:01:52 - Because initially we talked about grid cells, place cells, so on,
01:01:56 - and we talked about electrophysiological data. This is all happening in milliseconds.
01:02:00 - And now we are validating the model against fMRI experiments where we actually
01:02:05 - look at a very different spatial temporal window of response that's measured in seconds, right?
01:02:10 - So how do you overcome that challenge?
01:02:15 - And in some sense, the dynamics of the model is completely outside of the window
01:02:18 - of the measurement technique that you use to validate the model.
01:02:24 - Well, we made predictions that were appropriate for that kind of testing.
01:02:31 - We also use MEG and intracranial recording, which has higher temporal resolution, of course.
01:02:36 - But for predictions for fMRI, the most common kind of prediction is that,
01:02:42 - on average, during this task, this particular part of the brain is going to
01:02:46 - be a bit more active because there's neurons there that are doing something
01:02:49 - that are required for the task.
01:02:51 - And that's the kind of prediction that we made. though you may be referring
01:02:54 - to the sort of grid cell like experiment where we try to make a prediction for
01:02:58 - for what would be the time averaged bold signal as a function of running direction,
01:03:04 - over the whole trial um by noticing that the the grid cell firing patterns tend
01:03:11 - to be aligned across the whole population of grid cells we try to make a prediction
01:03:16 - for a difference according to to alignment with the grid axes or misalignment
01:03:21 - with the grid axes in terms of your running direction,
01:03:24 - which we could look for on this time-average data over the whole trial. Right.
01:03:30 - But that was a brilliant idea.
01:03:32 - How did you stumble into that? Or was it really like a sudden insight,
01:03:37 - like, okay, the orientation of the grids has a certain discretization.
01:03:42 - And given that discretization, we must see alignment or misalignment,
01:03:46 - and we can exploit that in our bolt signal?
01:03:47 - Was it really like... Well, I should say, so the other two authors who were
01:03:52 - both postdocs in my group at the time, Christian Dohler and Caswell Barry, had been wanting...
01:03:58 - So Caswell Barry was recording grid cells in rodents and Christian Dohler was doing fMRI.
01:04:03 - And he really wanted, the two of them really wanted to be able to see something
01:04:09 - in fMRI that might somehow relate to grid cells. And so we did discuss it for,
01:04:14 - you know, some years before.
01:04:18 - And Caswell noticed that these grid cells were aligned, the grid patterns were
01:04:22 - aligned. And so we tried to eventually...
01:04:27 - I had heard of quadrature filters,
01:04:29 - which are a good way of looking for a particular phase orientations.
01:04:32 - And so, you know, we put this method together and eventually applied it to data
01:04:35 - that we'd actually recorded over the previous four or five years for other reasons
01:04:41 - prior to coming up with this particular idea for looking for a grid-like signal.
01:04:47 - Right. It's an amazing result. But now it's the 60 degrees.
01:04:52 - Is that, let's say, the optimal, if you want, misalignment between grid cells?
01:05:00 - Is that where you would find the strongest competition between their responses?
01:05:04 - Or do you see the 60 degrees as some sort of canonical feature of their alignment
01:05:08 - in the human entorhinal cortex?
01:05:10 - No it's it's just that the actual firing
01:05:13 - pattern itself is um a regular
01:05:18 - triangular grid and so it has
01:05:21 - orientational symmetry of 60 degrees and
01:05:24 - so that's why every 60 degrees of rotation you should see a similar if you're
01:05:31 - looking ahead you're running in a straight line ahead then you should see a
01:05:35 - similar pattern of uh firing fields of all the different grid cells in front
01:05:40 - of you because it looks the same every time you go through 60 degrees.
01:05:43 - That's because it's a regular triangular grid.
01:05:46 - But that would mean any multiple of 60 should also work. Yes.
01:05:50 - Yes, it does. I mean, the control experiments in that paper was looking at things
