Skip to main content
Loading Explore guides…
Hippocampal Memory & Sleep: Cognitive Maps · Acedra
Home
Explore
Biology
Hippocampal Memory & Sleep: Cognitive Maps
Sign in to save
Chat
If YouTube asks you to sign in, open the video on YouTube.
Watch on YouTube
Reload player
Transcript
128 segments
Copy
Hide video
0:00
BORIS KATZ: --is Matt Wilson. He's a professor at the Department of Brain and Cognitive Sciences at MIT. Matt studies how animals learn and how they remember what they learn, and today's talk is hippocampal memory cognition and the role of sleep. MATTHEW WILSON: Great.
0:24
Thanks Boris. Well, good to be back here for the summer session here at Woods Hole. So I always like to give a talk from the perspective of, I'm an engineer, trained as an engineer. From the perspective of an engineer who's trying to understand what the brain can tell us about the fundamental nature of computation underlying human intelligence. There's going to be an ethics talk later. We'll be talking about AI. Obviously there's a lot of interest
0:56
in big data driven approaches that transformer models that can kind of emulate intelligence. And so there is a question, what does the brain have left to tell us? If we can emulate intelligence. What is it the brain has to tell us? And I guess Ed Boyden's going to be giving a talk later, and will be talking about new technologies that give us even greater power to interrogate and understand the nature of underlying brain activity. But I like to use the model system that I'll talk about,
1:30
the hippocampus, to try to give some both mechanistic but also kind of a higher level insight to think about what is it the brain does that kind of challenges the ideas regarding intelligence. And I'll just give you the little spoiler. We think about intelligence as being fundamentally about creating rational systems, that can take facts, that can make appropriate judgments, that make good decisions based on all the facts. That are kind of smart. And we think of human intelligence
2:14
as being the biological manifestation of precisely such a computation. But I think when you really look at human intelligence, and you look at the challenges that systems have, you realize that the approaches that are used to achieve this large scale massive rational intelligence are not really available to biological organisms. Organisms live in very narrow and small niches. They have very limited access to data. They cannot rely on repeating the past, because the future is these niches change.
2:56
You have to produce adaptive behavior in novel contexts. And that is that you have to do more with less. You have very little data, and yet you have to make decisions that reflect some broader understanding of the world at large. Which means you have to make models very quickly. You have to make and draw inferences from the limited data that you have. And so, and you have to do this quickly and efficiently. And so the idea is that the systems the biology has evolved
3:28
have been systems to do exactly that. Now the consequence of that, making big decisions, building models on very little data, does not actually lend itself to this, sort of, rational grounded system. Rather it leads to systems which are heterogeneous and irrational. And that is that you have many models, individuals, construct their own individual models based upon their limited interrogation of the world. And biological intelligence succeeds not because the individuals are so smart, but largely
4:03
because the populations as a whole are able to solve the problem collectively. And so we have to think of biology not as really smart individuals, but as individuals that are part of a population, that as a whole are able to do the best job they can, with the limited data that they have. So that will be the challenge. How does biology kind of extract data from limited experience in the world. And then build models, draw inferences, based on that. And so my talk will be in two parts.
4:42
The first part I think of as the interrogating the world. As the interaction with the world. This is the wake sleep model. The waking state, animals interacting with the world, and looking at how the brain represents processes information during this active state. And then there's the offline state, as I like to refer to it. That would include sleep, but also periods of quiet wakefulness or inattentiveness where we see the brain switching from a mode
5:13
rather than taking information from the outside world, is involved in internal processing. So there are these two functions take in data, and build models and do inference, and then these two states. The wake state, the sleep and quiet wakeful state. And we'll look at the biology, the nature of brain activity in these two conditions. And then try to extrapolate their computational function and relevance to this larger problem of building models with very little data.
5:53
So the system that I'm going to be talking about here. This is a picture of a rat brain. The overlying neocortex removed, and here you see the hippocampus, which lies in the medial temporal lobe. The hippocampus of this curved structure, the temporal lobes here on the side. And so if you think of the hippocampus, you just go into the middle of your brain, you've got these two curved structures. And in cross-section, if you take a slice through the hippocampus, the hippocampus
6:21
is composed of this very fundamental circuit. It's an older kind of cortex, and it has a simpler kind of structure. So this cortical structure, a three layered structure, in which you have a layer of cell bodies. So here for instance, if you take the waypoints in the circuit that starts from the adjacent cortex, the entorhinal cortex, the fibers from the entorhinal cortex come into the hippocampus, and go through the so-called trisynaptic loop.
6:56
They make three hops through these three primary subregions or subfields of the hippocampus. The dentate gyrus, and then cells in the dentate gyrus and absent cells in area CA3, cornu ammonis, 3. From CA 3 to CA 1, one of the subiculum and out. So if you want to count the three synapses, there's one, two, three, and then out. See, one has a direct output, but the subiculum also has is a primary output structure. You might notice that I don't know what's up with these biologists.
7:34
They didn't take like first grade arithmetic. They go from 1 to 3. What happened to CA2? CA2 is indeed a structure. In fact there's also CA 4. These kind of intermediate in the rodent. They're smaller, less obvious, less prominent structures. One interesting thing about CA2 is CA2 is similar to CA3 in terms of its architecture, it's a little bit different in that it doesn't actually-- it doesn't enjoy the kind of direct connectivity with the dentate gyrus, but it does have one of the signature
8:13
hallmark properties of the CA 3 region, which is strong recurrent connectivity. A lot of it's like a very dense associative network. The other interesting thing about CA2, which doesn't lend itself to a rodent talk, but would lend itself to a human talk, CA2 is the one subfield of the hippocampus that has evolutionarily expanded the most. So in humans CA2 is enormous. So you have a massive CA2 and rodents you have a very small CA2. So you can kind ask the question, what is it that, you know.
