Yale Psychiatrty Grand Rounds: September 13, 2019

Transcript

• 00:00It is my pleasure to introduce our speaker. Today, who's doctor George Strigoi and George is someone who did his MD and trained in both Medison as well as doing some clinical work in psychiatry in his early days in Romania and then he decided that he really wanted to go much more deeply into neuroscience and into what was happening in the brains of his patients and in general of.
• 00:30Of mammals and so he came to the United States and he trained with Boo Jackie at Rutgers and started to do really beautiful. Both Physiology and then also computational work to make sense of the recordings that he was doing. He went on, then to do a PhD postdoctoral fellowship with Suzumoto Nagawa at MIT and there, he had the rare distinction of having a 2 author paper where he was first.
• 01:01And his mentor was 2nd and we all know that mentors don't do that much, so really a beautiful paper showing an absolutely fundamental phenomenon in the hippocampus that had not been discovered before and that was that there were ensembles of neurons within the hippocampus. I'm not going to tease him too much 'cause you will tell you more, but they were already ready to encode trajectory's in space so that then when an animal moved through space. These pre played ensembles of activity.
• 01:33Predicted the later sequences of neuronal activity when an animal learned a new environment. Now his work here in dependently has extended that hippocampal work in that ensemble work. But what's really exciting about the work is although it's very fundamental about how all of us actually use our hippocampus? How we encode our movement through space. It's also important for understanding how the hippocampus participates in internal representations and when we think about patients with schizophrenia for example, those internal representation.
• 02:08It's become independent of the external world and so some of his work now in the Department in particularly in collaborate collaboration with folks who are doing human work is extending this fundamental neurobiology too? How do we understand processes that may go wrong? In psychiatric illness so without any further ado, I'd like to introduce George and welcome him to the podium.
• 02:41Thank you very much for the kind introduction and thank you. All for coming to the to my presentation. I'm really.
• 02:49Happy to be in the Department of psychiatry.
• 02:53I have a journey that started a while ago night are not the volume so a journey psychiatry who started quite awhile ago in medical school when I attended all five of My 6 Summers into the partner psychiatry was particularly.
• 03:13Female side causes unit and I've learned a lot from from that experience.
• 03:18Later on, I came to the US to do a PhD in neuroscience.
• 03:23And much, much later about 5 years ago.
• 03:27I give another set of presentations to the problem psychiatry, and made a few promises there in a in a 5 year plan that I hope today. Those people who are there also will appreciate some progress so I hope you're not going to be intimidated by the title or the contents of the work. I'll try to connect as much as possible and feel free to ask questions during the Clarifier.
• 03:51It's just so the title is neuron ensemble in underlying internal generated representations.
• 03:59And what I mean by internally generated representations of what everybody else should mean his mental representation of physical objects or events that not currently present so there. Several forms 4 of which are listed below. Some have to do with our path such as the case of memory. We can retrieve memories of events and physical objects that mattress currently present.
• 04:18Some have to do with our future such as the case of imagining and planning. Probably the most genuine form of these internal gender. Representation occurs during sleep when the brain is fairly disconnected from the external world and we have vivid representations about the world and finally not because we're in Department psychiatry, but this is always been there. Some of these internal generator representation. Take out normal form such as the case of hallucinations and delusions where subjects perceive.
• 04:50Objects that no one else can see.
• 04:53So where in the brain should we look for neuronal ensemble in neural patterns that may underlie and make lead us to understand better how the brain generates internal internal representation so.
• 05:07Sorry it turns out that the brain area called hippocampus that I had been working on mostly to understand the encoding of spatial information has been intimately related to all those 4 aspects.
• 05:21Primarily starting with the 1957 case of very famous patient now HM who lost the ability for a new memories after bilateral removal of is with people campus school was the surgeon.
• 05:36And Brenda Miller or investigated his ability to form new memories for many years after and this was done to alleviate an intractable epilepsy.
• 05:46Took about 50 more years to realize that the same brain area and actually in the same patients.
• 05:52Does play a critical role in imagining and planning particularly in those representations have something to do with whoever spatial content?
• 06:02As I mentioned we dream about things and we have representation during dreaming and this is now moving to the rodent type of studies.
• 06:11Back in 2001, Matt Wilson was next door. They might take while I was in some alternate was lab and we discuss interactive quite deeply.
• 06:21His lab has shown that temporal structure replay overweight people complain sample activity can happen during rapid eye movement sleep, which is primarily associated with dreaming.
• 06:31And finally.
• 06:34And finally there's a number of studies and as an increasing number of studies their associate's psychosis with Aper campus and it turns out that Scovel knew very well how to do that surgery because he was already operating on people with suffering from schizophrenia and by bipolar disorder and bilateral dissection of a program was actually attenuated quite significantly there.
• 07:00Generation of a delusion alusa nation, so we know we know this brain area is involved in it, and there's something that links all these all these factors with the Hippocampus.