01:05:55 - that were not multiples of 60.
01:05:58 - 30 also should not work because that in fact should be the optimal disambiguated.
01:06:03 - And that's what we compared, sort of 30 versus 60.
01:06:06 - But we also looked for you know, instead of six-fold rotational symmetry,
01:06:14 - seven-fold, five-fold, four-fold, eight-fold, and we did not see that.
01:06:19 - So that was the control for our method. Right. So why do you think the triangular...
01:06:25 - Well, actually, that's an interesting point. So if you were an engineer,
01:06:29 - which you probably are, I don't know.
01:06:31 - I'm a psychologist. You're a psychologist, okay. Well, if you were an engineer,
01:06:34 - then you might design some kind of navigation system or something to provide
01:06:40 - a metric on an X and Y right angular axes.
01:06:46 - So why do we have sort of axes that it's more like triangular axes with a 60
01:06:51 - degrees instead of 90 degrees?
01:06:52 - And um one i think um one reason
01:06:57 - you can see from um from dead
01:07:01 - reckoning so in sailors that were navigating they would um try to have multiple
01:07:07 - lines that they were integrating across and uh the reason why three is better
01:07:12 - than two because if i'm estimating my displacement in x and in y it gives me
01:07:16 - a location but if i'm doing it on three axes then any pair of them will give me a location,
01:07:21 - and I can check for my error by comparing it with any other pair.
01:07:24 - And so by having more than two axes, it's redundant, but the redundancy is useful
01:07:28 - because it tells you when you're accumulating error.
01:07:32 - Right. Well, an alternative is that a triangle is the optimal way to cover a
01:07:38 - sphere, which would be interesting for grid cells because that means it's sort
01:07:42 - of a never-ending, right, attractor system.
01:07:45 - Them um well you
01:07:49 - you could so it's close-packed and so
01:07:52 - you might think that you you have um you
01:07:55 - know 2d space represented on a torus in fact yeah and you have close-packed
01:07:59 - representation on that surface and and indeed you know hexagonal close packing
01:08:05 - is is is likely um is another good good reason for why that you might end up
01:08:11 - with grid cells that's right there are models of how how grid cells could be formed, say,
01:08:15 - from place cells in terms of compression or whatever,
01:08:18 - and you would end up with hexagonal closed packing.
01:08:21 - That's right. Although in some cases, if you do the PCA, certainly in rectangle
01:08:26 - environments, you end up with 90-degree grids also.
01:08:31 - So in our models, it's a twisted torus, essentially, with triangular packing
01:08:35 - because that's optimal. Yeah. Okay.
01:08:38 - So the other thing that you showed towards the end of your talk was actually
01:08:44 - an interesting effect that if you have connected rooms but not continuous rooms,
01:08:51 - and you measure the grid cell response,
01:08:54 - along these rooms, that they actually are sort of adjusting themselves with time, right?
01:09:01 - So how large can these adjustments be, right? So what kind of shifts can you observe?
01:09:08 - Because it seemed to look like if these are connected rooms,
01:09:11 - They converge onto still a consistent mapping, a consistent coverage of then
01:09:17 - these two rooms as if it is one.
01:09:21 - So what was the question? Well, what kind of modulation can you expect to see there?
01:09:27 - Well, so Francis Carpenter, who did this experiment with Caswell-Barry,
01:09:32 - he recorded for up to 21 days of experience of the rats walking between these
01:09:38 - two rooms along the corridor.
01:09:40 - And after that long length of time, the grid-like patterns were becoming,
01:09:46 - in the two adjacent boxes, becoming more like a global grid that covered both
01:09:51 - of them. but still hadn't got the whole way there.
01:09:54 - And so I guess what you could expect with experience would be adjustments that
01:10:00 - can cover up to half a grid wavelength so that the two grids can come into alignment
01:10:08 - whatever their starting phases are.
01:10:11 - It's interesting, in that experiment we had.