8:47
The hippocampus itself, and the overall architecture, has not really changed evolutionarily. It's just that some of the circuits have changed. And so you kind of think, well, what is it that CA2 does that might be particular to the kind of conditions, information or processing that are unique to humans. And so we can come back to that. But that's because I don't have any CA2 data here, so we're just going to have to focus on what we have. And that is so this circuit, CA3 and CA1.
9:24
Now CA1 is interesting in that if you look at the cells, the neurons in CA3 and CA1, they're almost identical. They look very similar. The individual cells. The real difference is the connectivity. CA3 has strong recurrent connectivity, and CA1 has almost none. So CA1 is like a feed forward network. So you have CA3 is this classical recurrent associative network. CA1 as a feed forward network. And in the early 1970s, John O'Keefe, who was a behavioral electrophysiologist,
10:01
placed electrodes into CA1 recording the activity, and of course you're probably familiar with the fact because he won the Nobel Prize for this discovery. He discovered that individual neurons, when you take animals and you allow them to explore space, these individual neurons express spatial receptive fields. As they fired at certain locations. The story, I always like to tell the story as, both kind of the history of science, but also a way of thinking about and approaching science.
10:30
So people have been studying the hippocampus for decades prior to O'Keefe's electrophysiological studies, behavior electrophysiological studies, and in large part because of its interest in the field of memory and human memory. Of course you're also familiar with the seminal case of the patient HM, wonder what a surgical procedure involving the bilateral resection of the medial temporal lobes, cut out his portions of-- went in kind of cut out portions of the hippocampus
10:59
and the adjacent parietal cortex to treat an intractable epilepsy. And following that, lost the ability to form any new episodic or autobiographical memories. So here was a part of the brain that was specifically and perhaps uniquely involved in the formation of autobiographical memory, which was contrary to a lot of theories of brain function up to that point, which posited the brain as kind of like the holographic theory. Theories of mass action. The idea that the brain is a distributed memory
11:29
system, that there's no one-- there's no specialization. It's just the brain does processing. Probably most famously advocated by Karl Lashley, who did these very important experiments in which he went in and just kind of scooped out different area-- locations in the brain, and different amounts of brain, and then looked at the impact of behavior. And his conclusion was the more brain you scoop out, the more-- the greater the impact the behavior and it doesn't make any difference what
12:00
part of the brain. You scoop out any parts of the brain. It's just it's the quantity of brain that matters. So this was really very contrary to that notion. And drew a lot of attention to the hippocampus. And so there was a lot of work. Electrophysiological work, people using rodent models. Putting in electrodes. And a lot of these models use classical learning theory approaches classical conditioning. But all of them had one thing in common for experimental necessity and expediency.
12:29
They involve head fixing the animal. And why do you have to do that? If you put electrodes, if you put little tiny electrodes that are small enough to record from individual neurons, then the tolerances in that recording are very, very small. So you have to be within tens of microns of cells. You can't really tolerate a lot of movement, instability. So everything needs to be fixed or secured. So all the experiments at that point had been done with head fixed animals.
13:01
Now what O'Keefe did was say, you know, this is a kind of a behavioral-- I mean a behavioral psychologist, but also a electrophysiologist. To really understand what the brain does, in natural conditions, you really need to study it in freely behaving animals. So he devised the methods that allowed for fine wire electrodes to be placed in the animal, and with a tether allowing the animal to move around. And that's the simple-- that simple step of allowing animals
13:30
to do what they naturally do revealed this fundamental property of hippocampal coding. And it really changed the overall view of the hippocampus. In part because it revealed this novel correlate, the spatial correlate. But it also fit with a larger kind of theory or approach to the study of cognition and intelligence. And that is this is the contrast between these so-called the behaviorist theory of learning and cognition, and that is the idea that all learning is just kind of--
14:12
is based on simple principles of associative learning. Stimulus response, reward. And so that is if you just like pair a stimulus with a response and you reward it, you reinforce that. It's kind of like the Hebbian-- the Hebbian synapse is a synaptic version of that rule. It's like reinforcement learning. And the argument was, well, that's how we learn everything. So everything is learned. It's all acquired. And it's all tied to reinforced associative learning.
14:42
But there was another school of thought that posited that actually learning is, we don't learn everything. That we actually know a lot before we learn. And that was we start with internal models, and we use our interactions with the world to both test, but also modify those models. And that is that their internal cognitive models, this is the cognitive approach. And the demonstrations of that. One of the classical demonstration of that, was the phenomenon known as latent learning.
15:17
Latent learning was classic behavioral theory. You take an animal, you put it in a box, you flash a light, you give it some food, to go to a certain location. It'll learn to repeat that. And so animals learn what they experience. This in latent learning, if you take two animals, you put one in a box, and you just let it wander around. You just-- it doesn't do anything. The other animal you leave in the cage, and then the next day you come out and you train both of them.