• 07:11So a few words about the hippocampus in the human is this C shaped structure located in the middle, temporal lobe.
• 07:17In a rodent it's also assists abstract shaped structure, an if we perform a cut around this level, we can reveal the intrinsic circuitry, which functionally goes by the try synaptic excitatory Circuitry International Cortext Adenta Gyros since information about.
• 07:38Pretty much all the sounds they walk into dental gyrus, then then projects heavy excited very projections, to see a 3.
• 07:47Where there's a lot of other sensitive fibers that sort of connect neurons and most likely form sequential activity that we see later in a CA one which is the final stage.
• 07:59And it's mainly the output of the foreground was towards the rest of them primarily to internal cortex and then we also be Cologne like really, really broadcasting for the rest of the brain so all of the experiments that I'm going to talk about today.
• 08:13Revolve around recordings in freely moving through behaving rodents rats and mice that from which the activity was recorded from dorsal say, 1 area.
• 08:24So this is the anatomy, but they?
• 08:28Neuronal ensemble, but the age or LFP type there. There's a very prominent 2 stage model memory formation.
• 08:36First argued by my PhD advisor project in 1989 in which he associated the two stages. The well known stages are memory formation. Encoding and consolidation with two very distinct electrophysiological patterns that occur.
• 08:56Primarily Napa campus at the time it was thought to be exclusively for campus an that is during encoding when animals explore and most likely.
• 09:06Encounter novel information there is a.
• 09:09Stater 12 Hertz oscillation called data in River campus with when the animals go into non rapid eye movement sleep and this by the way will be this little Camera for my talk from now on. I'm not going to mention that on rampart or even resting very different pattern seems to dominate the activity of the G level and that is the ripple oscillation about two hundred 250 to 200. Hertz and then at the same time is the so-called sharp wave which represents really.
• 09:39Very strong synchronous input from the seat rear end to the C1.
• 09:43This brings together a lot of neurons an looking into the content of that a lot of.
• 09:50For not going to describe later.
• 09:53So this is uh the ensemble level, but not at the individual level. But the individual neurone level. Similar discovery by the AC even those graphically work in 1971.
• 10:06Cold or named the individual place of individual neurons principle neurons in the in the CL. One area placers and that the same is true for CA 3 and allowed to logics, then today gyros so all 3 subregions of vapor campus.
• 10:20Are able to at the individual level be active in a certain area and the environment? Despite the animal traveling throughout this would be a square box in from above.
• 10:30And you can see the spatial reference that individual cell if we would record for multiple cell at the same time, it will see other sales mapping neighboring areas, sometimes overlapping area in in a way that the ensemble tessellates in Maps. The entire environment of the that that the animal travel through.
• 10:49So this is also time to say that individual cells are very interesting and exciting to study but it is the ensemble that really gives us?
• 10:57The full picture of how the brain encodes a map or even experience in space.
• 11:03The same phenomenon of glacial activity seems to be occurring in humans here is.
• 11:10Epileptic patients with intracranial.
• 11:14Electrodes and recordings being performed through that as the human performs virtual reality navigation task and here.
• 11:22This should be firing rate this will be the activity of of a neuron. There is maximal in this area just like it happens in the road and but not so much around.
• 11:31And for all this work and more Nobel Prize in medicine, or Physiology has been given in 2014 to John O'Keefe and to maybe it mother and Edward Moser, who did Seminole discoveries on other aspects of a spatial and memory encoding.
• 11:48So I already alluded to the fact that the ensemble seems to bring additional information into the picture compared to single cells. And here I'm I'm showing a cartoon and animated cartoon of an animal moving on what you can imagine being a linear track.
• 12:04And then place cells will be activated at sequential locations along the trajectory of the animal. I already mentioned they they like to fire with the spatial tuning and certain parts of the environment. So here I hate sales that form a sequence. This activity was believed and still is believed to be experienced driven because the animal is moving is awake and then a lot of sensor Inputs. This doing local activate the neurons as the animal moves when the animal goes to sleep.
• 12:31In a significant number of cases, the same neurons can fire in a much compressed manner. There's no time scale here, but about 20 times compressed in time.
• 12:40In the same order they had fired during the run, so because this occurred during slow wave sleep at this compressed manner. It was believed there internally driven and they were called replay because they followed the activity that just occured on the linear track.
• 12:58It was believed also that there is some critical processing that happens during sleep by which the sequences are being compressed, and then likely rehearsed autonomously during the sleep such that the network now learns about the experience damages had the other function of this would be to connect hippocampus with other brain areas and then.
• 13:18But he becomes will teach other braid hairs in.
• 13:22Understanding this type of information for long term storage is super canvas is known not to be.
• 13:28Diligently becomes I've known not to affect the very late the very early memories.
• 13:34Find a little bit of a technical aspect.
• 13:38We are greatly helped by the fact that.