01:10:15 - Perceptually identical, as far as we could make it, boxes so that the orientations
01:10:20 - of the grids would be aligned.
01:10:21 - I don't know if it might be much harder for the system with misaligned grids
01:10:26 - into two rooms to bring them into alignment.
01:10:28 - Although experiments by Jeff Tauby's group has shown the equivalent thing in head direction cells,
01:10:32 - that if you have two different boxes in which head direction cells have different
01:10:37 - directional tuning and you join them with a corridor that the rat can walk through,
01:10:42 - then eventually they line up and become consistent.
01:10:45 - And I would see these all as part of the same process of trying to build some
01:10:50 - kind of representation of space which is consistent to path integration between
01:10:54 - the different bits of space.
01:10:56 - And do you see that as an adjustment driven by, let's say, CA1?
01:11:01 - Or is it extra Hippocampal? No, I would see in that case,
01:11:06 - because it's the path integration by walking backwards and forwards,
01:11:09 - which is presumably driving the reorganization because that's what tells you
01:11:14 - the relative locations in the two boxes.
01:11:17 - So it's probably the grid cells trying to, if you like, align according to their
01:11:22 - movement-related inputs that
01:11:24 - drives the slow remapping that you see in CA1 in this type of situation,
01:11:28 - where place cells that are originally firing identically in the two boxes,
01:11:33 - begin to slowly disambiguate the two boxes.
01:11:37 - But there's still some monitoring process that can then check whether the grid
01:11:42 - cell response becomes more or less regular at some periodicity, right?
01:11:48 - Otherwise, you won't get the correct alignment because there's a phase and spacing
01:11:51 - and orientation that has to be aligned.
01:11:54 - Yes, I'm not sure there's anything checking. it's more like simultaneous localization and mapping.
01:12:00 - So in robots, you know, you come into a new environment, you try and you've
01:12:04 - got movement detectors, you've got sensory inputs, you try and build a map that's
01:12:08 - consistent with everything.
01:12:10 - And I think in this situation, perceptually identical, the perceptual feed forward,
01:12:15 - you know, boundary vector cells, and so on to place cells was dominating and
01:12:18 - making identical replications of patterns in both boxes.
01:12:22 - But eventually, it manages to do something like slam
01:12:25 - and make a representation which is
01:12:27 - consistent with the path integration okay but
01:12:31 - i would see it as you know you can't put one
01:12:34 - before the other it's trying to come to a compromise of both
01:12:36 - and there's nothing checking okay but the
01:12:39 - plasticity sits in the entorhinal cortex driving
01:12:43 - this or the or in the hip or in the ca1 or ca3 well
01:12:46 - there must be plasticity uh i
01:12:50 - was going yes i'm not sure actually i was what
01:12:54 - i was going to say was there must be plasticity between the play cells and the
01:12:57 - grid cells because the alignment must be but i'm not sure now actually it may
01:13:02 - be that um you know the play cells are trying to anchor the grids to the environment
01:13:07 - and the grid cells providing path integration input to the play cells it may be that.
01:13:13 - That doesn't require plasticity, and that the plasticity happens as the grids, if you like,
01:13:22 - pay more attention to their movement-related inputs than their sensory-driven
01:13:26 - inputs from place cells and become more path integration-driven.
01:13:30 - And that, as a consequence, changes the place cells to become more consistent with that pattern.
01:13:35 - Exactly. And that actually didn't require plasticity between the place cells
01:13:39 - and the grid cells. Right, yeah.
01:13:41 - And on top of that... But on top of that, then you must have plasticity with the two kinds of inputs.
01:13:47 - So I've posited the movement-related input that we think is driving the grid
01:13:50 - cells and the environmental sensory input, which I think is driving the place cells.
01:13:55 - There must be plasticity in both of those inputs, because in the end,