15:43
So now they're going to learn a task, but in one case the animal has had passive or latent exposure to the environment. In the other case, it has not. And it turns out the animals that had latent exposure to the environment. learn better. So they learned something by just exploring the space. There was no reward, they weren't learning any particular behavior. And so this idea that there was latent learning, building based on a cognitive model, and then O'Keefe's discovery of place cells,
16:12
and this fit into this notion that the hippocampus-- yes, question. Was the question what do you mean by latent learning? Exactly, no external reward. So it's essentially learning without reward. That was the challenge to the behaviorist theory. And so the idea is, look, I'm wandering around here doing things. I'm not learning anything until I get reward. If I get rewarded for something, then I increase the frequency of that behavior, and behaviors that don't get rewarded
16:38
are decreased in frequency. And so this is the idea, that's how we learn. We just do things we're like sheep that follow the brightest light, the sweetest reward. That that's how behavior is driven. As opposed to whatever we do, we're constantly learning. We're learning by, again, testing informing and refining internal models. But what are those internal models? And one of those models that O'Keefe in along with Lynn Nadel, in a very influential book The Hippocampus as a Cognitive Map, the idea
17:21
is that we learn on a foundation of spatial representation or maps. And that is we have maps of the world. And then we learn things on top of that. But the map essentially is-- it's there. We can refine it, we can kind of tweak it. But the substrates for formation of maps is there. And here comes the hippocampus, there are these recordings, and you have these cells that look like they encode spatial locations. Now I always like to point out, encoding
17:52
is like it's tossed around, it's a little computational buzzword. Oh, neural codes for this, codes for that. But codes are something we can also talk about representations. Codes have a very particular meaning and requirement, and that is codes are directed transformations that are intended to communicate information to a receiver. And that is that a code is only meaningful if it gets decoded by someone. Simply transforming information does not mean it's a code.
18:26
And so the challenge for the hippocampus is both to point out, a, is there a spatial presentation, is their transformation into the spatial domain. Well, you think about that. Oh, look, I see like place cells. Isn't that-- the split that's-- like a it looks like a transformation. Well, it could be if the spatial information wasn't already there in the input. If it's already there in the input, it's not being transformed. And if it's not being transformed,
18:56
then you have to ask the question, when you what would be the coding imperative. If you already had the information input. So to understand the nature of the hippocampal code, you kind of want to know what's going on the input, the entorhinal cortex. And so this is where the companion prize the Nobel Prize that John O'Keefe received for discovery of place cells, was given to Edvard and May-Britt Moser, who discovered for the discovery of so-called grid cells in the entorhinal cortex.
19:24
And that is a representation of what appeared to be a largely kind of Euclidean even like Cartesian like grid like, regular grid like structure, in which metric distance information was represented. And then these grids were conveyed into the hippocampus. And so the hippocampus, as we'll see, forming representations of places in space. And so we can again think about the difference between a location in space, a spatial representation, like a map, and then a place representation
19:59
in the hippocampus. Now those two things are obviously going to be related. We think about this room as enjoying both a place representation and a spatial representation. You can think about the spatial representation, it's like the geometry, the coordinate system. You know where is this located if I took out pulled out a map MBL, where is it located. But the place representation. So place is, you can think of this as being the combination of location and context. Now what is context?
20:31
Context is this nebulous term. It's the thing that makes this, the experiences that occur in a particular space, unique and memorable. And so you've been in this location. You've been sitting probably, as I always like to point out whenever I teach classes on this. Like in the first class the one thing I tell the students. One thing I'm pretty sure is going to happen the next time we have this class is that everybody's going to come back and they're going to sit in exactly the same place.
21:00
You probably come back and you sit in the same location. There's of a tendency to, I don't know. You just feel more comfortable. Well, yes question. So this is the if you change the location. So let's say you come in and you move to another location in this room. The cells that carry that information, I'll say encode that information, will be different. And that is that by looking at the hippocampus, I will be able to tell that you actually moved to a different location. Great question. Exactly.
21:34
So there was a lot of work after the discovery of place cells, there was a lot of work trying to understand exactly what control place cells. And the obvious thing would be, well, it's like the cues. It's come in you see certain things it's the configuration of cues. And there's a distinction between referred to as local or proximal and distal cues. Distal cues would be things like off in the distance, if you're out you can think of buildings, landmarks.
22:00
Proximal cues would be exactly this things like chairs, the design and the flooring, and manipulations of those two cues demonstrated that they both have like different impact on the spatial coding. I won't go into too much detail on that. But you can think of there being two properties of a map. One is like the actual layout, the geometry, the metrics. And then the other of course is like the orientation. And as you have to orient a map in order for it to be useful.
22:35
And while proximal cues had a lot of impact on exactly what the distribution of cells would be, and you could like by shifting cues you could kind of distort the distribution of the fields, the distal cues would have great impact on the orientation of the map. And that is you used it to align the so-called directional system. So there are cells that carry information, compass head directional information. And those are combined with this spatial geometry to give you the layout.
23:07
But the larger questions of this notion of place. So what does it mean, place? What is context? And some interesting experiments. So you can take the same cells, bring the animal, you come in today. Come back tomorrow morning, record the cells. And large body of work pointed to the stability of these place cells. Same cells fire, same location, the map is preserved with repeated exposure. But there are other things that could change, or so-called remap those cells.
23:40
You come back into the same space, but I change some condition. Now, what would that condition be that I changed? So for instance, I could change the time of day. Let's say you come in here in the morning, you have lectures. And then you come in the evening, same space, but different talks, maybe it's different topics, maybe it's like another course. And what you find that is sufficient to induce remapping. A different set of cells. And so you can ask the question, how can you
24:08
have a different place in the same space? Time of day. That can be a determining factor. It could be the content. Let's say you have instead of me up here talking, you're going to have Ed Boyden up here. He's going to be talking. And the question would be, are you going to have the same cells, the same hippocampal place cells going to be active in the same way, for a slight difference, slight change in the cue. And that is just the one, just one person. Everything else is the same, one person.
24:39
And so this idea that you need to have these maps that are associated with context, but context is not something that you can tie to any one property. It's not like, oh, it's like if the color changes, then we change the context. It's anything that might be relevant that might require some kind of novel encoding. If I have new information, I need a new-- I need a new map. And so that we can call context. And so this circuit, and there was a lot of work
25:10
and a lot of interest and developing computational models that tried to confer the properties of these circuits. These computational functions onto the properties of these circuits. And so this idea of making very small changes, very small changes in the input, leading to very large changes in the output. That's like a different map for morning and evening, one speaker versus the other. And the idea is that was like separating patterns. Small changes input large changes in the output.