• 13:41Neural ensemble neuron assembles in Depot campus like to fight in synchrony and then like to rest in synchrony too. So we have this up and down states within the campus during slow wave sleep, which allow us to establish boundaries to synchronous activity where the there. We can look for content and we call this frames of activity in this term will kind of.
• 14:03Recur problem the meaning of my talk.
• 14:07Later on, just about 4 years later.
• 14:12We realized that in fact, these type of compressed, temporal sequence activity is not specific to sleep. But in fact, is occurring as the animal is exploring in our case, there was a familiar linear track.
• 14:25So here there's a cartoon description of potato isolation and then.
• 14:32Squared linear mazes more onto that little solution and this will be around 125 millisecond and you can see here depicted in color. The the same cells that fires in a compressed manner in asleep after they had fired already during the run, so with that. I'd like to propose and we propose at the time that an animal model of internal representation of the external spaces. This compressed, temporal sequence of neuronal firing laser activity in the road in program was that can occur during awake.
• 15:03Explorer 30 States and also during sleep.
• 15:06So how did they emerge? Where does it start?
• 15:11Is it all created during the run for the very first time and then as people have argued replay during internal generated states in the campus.
• 15:20It turned out that if we take naive animals and let them run for the very first time and Marina loaded already did this finding on a linear track. We express' we analyze the data and then can order the place cells based on the location of the big firing.
• 15:35And can we expresa relatively long sequences about Harry cells here and then we if we look. This is a naive animal. If you look in the slowest it before the animal has ever run on this, or any linear track and then we use more sophisticated procedure rather than displaying the spikes. We decode the activity from the ensemble of neurons. We call it by isn't decoding.
• 15:56We observed trajectory like sweeps during the sleep through something that looks like the environment that will the animal run next so the position decoded position will be on the Y axis and then.
• 16:13The decoded time or the time it happens in a compressed manner.
• 16:18In the
• 16:20Just about 102 Seven 800 millisecond long on the X axis so these are the heat map descript. The decoded position virtual position of the animal on the linear track and you see they look like projectors, so this was done in a 2011 now in my lab. We've reproduced these effects it. In fact, on a new data set and this is actually a data set recorded in the particular completely independently and published as partly replication of the phenomenon in 2016.
• 16:50We re analyze their data and found the same thing to happen in a different data set different set of animals an quite interesting. Lee across to environment so the sleep was in one environment and then the run was in a different environment, so this speaks to the network reconfiguration.
• 17:06And sort of argues that the cognitive mapping is driven by external environment is probably secondary to.
• 17:13To the network reconfiguration, so we propose granite through the development of ideas in this field, we propose that the hippocampal network generates preconfigured patterns or configuration of patterns that are later being selected, and used to encode Noble information.
• 17:33So.
• 17:34This is a cartoonish description of what I just said. This is the dominant it was the dogma in the field prior to 2011 by which externally driven inputs, extended driven activity occurs in Naper campus is the animal runs on a linear track and if one looks in the sleep following that experience.
• 17:55Sees that the sequence of play. Selectavision is now replayed as the major remaining the only thing that happens in that network without much any other sequence occurring.
• 18:06If one would look as people did being asleep before when will not find such patterns of activity so this was dubbed a blank slate or a double as a type of network in which everything is created.
• 18:19The novel during an experience and then replay it for a few hours and then the board is or erase again ready for a new type of information to being called the next day or next experience.
• 18:29So that was a pretty powerful model in quite intuitive for a lot of people and.
• 18:38We have to somehow change that that vision be given our data.
• 18:43So what we're proposing and we propose back in 2011 and a few years later is that indeed the place cells? Do fire in a sequence is the animal runs and we do see this replay of activity, matching the way the neurons head fire during the experience. But there are other sequences, there occur in the network that are not significantly correlated with this sequence.
• 19:03And even more importantly, the same neuron had fired in the same order in asleep before the experience so the network is pretty confident ever has those patterns on going all the time and it's it's providing this patterns and then associating that type of neural activity with.
• 19:19With the external environment stimuli from the external environment and then becomes.
• 19:23An index for retrieving that type of information later on is like if you like a cell Phone number that here may not have a lot of meaning.
• 19:32But once it becomes yours, then it means a lot of different things. But the Phone number was there before you was not created necessarily for you.
• 19:41A few remaining questions were still on the table at the time the argument here was that everything here is 100% replay of that activity and it's simply our inability to detect the significant correlation of all the other patterns.
• 19:56But they they they are all 100% replay so we couldn't really address that here because we had the animals exposed a single linear track. So we did a different experiment in which we let the animals run on 3 different tracks and they lead to 3 different types of activation of of neurons.
• 20:13And then we found a replay here in red for track one in yellow.
• 20:21For Track 2 and then in blue for tractor E and then we found pre playful for that, like so we
• 20:28Proposed in the study that in fact, the sequences exist.