01:13:59 - a place cell in one of the environments will be firing differently relative to its sensory inputs.
01:14:03 - And initially, the grid cells will be firing inconsistently with path integration
01:14:09 - because they're driven by the sensory input.
01:14:10 - So the plasticity in that model would be not between the place cells and grid
01:14:15 - cells, but between the place cells and the central input and the grid cells
01:14:18 - and the path integration. Right.
01:14:19 - But also because this emerges relatively slowly, right? It takes a long time to get this alignment.
01:14:25 - So how many hours does the animal spend between these two? Well,
01:14:28 - that would be a couple of hours a day for many, many days. Exactly.
01:14:32 - And so one valid question is if you made the task depend on knowing where you
01:14:38 - were in each box separately, maybe it would happen faster. Who knows?
01:14:42 - That might be worth trying because it's a horribly long experiment to run in
01:14:46 - its current form. But we don't know. Okay, great.
01:14:50 - So we talked earlier about the concept cells, and in some sense it's now...
01:14:57 - The hippocampus is almost making Jennifer Aniston more famous than she was ever before.
01:15:02 - But we also have a new concept cell, which is now representing the concept cell of Jennifer Aniston.
01:15:10 - So if we start to think about generalizing away from spatial cognition and just behavior in space,
01:15:17 - and we start to think about other domains in which you can use these capabilities
01:15:21 - of the hippocampus, we still would need some sort of path integrator that drives that system.
01:15:27 - So imagine we're going to apply this now to, let's say, a complex cognitive
01:15:30 - task, and we're going to use the computational capabilities of the hippocampus.
01:15:36 - What would then stand in for its path integrator if I'm not moving in physical space anymore?
01:15:42 - More well i think as
01:15:45 - we um discussed with this stretchy bird
01:15:48 - experiment and the grid cells if you have a
01:15:51 - nice grid cell representation it's very powerful and can represent the vector
01:15:56 - 2d vector relationships between uh any two points in a very large space this
01:16:02 - could be very useful for mapping all sorts of things concepts for example where
01:16:07 - the vectorial relationship would
01:16:08 - be on vectors of neck length or leg length or who knows how much I like them
01:16:14 - versus how much they wear expensive clothes or anything.
01:16:19 - But it would be a powerful system for representing whatever that conceptual space is.
01:16:23 - And although path integration was probably useful for wiring up these grid cells
01:16:28 - in the spatial situation, if you have that representation somehow.
01:16:33 - In this non-spatial situation, and maybe it's developed as a PCA of these non-spatial firing patterns,
01:16:40 - or maybe it's somehow the conceptual problem is mapped onto space implicitly.
01:16:46 - Either way, you then got a very powerful system for understanding relationships
01:16:51 - between different concepts.
01:16:54 - Okay. So Neil, to finish up,
01:16:58 - okay, you come from physics, went to theory, and then to experiments,
01:17:04 - now you combine this in animals and humans, and you also have been part of some
01:17:08 - really great discoveries in this field,
01:17:11 - and which also was recognized recently by being elected as a fellow of the Royal
01:17:16 - Society, which is a big compliment to your work, so congratulations.
01:17:22 - But the question now is, So given that experience, what would you see as Neil's
01:17:29 - law in the study of the brain?
01:17:33 - I don't think there's any such thing as Neil's law. I do think that if you're
01:17:38 - trying to model something complicated like cognition.
01:17:49 - Then at the level of neurons, you need to start with the simplest possible model
01:17:54 - because we have so little idea about how neurons do represent things like cognition.
01:17:59 - Obviously, in the spatial domain, we have some great clues from all of this work we've heard about.
01:18:05 - And so i tried to always use
01:18:09 - the simplest model um and also
01:18:13 - it might not be a mathematically tractable model although