25:39
And that was a function that was ascribed to the dentate gyrus. Now on the other hand, as you point out, OK, you come in here. Sometimes maybe some people are different, sometimes you have a coffee cup. There are lots of changes. Many changes to the context that are not relevant, that don't indicate that new representation needs to be formed. And so in that case, you need to be resistant to changes, large changes in the input leading to no changes in the output-- small changes output.
26:07
And that's the so-called pattern completion. And so this pattern separation function, and maybe after we can talk about what all the properties of these circuits dentate gyrus that were argued to confer this property, and then CA3 was largely the recurrent associative connectivity giving this pattern completion-like function. And so that was like say OK, pattern separation. Pattern completion. This is like getting the context right. So it's like, OK, I got a set up--
26:34
it's like pulling out the right notebook if you're taking notes. It's like if I'm taking notes. I got to have the right notebook to write my stuff down right I don't want to write pull out the notebook for the evening lecture, and start taking notes everything gets scrambled up. So it's a way of establishing these buckets for encoding information. Now while you're in that context. That's just the starting point. So you can kind of think about that, oh, like that was the big problem.
27:03
It's storage and retrieval pattern encoding, representation, storage and retrieval. And a lot of work in associative memories focus on exactly that. How do you use these circuits to encode and retrieve information? They talk about capacity, fidelity, all these things. But when you really think about it, what is the information that you're putting into these notebooks? That just kind of ensure that you've got the right context. You pulled out the right notebook.
27:30
But what you actually put into those notebooks are, it's the time series of events, the experiences that occur within that context. And so you have to have these two functions available to you. The ability to have a stable notebook, a context, a representation, a framework on which you can encode information, but then you have to have the ability to encode, time sequence information, so-called episodic information. In an instance unique way.
27:59
And that is it's like one trial learning, episodic memory as you're able to encode one shot memories things that occur exactly once. And so we'll see how the hippocampus tries to achieve that. Spatial representations, contextual dependence on this place, and then the ability to encode time sequence information using the dynamics of the hippocampus. Follow up question or no? Yeah. Great question, right. So this question, what triggers remapping? A lot of people have studied this.
28:35
I just had a postdoc who recently published a paper on this. Taking a Bayesian inference framework. And that you're trying to do inference, that you never really-- you don't know, right? Is this a different context, is it not? Is this change important, is it not? And so as you accumulate evidence that, in fact, the information that you're getting actually forms more than one-- it doesn't fit well into one distribution. So you need to have multiple distributions,
29:07
and so at that point when you decide, you know, I need a new map. That's the trigger. And there's no one rule, but the idea is that it depends on your experience. The more experience you have, the more evidence that it really does seem like I come here in the morning, I come here in the evening, and the talks, they seem different. They really don't seem-- I can't really lump them together. And so as you gain more experience, you decide, actually these are like two--
29:39
I thought these were just one lecture series. Actually two lecture series. I'm going to split those. And so at that point you actually get the split. So this is what you find. That if you add-- put an animal into an environment with uncertainty, as that uncertainty is resolved, there can be triggers that will lead to precisely that. This precipitous a kind of sudden remapping. So there's also, you describe mapping. There's a total remapping so-called global remapping.
30:13
Global remapping is like complete reconfiguration of the place cells and the spatial representation. And then what's referred to as rate remapping. And that is that the cells fire in the same location, but the firing rates vary. And so you can kind of think about this as like in a vector space, global remapping is like completely changing the angle of that vector. Rate remapping would be changing the length or slightly the angle in one space. And then the other is partial remapping, which
30:51
is what you're referring to. And that is you don't have to change everything, or in the case of rate remapping, you don't really change-- you don't change anything, you just change the kind of scalar properties of the vector. Partial remapping, you can think of decomposing this larger vector space into smaller subspaces. And so you can change some parts, but not other parts. And what this suggests is that the overall representation is not like one monolithic representation. One map.
31:25
You can think of as being composed of smaller submaps. This compositional notion. And so the question is, is there a logic to the way in which you might want to break down a map? How do you break it down? If you're in a conventional map you break it down and it's like a geometric or kind of a spatial breakdown, right? I have maps of adjacent locations, and largely the breakdown of those-- how I decompose that is based on the relative independence of the information within each of those maps.
32:01
And so there are going to be things that are going to be sufficiently described within that individual map, and then there are going to be transitions or interactions between these maps which could be interchangeable. So maybe I go from this room to, there might be five other rooms. And so you know I might want to be able to kind of swap parts of the map in and out. And so this idea that there can be generalization. This is one way of thinking about it.
32:28
Generalization is taking the rules and representations from one experience, and then translating them into another. The idea that there's some equivalence even though the context seems have changed. There's still stuff that I can use from prior experience. So all of these ideas. The ideas of remapping, global rate remapping, partial remapping. The idea of potentially compositional representations that might reflect the use of these cognitive maps
33:00
to serve a larger function, and the larger function that I would argue the system is trying to in the hippocampus contribute to is essentially kind of building a generalizable model of the world. That's what animals are trying to do. Limited data, I need to build a model of the world. Once I have a model, then I can use that model in a generative fashion into both predict and direct behavior, even though things change. That is, it's going to be no-- model is going to work even if there are
33:33
minor changes to the context. And so that's what animals, you could say, are trying to do. And then the question is, how do they actually doing that, and how is that reflected in the activity of the hippocampus. So any other questions? But that's the big-- that's the top down view. So we had definitely spatial navigational deficits. Part of the uncertainty as to what that told us about the hippocampus is that the lesions also included the adjacent entorhinal cortex.