• 20:34Prior to the very first experience on a linear track and they're selected online and associated with stimuli from this very environment to encode that information and they're later replayed.
• 20:43A few questions actually a lot of questions have been.
• 20:47Raised by this type of a model some of which I've brought into the lab here at Yale and in the?
• 20:55Next part of the presentation I will show 2.
• 20:58Completed studies that that will address and explain and hopefully solve those issues, so one is what is the capacity of the hippocampal network to pre play or preconfigured future patterns of activity? Is there a limit given that we can detect those significant events. Is there a limit to the capacity after which may be confusion occurs in the may be repeated experience needs to this to be performed.
• 21:26So network reconfiguration network capacity is one aspect and the other one is given. This a pipe patterns occurring before can, we move from the core relative domain. Can we move to the predict if a man can, we look at those patterns that recur during sleep an?
• 21:46For the time being, ignoring anything else that happened in the environment predict some part of the of the way.
• 21:52How the neurons will fire in pretty much any next environment?
• 21:57So this is a?
• 21:59I have one slide in between, but then I will I will go to that description. So here is what we found in the in the 3 tracks in the previous slide was a cartoon we find around 7% of the frames there occur during the same sleep to be correlated with each of the 3 tracks and a lot of capacity of the network was left.
• 22:23In Gray area, possibly too quickly code new information. So we thought they the network is pretty efficient given at least the way we run the our experiments.
• 22:33There is not a lot of overlap between these the content of these frames so there.
• 22:39Probably activating different tractors and it was sequential activity across time during the sleep.
• 22:45Is the cartoon of how we envision the network sequential activity? This is during the sleep individual cells?
• 22:54I often use this analogy with with this subway map and we're not too far from New York or Boston so which we probably witness in reality, what it what it's it's a special map.
• 23:06In the station so each cell is noted by capital letter and in each cell could be envisioned as a station and then a sequence through those stations would essentially.
• 23:19Denote the line so here is the red line running through several stations here is the blue line. This will be tracked to you. Notice that sells BNC can participate in Encoding multiple information. This is another way of efficiently. I guess storing and Encoding and storing information.
• 23:39Finally, another example of the Green Line and just one Gray line that we find occuring during the sleep activity and then it's not yet allocated in this in this design of 3 only.
• 23:53So that the fact that we can compute the percentages of individual frames allocated to future experiences allowed us to generate in principle. I an estimate the capacity of the network. So here are the 3 points shown here. They each take around 7% of total about 20% of of network capacity and we simply do the linear extrapolation of the activity and found it.
• 24:20It is unlikely that the network a pastoral exhausted before the animal has been on 15 tracks in each others to directions, which are this. The sales are directness.
• 24:32Specific they have specific directions productivity, so this will be 30 templates so.
• 24:39There's a static view this does not assume any plus this day going on and does not assume multiple rooms and.
• 24:45We thought, This is a good approximation of what might be happening certainly bigger than 0 as people. Speaking out because people have strong before, but he left the left some questions about the size of this capacity.
• 24:58So we decided in one of the study to address.
• 25:03The.
• 25:05Trade off we think between speed of encoding information, which people it could be useful for an then capacitively network, which may appear to be limited at somebody.
• 25:13So another way to display just what I showed about 2 slides ago with the with the hexagons in the cartoon is this way, so each circle is a cell and then the black cells are cells that are active in a certain environment. This will be during sleep. This will be during the run an the internal model that we proposed of extended sequences.
• 25:37Will be described just like that? You have a long sequence of activity during the sleep and then very similar not identical, but very highly correlated sequence during the run, so this will have a very high efficiency of encoding information because there is not a lot of pluses to a new one shot learning can occur.
• 25:55But you have as I mentioned somehow low capacity of about 15 tracks.
• 26:01The.
• 26:03Dogma in the field before 2011 with externally driven activity will basically have a blank slate in the sleep. So this will lead to very high capacity of encoding information you just encode on the run anything that.
• 26:17Drives the selectivity and their multiple of those but whoever low efficiency because you most likely need to establish trust is to be repeated activity.
• 26:26So we were wondering.
• 26:28Weather amid model would be able to achieve high capacity in high efficiency and a good analogy could be with language here that.
• 26:38You have long sentences that exist in the vocabulary and just use those and just make it up. Today right give it a few times and then in the you, you deliver it with some editing every time.
• 26:50Were you created the novel right? That is not going to be super efficient or you? Can have slides or you can have words in other vocabulary that exist. You don't need to create them letter by letter on the spot, but then you can combine them in multiple possible ways eventually and create new centers in right now. Those with pretty much the same vocabulary. So we were inspired by linguistics and we we tried to see whether this model is true.
• 27:18So.
• 27:20Back to the set of hypothesis and questions that I launched within the last three slides or so. the Super Campus expressed productive codes? Is this strong correlation between a sleep in future run.
• 27:33Possibly leading to some predictive code, or is just a quality of thing.