01:18:15 - it would be great if if it was but it doesn't have to be it's not clear the
01:18:19 - brain is going to be mathematically tractable and it needs to make experimental
01:18:24 - predictions and indeed experiments have to um have something to say to to theory
01:18:31 - other you know they each only exists with the other in some useful sense.
01:18:35 - If you do an experiment, it doesn't impact any theories.
01:18:37 - What was the point? If you have a theory that doesn't impact any experiments,
01:18:41 - you know, maybe it'll be useful one day, but it's not that useful right now.
01:18:45 - But wait, the law of defining laws is they must fit on a t-shirt. So what's the law?
01:18:56 - Uh keep it well i
01:18:58 - think i think there's a um somebody famous said something
01:19:01 - like you know if i could capture my contribution maybe
01:19:04 - it was feinman in in one sentence it wouldn't be much of a contribution okay
01:19:08 - i don't think that's true i think there are some great discoveries which can
01:19:12 - be captured on a t-shirt but i'm not sure but i thought you were saying you
01:19:15 - know i thought you were saying keep it simple yeah that's what i that would
01:19:19 - be fine keep it simple keep it grounded in experiment.
01:19:22 - Okay, good. Great. It's not a very exciting t-shirt.
01:19:25 - Well, it depends what else you put on it.
01:19:28 - Or who's wearing it. Maybe Jennifer Aniston. But look, five years from now,
01:19:33 - I'm going to smuggle myself into the UK because by then, after Brexit and the
01:19:38 - whole disaster that goes along, I will not be allowed to enter anymore in any legal way.
01:19:42 - And I'm going to come to your lab because by then you'll still be there.
01:19:47 - And I'm going to check whether a specific prediction you made today was actually verified or falsified.
01:19:53 - So what's the one prediction you really would like to see tested in that five-year frame?
01:19:59 - Well, there's a few.
01:20:05 - I mean, I didn't really talk about it because there wasn't time and it seemed
01:20:08 - a bit complicated, but I really would like to know if.
01:20:13 - Temporal coding and theta phase procession has something to do with path integration.
01:20:18 - And it may take nearer to 10 years but I hope that one point we'll be able to image grid cells,
01:20:26 - and their dendritic inputs to work out what it is that's making them fire in
01:20:30 - a grid like pattern and there's a clear prediction that you should see oscillations
01:20:35 - of different frequencies in these different dendrites and I would I would like
01:20:38 - to see that at the more cognitive level there's some
01:20:43 - applications to what happens in post-traumatic stress disorder in terms of different
01:20:48 - forms of memory supporting imagery or impacting on them in different ways.
01:20:53 - And it would be nice to see if some of those predictions have come out that
01:20:57 - this model of imagery had some clinical relevance.
01:21:00 - All right, great. Neil Burgess, thank you very much for this conversation. Thank you.
01:21:08 - The CSN Podcast was produced by the Convergent Science Network of Biometrics
01:21:13 - and Biohybrid Systems, a project funded by the European Sevens Research Framework Program.
01:21:21 - For more interviews, recorded lectures, or upcoming conferences in the field
01:21:27 - of biomimetics and biohybrid systems, go to csnnetwork.eu.
01:21:33 - Music.

CSN Podcasts

Neil Burgess on boundary vector cells and place cells

Season 2018

Description

CSN Podcasts

Episodes

Ton Coolen on immune networks and neural networks

Paul Verschure & Tony Prescott on synthetic psychology and robot models

Giacomo Rizzolatti on mirror neurons and action understanding

Francesca Cacucci on hippocampus development and grid cells

Bechir Jarraya & Lynn Uhrig on anesthesia and consciousness

Timestamp

Be the first to leave a comment

Leave a commentCancel

Legal

Follow us on

Neil Burgess on boundary vector cells and place cells

Description

Episodes

Timestamp

Be the first to leave a comment

Leave a commentCancel

Tags

Legal

Follow us on

Login to enjoy full advantages

Please login or subscribe to continue.

Go Premium!

Enjoy the full advantage of the premium access.

Stop following

Cancel subscription