34:05
So the entorhinal cortex you can think of as if it was critical for providing the spatial information to the campus, that could make it difficult to form maps. But there was a lot of work, a lot of arguments about the necessity of the hippocampus in navigation, because HM, highly intelligent, was able to function, to carry on a conversation. So all those other cognitive capacities seem to be intact. It was simply this loss of the ability to form kind of short term anterograde new memories.
34:39
And then there was also this loss in limited loss in retrograde memory. And that is for relatively recent, well in the case of HM, in humans it could be years past. But couldn't remember what he did yesterday or last week. But if you ask questions about did he remember things about his childhood, he could relate-- tell you things that he remembered from his childhood. So this gave rise to this idea, it's like this dual memory system. The idea that we have a system for forming short term
35:15
memories, and then those shorter term memories get consolidated into longer term memories, and that the locus of those two memories is different. Hippocampus is the site where you form short term or immediate memories, and then those get translated and transferred into other parts of the brain. Let's say into the neocortex. And so there was this idea that you have you know memories being slowly, gradually shifted to other parts of the brain. And this was one of the models for sleep.
35:47
Sleep was-- OK, sleep is the time that you transfer memories from the hippocampus to neocortex. And we can talk about that. Yeah. Yeah, I always like to point that out. I think it's really I think it's really an important observation and insight. And that is that damage to systems that are involved in forming memories of the past also impact our ability to imagine the future. And so it's this idea that imagining the future involves a generative process that relies on this memory system.
36:21
So what is it about the hippocampus that would tie like past and future? And so that's kind of the segue into the talk about the electrophysiol, what actually goes on in the hippocampus. And what we'll see is that while that initial observation of place cells and spatial representation in the hippocampus was correct, there was a follow up observation also made by John O'Keefe in the early 1990s that involved the encoding of temporal order information.
36:49
And so what he observed was that place cells didn't just fire when the animals had a location, that the timing of cells with respect to an internal oscillation, the theta oscillation, carried temporal order information. And so when you decode the full population, what you see is the hippocampus is actually every 100 milliseconds or so, 10 times a second, it's repeatedly representing and reflecting your experience going from the immediate past to the immediate future.
37:20
And that is a tying past, present, and future together. So you're doing this constantly. You're remembering what I just did, tying that to where I am now, and then you can say, imagining or projecting, predicting what the future state is going to be. So the hippocampus, by its nature, is constantly tying together in this temporal domain. Past, present and future. Now the question, OK, how do you take that, limited, very limited, and that the time scale of that kind
37:52
of past, present, future extrapolation, is on the order of a few seconds of you could say like the kind of cognitive behavioral time you're imagining. You're remembering where I was a few seconds ago. Imagining where I'll be a few seconds from now. And it's not an absolute time frame, but it's still it's relatively fixed. So thinking about, well, how would you go beyond that? And so that's where we can kind bring in some of these other mechanisms that
38:21
suggest well the hippocampus is able to, in fact, take information about time ordered events, and then use them to both recall things in the past, but also project forward into the future, using some of the properties that I'll point out. One is this encoding kind of time sequence encoding using the dynamics of brain oscillations, and then the use of these different brain states, in particular the quiet wakeful state to kind of express these discrete state transitions that
38:54
would have the ability to kind of transcend the real time constraints of just a few seconds. So you could imagine linked together events I could think, oh, I'm here. I could get to my car, and then I could get to Boston. So from here I can imagine getting from here to Boston in like three steps. So that would be the kind of temporal order linkage that would be possible given the properties of the hippocampus.
39:23
So this happens pretty frequently, I think. Actually let's say I got to slide one. Which is sometimes more than I get to. Kind of get just to the intro, to kind of introduce the basic topic, and we'll take a break at 10:30, so but the first half of the talk really was to try to fill in a lot of the details of what I just went through. Right? And so this idea of the hippocampus involved in memory, and then the observation of spatial encoding, the link between spatial memory and episodic memory.
39:57
It's not an obvious link, if you think about it. Why would damage the hippocampal, why would it impact these too. And it does impact this even in HM, in humans. It was definitely a spatial navigational deficit. Show them how to get to the bathroom, you come back 10 minutes later, he will not be able to get to the bathroom. Right? He just can't-- would not be able to do that simple function. And then the argument is that they one critical feature of these two forms of memory is that they
40:21
involve encoding of temporal order or temporal sequence information. You have to remember events in the order in which they occurred. In one-- and that's fundamentally important. You think about it generally, because time order, from time order, and this is my what, why, and how model of cognitive function. The three step. The three step program to intelligence. The first thing you have to do is like the encoding step. Like what happened in your evaluation of the world.
40:50
You have to take in information, you have to know what actually happened. But you also then from that experience, you have to try to infer why things happen. That is you have to infer causality. And inferring causality requires that you get the order right. And once you get the order right, once you've identified the causal agents, what are the critical things A actually causes B. I mean there's a lot of stuff out here. Most of it is irrelevant.
41:14
Some things are relevant because they have predictive power. Now I can, with this refined model, I can invert that and use this in a generative way to determine how I want to see. So now if I specify an objective, I can use the model to figure out how I would achieve that objective. So there's the what first, and that's where the encoding comes in. So the first thing I have to do is I have taken information, and I have to preserve temporal order.
41:42
Otherwise I can't do the next step, the causal inference step right? And so the approach is going to be simple. We're going to take, this is O'Keefe, used a variant of this, putting a little fine wire microelectrodes into the brain. The shafts are insulated, the tips are exposed allowing you to pick up electric fields generated by nearby neurons. And just given the size and the geometry of these electrodes, so-called tetrodes, because they have four channels,
42:13
you're able to pick up the electrical discharges or spiking activity from very small dipoles that are generated here. For small dipoles you have to be like really close to cells. So you can get really close to cells. What this means is you can see the discharges, the electrical discharges from a bunch of cells in the proximity of this. The tetrode geometry means you can resolve the spatial location, which means you can pull out individual sources distributed around this.