• 27:39The second what is the underlying Arolla Syntax and then here I mentioned this chunking of information in a shorter neural sequences and finally I have to get to that by the end. When and how does temporary compressed sequence coding emerged during the animal development right? How does this activity emerges even?
• 27:59So the first 2 questions were answered in a study led by Cafe Lu Associate in my lab.
• 28:07And here is the classic design that we use a lot of people use in starting place of sequences that is to put to let animals sleep in a naive state and this will be the pre run sleep. Everything is related to this run, then the animals run on a linear track is 1.5 meters long. So it allows expression of a a good long sequence in a good number of cells.
• 28:32On the track and then we let the animals sleep and we look at replay replay, and we can study the predicted patterns from pre to run.
• 28:42We let the animals run again and sleep again and then we also let the animals around on the next 2 tracks. As I shown and then they slip again? So this time, then income income passes that all that activity.
• 28:55But for the time being, a lot of activities related to the program slip. I don't know activity. So what did we find here is something that we've reported but we look at the patterns slightly different so this is a collection of about 76 else. Not all of them shown and the focus is or not two cells that like to fire in this order so in that describe your dinner at the Top cell likes to fire before the bottom cell in this frame and this frame.
• 29:23Ann.
• 29:26Fires in the opposite order in the middle frame and then again, resume its preferred order. What looks like before order and then their frames in which one so far is the element does not so this is a good?
• 29:37Representative sample of what happens during the sleep.
• 29:41So we were wondering how many, the activity that given cell that we decided to look on. It depends on how many cells being active before in that particular order that we were looking for.
• 29:55And.
• 29:56Given our original description of the extended replace sequences. We were thinking. The number is is very large what we found instead is that the dependency order meaning.
• 30:08The group of the number of cells that proceed. This cell in activity during sleep. There is repeated over chance is actually 1.5. So it's between one into cells it like to file.
• 30:22Like the fire before a particular cells, so you would have at most a triplet.
• 30:26They would they would like to recur higher than my channel so that was the starting point.
• 30:31Of indicating that maybe the network does operate in this chunk mode word versus long sentence.
• 30:38Model so.
• 30:41We build a Markov chain model in a transition matrix and all you need to know from this is that.
• 30:49When we look at order preference in sell order firing we find that we find that preference so some values are highest levels are low for instance, this 50.
• 31:012 likes to fire a lot after sale 3132.
• 31:07And this is not always the case so we get a lot of combinatorials here and.
• 31:13Can get a probability of sale be filing update and so on for for all the 7080 sales every cord from?
• 31:21So we try to use this transition matrix and.
• 31:26Basically build a long sequence by multiplying this probabilities, so if they proceeds be.
• 31:31But the .7 probability and be proceeds, see with say .5 probability that we multiply this probabilities to.
• 31:39Get at the probability of ABC happening. So so that in that can be allocated to the 30 still long sequence or so.
• 31:47And then we try based on this predicted pattern too.
• 31:52Investigate whether we can at least reproduce the sleep sequences. So it's Li predicting its own activities. So we testing. This model and we find the numbers that expressed the percentile. It's over a random distribution of a million possible dummy possibilities or sort of possibilities.
• 32:11So we find that the slip sequences, literally rested 100% file. I mean, it slipped predicting sleep, whereas randomly generated shuffled type of sleep activity. It's much. It's much worse at predicting its own sleep so.
• 32:28The metal works, but it's not much scientific advancement here.
• 32:32So we decided to use this model in trying to predict a sequence during the run based on the.
• 32:38Pairwise activity of multiple sales during the sleep before.
• 32:42So here is the sequence of place cells and these are just for example, the two cells that like to fire and sleep in this order and we generated a million possible sequences outside 1,000,000, - 2 possible sequences outside these 2 run sequences direction. One and then direction do not shown.
• 33:04So we generated distribution of probabilities.
• 33:09Of Placer survey place on sequences from sleep.
• 33:13And we plot it as a log.
• 33:17Value and then it's not log normal lognormal distributed so if you take the log. It's appears normal distributed.
• 33:24And then we added the probabilities of the exact.
• 33:28The probability of predicting the exact 2 sequences from sleep and we assess where do they land in this distribution and it turns out? They they have a pretty high probability. Compared to the shuffle ones there, not the highest predictable probabilities. But they are among the very highest that the network and do so that.
• 33:48OK, that was evaluated here for 6 animals in 2 directions of movement soul ended. We ended up with 12 templates to be tested just like that, and and here we were plotting is the percentile.
• 34:00Where this real sequence lands on this distribution of a million possible cases of play some sequence of equal length and we find initially surprisingly but.
• 34:13Now, not that surprisingly that sleep is pretty good at predicting the order or constraining the order in which the place. I'll fire in the future run at above 95% are in every single animal at least for One Direction.