42:42
So you can get lots of cells from a single four channel electrode. You put in a bunch of those electrodes into a little array, you put the array on the animal's head. You stick this-- you put the tetrodes down into the hippocampus, and into one of the properties of the hippocampus, as I mentioned. It's this older simpler architecture or cortical structure where you have a primary-- one primary cell layer where all the cells are packed in. And so that's great.
43:10
It means it looks like, hey, all the cells are there. So if I put the electrodes down where all the cells are. I'm going to get good recordings. Other parts of the brain, cells are a little bit more distributed, so it's harder to take advantage of that, because again the electrodes are really just going to pick up stuff that's near the electrodes. So if the density and distribution of cells differs, it's not going to be as effective. Also the nature of electrophysiology
43:34
requires this dipoles be oriented in a particular way, which means if you have good layered structures, it's easy to-- it's easier to pick up to position the electrodes to pick up. The spiking from lots of cells. The cells are oriented in different ways, like in nuclear structures. It can be harder. So in a sense the hippocampus is the ideal structure to get lots of cells. And so when you do this, this is just an illustration of actual data. Each point here is an action potential
44:06
where it's just where the axes here are the amplitude, the peak amplitude of a spike. Here it's four channels. So it's four dimensions. You're just seeing two of those dimensions. So you pick two channels. And the nature of the electric fields produced by these things is that bigger fields-- closer to the cell is, bigger the field is. So stronger the field. And so when you see, for instance here, a bunch of points, which means a bunch of spikes that have a large amplitude on channel one,
44:34
and a small amplitude on channel two, that means the cell is like close to channel one, and far from channel two. And so different cells are going to have different amplitude profiles. And of course this is in four dimensions. So if you do this, you pull out a dozen, a few dozen cells from each electrode. You put in a dozen, 20, 30 electrodes. You can get a few hundred cells. And so back in the day, and this of course is going to be talking about new technologies,
44:59
and a lot of people, including our lab is using new technologies for doing large scale imaging, which is great. I mean this really is the future. But the one thing that the electrophysiology, the use of wires affords is, that it's actually looking at one thing. It's looking at the electric fields generated by electric currents that are the result of the opening and closing of particular channels. So you're looking specifically at electrical activity. And you know, unlike other methods
45:32
where you have to define sensors, devise and define sensors that will report different states of the cell, and the most popular of which are these kind of secondary indicators of activity, like calcium, which are not directly reflecting. See actual spikes. You see the consequence of that. And so this remains the most direct way of looking at the actual spiking activity until, and Ed will be talking about the new generation of voltage sensor genetically coded voltage indicators,
46:07
voltage sensors that could provide the same kind of, the same signature of electrical activity. And so that's certainly is the future. But armed with just this basic, what I call garage neuroscience, because you could literally go into your garage, if you had the right wires. You could make these things. You could make these devices, and you could do this at home, because this doesn't require it doesn't require a cyclotron. Doesn't require a 70 magnet.
46:36
It's just need some wires, you need steady hands.
46:42
And once you do that, you see, you can kind of see this activity. And one nice thing about this approach, these wires, once you put them in, because they're flexible wires, they remain relatively stable. Over time, days, weeks, months. So you can carry out chronic long term recordings with electrodes in animals that are moving freely, just as O'Keefe had done in those seminal experiments in the early 1970s. And so when you do that here, this is just an illustration of the property
47:10
the phenomenon of place cells. This is just showing each individual point here as an action potential, a spike. They're color coded based on this color coding. And so the idea is that when you see spike amplitude distributions like this, you can go in and in this case manually identify regions or clusters which would identify these things as units. I'm careful not to refer to them as cells, because you don't really know they're cells. Just says these spikes seem to have come from a location.
47:44
A fixed location in space. Right? They came from a particular region. The amplitude profile reflects the relative distance from those electrodes. But you see there are variations in that amplitude, and that's because the spike amplitudes themselves can vary. There was a question? Question, no question. No. OK. And so these could say, well, units what else could they be? If they weren't cells. If they weren't actual neurons, right? And neurons the only-- those are the only units there are, right?
48:20
The little ball, and you know. Well, yes and no. And so like many of the neurons in the brain, while we kind of think of action potentials and spiking activity as coming from the cell bodies, or the initial segments going out the axons. The dendrites or the receiving structures are also capable of producing action potentials. And so you can get spikes that are generated in the cell bodies that you don't see in the dendrites. Spikes that you see in the generated dendrites that you don't see here.
48:49
So really, the cell itself is kind of a more complex computational unit. And this, you'll be able to see that. And as you can have spikes here that might be generated at different points, at but that's not because they come from different cells. They just come from different parts of the cell. And the one way in which you would see that would be, for instance here, this change in amplitude could come from actual change of the amplitude of the spikes,
49:21
or a shift in the location and action potential that's actually generated from different locations. So this is like a change in the relative location. And so all of that is actually going on. You're seeing that. You're seeing both dendritic and somatic action potentials here.
49:39
Amplitude. So this is like the actual amplitude. Like the points here. You just, you're recording the electric fields is just the voltage, the local voltages. When the voltage reaches the peak, you pull that point off and you drop it on here. So channel one, this would be the voltage on one channel. Voltage on channel two, and so it's just the relative channel one, channel two. That's what you're looking at here. And so different cells generating action potential will produce different profiles.