• 34:29So this is for very first run, which is adjacent to the sleep that the animal has just they will just recorded from when we went back to this time and kind of run activity, but all refer to the verifier sleep. We find a very similar pattern. This will be the chance level this will be the.
• 34:48Distribution if there would be no predictability.
• 34:52Future run sequence from sleep proceeding sleep. But here we find a very skewed very close to 100%. I'll distribution for all run activities.
• 35:03That followed the slip session there is adjacency. Temple adjacency effect in which the first sequence is better predicted, and the following ones. But they both they will significant and finally we validated is finding on the other data set that we've been working on from the Jackie Group.
• 35:21They recorded from 4 animals times 2 directions of movement. There was 8 templates 8 sequences and we find 6 of them to be at 190 above 9595% that so the metal works across multiple datasets an it's now.
• 35:37Hear from a lab in Oxford, it works for them, too, so that it's.
• 35:42We hope to be able to use it and more people to use it to decode activity on multiple aspects of brain.
• 35:49Research so.
• 35:52This indicates the sleep has says the ability to generate patterns. That constrain how the network will fire when the animal goes out in the world to run regardless of the external environment, so there's something that the external I cannot bypass entirely.
• 36:08So we decided to look at the prediction of the predictive coding from its complimentary part which is the prediction error. So we
• 36:20We decided that external environment when I decided that we realized that external does influence. This this sequential activity.
• 36:30And we try to understand what happens with that change in.
• 36:36And predicted activity so here we're starting from the run sequence and this is a cartoon, but really look looks just like that. There they look like that. We have a sequence of place cell firing abcd and so on.
• 36:52That occured during the run.
• 36:56But was not predicted that 100 percentile from the activity during sleep. In fact, the sleep predicted. This sequence to occur. Maybe the at its best right so we know something must have happened.
• 37:09Between sleep and run such that this cells, called see that was predicted to fire in this location between G and age.
• 37:17Now fires between BND and we essentially got to this by exhaustively moving the place. One place sell at the time and recomputing. The predictability from sleep and we found this. This factor right so we decided to use this editing of network activity to test.
• 37:41The prediction error signal so in fact, we divided the pairs functional pairs between adjacent neurons in 3 classes.
• 37:51One we call intrinsic unlikely so.
• 37:54This was not predicted by the intrinsic patterns of activity, so was unlikely by those by those patterns. There's a bit of a phenomenal description is not have any method would do it, but it's a blank intrinsic unlikely type of functional connectivity.
• 38:08Or you can call it also new functional connection.
• 38:11There's a class of connections that remain unedited. They didn't seem to contribute to this prediction error and finally there. Some connections there were lost the connections that were.
• 38:24Sort of indicated to be strong enough during sleep that did not last till the next run session, so we call this intrinsic likely.
• 38:33Right so we have these connections here between BND which was lost because she is now in between. And these 2 connections were lost because see moved and then 3 new connections were were created and then a lot of them remain unedited.
• 38:48So 1st question was where does this editing happened preferentially an does it happen preferentially and we find that particular at the end of the tracks. This is the middle to end either 2 and then they kind of flipped so we compare middle and ends and we find that most of the editing either in as we call it location extractor location, insert happen at the ends compared to the middle of Jackson at the ends this, where this is where the reward is which has been previously associated with protection error signal.
• 39:15And also there more cues there supposed to running on a track.
• 39:21So we decided to look.
• 39:24At the predictability of these functional connection from the slip before compared with asleep after.
• 39:30If experience change something in the network is there any type of preferential consolidation or increase of those connections over the others?
• 39:40And again surprisingly, but probably not so surprisingly the intrinsic unlikely connections were stronger in the post run slip compared to the previously.
• 39:50But not the unedited which didn't seem to change.
• 39:54Value and then not intrinsic likely, although in the book, Jackie data set. They actually went down so there's a penalization on the fact that they have not been used.
• 40:04In the run, so here is a summary of the data showing that.
• 40:08This is the difference between post and pre run.
• 40:12Activation of this of this functional connections.
• 40:16So a simple cartoon model will show the network has a lot of potential connectivity between neurons when the animal runs this will be during pre run sleep?
• 40:27When the animal runs this is the sequence. There is being activated a lot of it comes from constraints living sleep, some of it comes from.
• 40:37Factors.
• 40:39Postley probably from the external environment and then these connections are stronger. The thickness of the line will be showing that these are strengthen compared to before but not the rest.
• 40:51So.
• 40:54We decided to go further with this type of analysis in the predictive coding and prediction error and understand the nature of the chunking of activity and 1st of all.
• 41:07Demonstrated that exists So what we did. We took the long sequence of places and we chopped it into and kept the order of place cell firing within each half, but then we swap the haves.
• 41:17Such that there's a new connection here, but all the other connections are the original ones. They look are doing the run.
• 41:24And we did that for smaller and smaller chunks of Placer activity and we did that multiple ways of one of them was by number of cells, including in each chunk essentially looking for? What is the size of the essential chunk that when?
• 41:39Reduced affects the predictability from sleep an?
• 41:43To our satisfaction we find that initially, dropping the cells. The place of sequences in half and simply swapping the order of the two house.
• 41:53Or 3 or anyway, bided by cell number will be about the same thing did not affect the predictability so that that's a sin away robust effect but when we?
• 42:04Essentially affected the organization of place cells as it occured to 2. Three and 4 cell type of size right so we chopped? Which of the network too?
• 42:162 details we probably affected something that was a building block. There was required for this type of predictability to work.
• 42:26The same thing as shown here is a difference too.
• 42:30100% tile.
• 42:32Essentially, the same thing.
• 42:34So.
• 42:35This would indicate that there is something about a triplet plus minus one cells that that could be a fundamental unit of organization in the hippocampus and then we went on to look for them are are groups of neurons firing in exact same or the repeating higher than by channels and what is the size of that junk and we find again this is a this plot against percentile among shuffles everything signific will be above the dotted line and we find that triplets are indeed the most frequent.
• 43:06Sizes would chunk chunk of sequential activity that occurs higher than by chance, but also doublets and then.
• 43:14Groups of 4 cells so to group these.
• 43:18Short sequences into one word we use the word tablet.
• 43:22That is not a triple is not a quadruplet it. It's something that that hopefully will will define what is happening here?
• 43:32How does it look how?
• 43:35How the templates are tablets look in asleep and in particular with regards to how these neurons this tablets will be played during the run's place else.
• 43:46We marked all the places we found in this particular animal by the location of the two cells there occur during the sleep.
• 43:56In this order.
• 43:58And with a certain time lag how they ended up being places. It's it's too much here to mark all of that. I think it would be a lot of animation required. But we are very interested in those that ended up being adjacent.
• 44:09So the sales that, like to fire 1 after the other and they were playing. Also, 1 after the other's place cells and there's a number of those, though not all of them have that.
• 44:19So then we computed the correlation between try to estimate if there is any correlation between the distance in time. During the sleep and the distance in space. They are actually being decompressed right. This time they're not being compressed that they've been decompressed from uh.
• 44:38Compressed way of firing during the sleep into a place cell sequence during the run and we find that correlation exists for both 2, three and 4.
• 44:48Long sequences.
• 44:52Alright so this looks like a very interesting and never been shown to be honest never been probably conceived that the brain network in the C1 area functions. Not quite in long extended sequences.
• 45:04Which they caught the whole network but in fact in this small chunks and we were wondering whether the size of what the group of limited group of known as they were recording from could be a factor in determining the size of cells.
• 45:15So again we went to the our data set also to the our friends data set which had more cells.
• 45:22And.
• 45:23Plotted the correlation or plotted.
• 45:26A scatterplot essentially of average tuplet length as a function of number of sales recorded and found that that number stayed around 3:00.
• 45:35Weather we had 30 sales recorded over 100 by cells being recorded, so we've done this phenomena's newer codons because there seems to be some.
• 45:46Relative rigidity to this, the size of the minimal group of neurons that do fire in more more than by channels in the exact same order.
• 45:57In the network.
• 45:58And then since triplet now seems to be the unit then we simply play the 2nd order. Markov chain model in which we use, not just the previous sale. But the previous 2 cells to predict activity book to constrain activity during the run at the sequential level from sleep and we find that that is also the case. These are the percentiles operate ability.
• 46:18So coming back to our model, we can now reveal what were sitting before that we believe there is an internal organization there is interplate motives that can ensure the network to have high capacity high efficiency as opposed to the other two models, which wich gone.
• 46:36In words supercouple network generates prediction and productive in prediction error codes.
• 46:42If.
• 46:43If you navigate very well in the sensory type of literature. Predictive coding is much more use. Their I think with the first uses in the pro campus, but is.
• 46:53It was a pretty attractive type of term to use an second neuron 2 plates are hypothesize to represent the building blocks of people compliment organization and temporal sequences in the middle model.
• 47:05An A question there is always interesting to address preconfigured sequential patterns are used to encode normal space information. Not just there to show the network is configured in some way but it's actually being used to encode novel information.
• 47:22So.
• 47:27If if a lot of structure is present during the sleep before an experience but also in asleep after the experience I already alluded to this selective plasticity.
• 47:39What is the role of experience so one is that but if we still want to look the way most people do but they extended sequences? Do we find any signs of plasticity so here would be?
• 47:49The decoded projector of the animal from the sleep before run and here is the decoder activity. These are examples from the slip after the experience and you see, there are fairly similar, but not quite identical.
• 48:02In a sense that the replay. This would be occurs significantly more more frequently than the pre playing so the network represents better.
• 48:14The spontaneous activity during sleep represents better the recent experience, then it was representing it before.
• 48:19But Interestingly and these are different methods to look at it. Interestingly, the difference that we would.