50:08
Big here, maybe small here. And it's that profile that's giving you this clustering of points. So you break down these clusters, you color code them. And when you do that, you get something that looks like this. And this is something that was tested on a so-called linear maze. So the animal in this case is running in two directions. Gets a food reward at each end, runs here and it alternates going back and forth like this. And what you see when you constrain the animal's
50:32
trajectory or path in space to follow these repeated-- these repeated sequences, and that is that the cells now are not simply reflecting the location on the maze, but also the direction. They acquire what's referred to as directionality. So this yellow cell, for instance, will fire as the animal runs through here. You get this each one of these spikes. This is now accumulated over like 10 minutes, so this is like the average activity. On each and every pass when the animal goes through here,
51:02
you'll get this, pop, pop, pop, pop, pop, pop, pop, pop, pop, and then it'll go through, and then the animal comes through here. Now when the animal goes back in this direction, you don't get the same response. Now the cartoon version is, it only fires in this direction. Pop, pop, pop, pop, pop, and then it doesn't fire this direction. In reality you get a difference there's like a different firing rate higher, much stronger response in one direction and the other. Great question.
51:26
So this is, can you get place fields without reward. That was the fundamental premise of the latent learning experiments. You put an animal in a box, no reward. And they learn something. You put an animal in a box with no reward, and you get place fields. Place fields form. And so they form very rapidly, within like just a few passes. Like one pass through a location is enough to lead to the expression of a place field. But then over a period of repeated passes,
51:55
seconds to minutes, will get stronger. But nonetheless the map itself is established like one shot. That's the idea. No reward. What does reward do? There have been a lot of experiments, recent experiments, some older experiments that suggest that reward is like a gain factor. So it can increase the density and firing rate of cells. You get more cells around locations of reward, and that the firing rate of those cells is modulated. So you get a stronger response at reward location.
52:24
You overrepresent reward locations. But you don't need it to form places. So it points out two things. One is there's just like this, again, this kind of the spatial layout. But then there's the stuff that happens in space. And so in this case, the stuff that happens in space is like moving in a particular direction. And so when the animal runs in this direction versus this direction, now you say that's actually like do I need a different code for the two directions?
52:54
Because different things might happen. This direction might be different, might be functionally or behaviorally distinct, in that I might have to do different things in this direction versus that direction. So that was the idea of context. So in this case, you can say, direction of movement serves as a contextual signal for remapping the cell. Now it doesn't just fire at this location. Only fires at this location and this direction. Yes question. It's is a question that had been raised early on.
53:27
You think, oh, this directionality. Maybe it's just because of differences in cues, right? The animals pointing in this direction, it sees one thing. And that actually was an early theory, the so-called local view theory. And the idea was, yeah, the directionality is because in this direction you're seeing one set of cues, this direction you're seeing another set of cues, and there was the idea that place cells form based on the configuration of cues. It's like an associative learning thing.
53:51
So I have a cell here that's only going to fire when I see that pillar, and that speaker, in a particular-- with a particular geometric arrangement. So it becomes dependent on cues. And the arrangement of cues and the two directions, obviously different configuration of cues. Now there's a one simple challenge to that idea, is, oh, yeah, it's just the-- it's the configuration, distinct visually polarized configuration cues. Because we just turn off the lights.
54:20
If you turn off the lights when an animal is doing this task, you know what happens to these place fields? Almost nothing. It's like still they still fire, same location, same. Everything's the same. You definitely don't need visual cues in order for these cells to fire. The other thing, what's interesting is, so you can do this if I take this maze, in this case, and I just have a track that goes up like a really strong u-shaped track, what
54:46
you find is there will be place cells that are directional, but now it's for the turn direction, right? They'll fire when you go around this curve, this turn, this way, right? Counterclockwise and not clockwise. Now that challenges the view hypothesis at all. And the idea of specific configurations of cues is driving it, because in the one case, I'm seeing all the cues-- I'm seeing both directions, right? I'm just seeing them in a different order. This way or this way.
55:19
You know, when I go around this way. And then this way, this way, this is the same. It's basically the same cues. It's not one set of cues. It's not one view. Both views are actually part of both trajectories. And the difference is it's this the clockwise counterclockwise. So it's like, eh, you know, it can't be some like static property. It really does have to be something about the time sequence. It's something about the time sequence that must be relevant.
55:48
100% and all of these things will be used. In just a little-- yeah, just a little side note. So rodents kind of get short shrift when it comes to this idea. It's like, oh, yeah, mice, rats they're blind. Like primates. Primates. And for rodents it's all about smell, right? Smell first, touch, somatosensory second, and then vision is like way down there. When you actually do the psychophysical experiments and then you put them in environments
56:16
where you pit visual, olfactory, and somatosensory information, much like primates, vision is actually a dominant sense. So they use vision over other cues if it's available. The difference is they don't require that. And also the nature of visual information, indeed for like distal cues, the vision is relatively low acuity. But for very proximal cues, in fact cues in which you can combine like vibrissal, tactile, and visual information, they can do very high acuity proximal object recognition.
56:54
Just like primates. It's just that you have to put everything-- you got to put everything very close. And maybe when we get to it, I'll talk about these experiments actually recording in the visual cortex, where we found it was very much like the primate-centric view. When you record the visual cortex, you put a rodent in a large room with distal cues, relatively impoverished environment. You find visual responses are terrible. Very few cells fire, they're not really strongly kind of cue
57:25
dependent. But if you make one small change, and that is instead of putting the cues out on the wall, you put the cues down on the floor and close to them, right? Proximal visual cues, now all of a sudden, now you get most of the cells fire, they're very selective, and you get all-- I'll talk about the reactivation during sleep. So you get the real visual code for rodents is something that is tied to the thing that's most ecologically relevant to them, and that's it's proximal cues.
57:53
Because that's where their world is. Primates. Yes, they rely more on high acuity distal information. But it's not a vision thing. It's not like rats are not blind, basically. That's it. Right? And so if the visual cues are available they would use them, but here it's just the time sequence information. You can turn the lights off, and they will use even in the absence of all of these cues, olfactory, somatosensory, visual, they can still do this using the integration of self-motion information,
58:23
or so-called idiothetic information. And that idiothetic or self-motion information can include vestibular information. They can detect linear and angular accelerations. It can also include like motor efference. They can keep track of their footfalls. The kind of strategies that insects use, for instance. They don't have a-- insects don't have inner ear, no vestibular system, but they can keep track of heading and distance by counting footfalls. Rodents will do the same thing.
58:58
And so which is why you can have, for instance, put a rat in a running wheel, and you still see place cells. It's like, oh, they're not moving, but they're still running, and so they can use that the motor efference together. And all together, that has argued that the hippocampal system, the spatial representation, can be sufficiently driven by what's referred to as the path integration system. That they're using a system for like computing spatial location
59:27
by integrating motion signals to figure out where they are. And that's how you build the map, through path integration. Yeah. So little segue. But so I was wondering whether they would jump on the running wheel, so like these running wheel experiments that were really designed initially to challenge the idea that the representation of the hippocampus were purely spatial. And the argument being, well, you put an animal on a running wheel, and they're not moving. Their location is not changing.
59:59
So if there's any non-stationary in the hippocampal representation, can't be space, right? And so when you do this, initially like your-- Buzsáki found this, and Howard Eichenbaum follow up with it this idea that cells in the hippocampus seem to be able to produce non stationary responses that are more correlated with time elapsed than elapsed location. OK. And so they referred to these cells as time cells rather than place cells. And this is like there's a whole history going
1:00:30
back right from the beginning of O'Keefe's discovery. The challenge of the notion that the hippocampus is primarily, or some would argue exclusively involved in representing space. The so-called non-spatial hypothesis. Like yeah, no the hippocampus is nothing special about space. Space just happens to be the thing that is always there. And so of course you always see it. So there's a great effort to devise experiments which took space out. Animals doing things without spatial correlates.
1:00:59
And the short takeaway from that experiment is that like three decades more, or up to like five decades of research, have only really strengthened the argument that there is something fundamental about space. It's never-- it was never the case that O'Keefe argued that it's space and only space. His argument really was that there was something special about space. That whatever else you include in the hippocampal representation, space is always there. And that largely has been the, I would say,
1:01:30
the conclusion from most of the experiments designed to try to remove space as a relevant correlate. And that is you don't need space. A task that don't need space, and yet the hippocampus still encodes it. So that to me suggests yes, there is something special about it. And what would that-- why would space be space be special? I think is an interesting question. Now of course, there are many other functions that the hippocampus has been tied to that don't seem obviously spatial.
1:02:02
A lot of interest, for instance, social cognition. Social cognition is something that you could argue, and has been argued. Well that's like a CA2 thing. CA2 might be more specifically involved in social cognition, is elaborating humans. Doesn't involve space, and yet in different parts of the hippocampus, everything I've talked about right now. There's this long axis of the hippocampus. The so-called dorsal and ventral axis in the rodents, or anterior posterior in humans.
1:02:30
So the hippocampus in humans, it's kind of, yeah, as you evolve, and you get lazy, the hippocampus kind of slides down, because it's kind of-- and so what's dorsal in the rodent is posterior in humans. So in the human literature talking about anterior hippocampus, is really in the rodent literature you would talk about as ventral hippocampus. And ventral hippocampus has a different kind of connectivity, it connects the prefrontal cortex. It's been associated with emotional learning
1:02:59
and social cognition. And yet when you look in the ventral hippocampus, you still find place cells. And place representation is still there. It's different, place fields tend to be larger. They may have certain kind of symmetries. But place is still there. So that was always the O'Keeffe argument. Whatever else is being encoded, place is there. Now why would place still be there? Why would you use-- let's see-- the idea of place. And again, I kind of come back to this notion
1:03:30
of a path integrator. So a path integrator is largely, I construct simple equation of motion. it's the x dot idea, right? If I have x dot, something that's changing as a function of time, state changing a function of time, I can integrate that to get my state. Space is an obvious one. The x dot is just-- again, it's like self-idiothetic, self-motion information. But x dot-- x can be anything. It doesn't have to be space, right? And so it's the idea that path integration,
1:04:02
the path can be a path in literal space, and semantic space, and social space. And so it's the idea of using-- the hippocampus using this derivative information. It's representing change in state in some way. And then use the same principle. The same principles that work in space to do path integration. So the path integrator works in space, you can generalize to anything else where you have an xd ot. And so it kind of comes back to this-- well, we'll see if we get--
1:04:37
this idea of utilizing gradient information. And in particular, the inputs to the hippocampus, reflecting gradients, or these like x dots being able to compute the derivative of those things, being kind of fundamental to the computation that the hippocampus is going to perform to try to keep track of time sequence information. Could be a sequence of locations. It could be a sequence of words or letters. It could be a sequence of social interactions. It could be a sequence but you need the x dot.
1:05:09
And the nature of the representation that can exploit that will be the common property. And so that's really that's the argument. So I tend not to get too bogged down in this, is it space, is it not space, is it. And then the idea of time or time cells. Clearly time is the critical-- is the one common thing. That's why in the opening slide, the slide one that actually got to, that was the argument. It really is-- it's the encoding of temporal information. How is that done?
1:05:46
So I posit this time sequence information, but I think more fundamentally, it really is this x dot time varying state.
Here to help you learn
Sign in to keep a private conversation with your course tutor.
Ask me anything about the material…
Sign in to chat