• 48:25Tentatively called plasticity as it must smaller than the way that the amount of the configuration that ever had.
• 48:33It's pretty much like your cognitive.
• 48:37World before and after listening to my talk is is a.
• 48:44Is pretty much the same plus plus the little plasticity that my talk induced into your but I would not?
• 48:51I will not be fooled into thinking that you're a completely you know cut ours. Yeah, completely different person or your your cognition is fundamentally different so keeping the scale. We think and experience on a linear track and this is the very first time, naive annual runs on the leaning back. I think it it gets as normal as one can.
• 49:12That does induce some pluses, but there is minimal compared to the network reconfiguration.
• 49:18So quick very quick summary of what I've shown so far in terms of compressed, temporal sequence of firing. This is the phenom. This is how play cell sequence looks like when we decode the activity during the run.
• 49:33We can estimate the the error or the precision of that method by the decoding by plotting the actual trajectory of the animal. This is the position in this is time as the animal moves from one end of the tractor another and here with the heat map unplugging the decoded position of them based on ensemble are active neural activity, taking all the Spikes is this is not a place in people.
• 49:59Anything like that, so we see that the metal is pretty good, and then the phenomenal preplay that depicted already trajectory like.
• 50:07Processes before the actual experience I'll introduce briefly the data sequences, which is related to the cartoon that I showed you this is what we believe is happening during state of a solution in which past current and future locations. This is distance in time are being bound within a data oscillatoria event to.
• 50:30Potentially induce plasticity the one that we've described.
• 50:35And then as the animal goes to sleep. There's the phenomenal replay by which the network depicts projectors the animals has taken.
• 50:43And finally the plasticity would just described by which replay represents better the experience then.
• 50:50So we think that these phenomena here, particularly and data sequences play a role in encoding of novel information. At least the special one and then the plasticity begins the process of consolidation, which is a sign that some process of consolidation of normally newly encoded information occurs.
• 51:08So for the remaining of the talk I'd like to investigate. What is the development on timeline of all this phenomenon and link that with the ability of 4 new memory as as we and our audience develop.
• 51:22So this is what roughly what was known before we started experiment. Uh this is these are audience around post Natal Day 14, their eyes are opening so visual information can start impinging on the network and around P 2324.
• 51:38Another log of what we call in humans, infantile easier the end of that period has been described so the around this age experiences are able to create long lasting map episodic type memories on in in rats.
• 51:54And on the electrophysiological level around post, Natal Day 17 place sells the Scion place cells appear to mature their place filled meaning they fire in a particular location, but not throughout the environment and finally the grid cells, which send a strong input to play sales actually developed later.
• 52:14So these are important timelines, P1517 and P2122 and then be 2324 they were going to use in our experiment.
• 52:24So these are experiments were led by Guzman for Oak and IMP student.
• 52:29And there's quite a bit of methodology here, I thought. Maybe it's worth using one slide to describe it. We've used Silicon probes mobile Silicon probes one implanted on each side of the brain.
• 52:40Each of 32 recording size total of 64, the configuration is like that, about 20 Micron. In between the sites and this is how they target CA. One area in one of the boxes for that exactly the Age.
• 52:55So we implant them about 3 days before the recordings. We plan to start recording and then we scan. Those ages is a total of 19 successful developing animals that were recorded from each being naive to to the experience right. We're not when I considering second day.
• 53:14As part of the experiment for this part.
• 53:16So we started but well, we ended up with twenty one.
• 53:19Recorded rats from P-15 all the way to be 24 hours, their first day.
• 53:24Probes are lower to the sea owner of the campus and this is a good case, but not the only one in which we record summer. Currently, 67 neurons distributed across all the recording sites. This is a good hit of 8 out of 8 sites.
• 53:38Drinks being recording cells.
• 53:41And one stable recordings are obtained and we should experiment, which is the typical pre run sleep. Dinovo first time ever running on a linear track and then post run sleep?
• 53:52One important question was whether the rats at this age are even performing this behavior. And here I'm showing location of their location on the track versus time passed, you can see they initially rest a little bit. At one end, where we place them and then they start running they rest again.
• 54:08And then they keep running and they do it pretty much like an adult would do better sometimes.
• 54:14So there was never an issue starting from Peanut MP 15.
• 54:19And then we plotted the place cell sequence is just as I showed for the adult but across development again. The novel experience all the time.
• 54:28As you can see the places are a little bit detuned, they mature.
• 54:32As shown before around P17. Even then, there's still room to improve their tuning so we group the animals into.
• 54:43They groups so P 1516 and this will increase the cystic power and reduce the number of animals for phenomena that we don't think happen at the single day level.
• 54:55So the 1st question was to test where the during the run. Our bodies in decoding algorithm can detect can can.
• 55:04Depict the trajectory that the animal has taken so run.
• 55:08Decoding.
• 55:11So this in yellow is the position of the animal you can see that NNN heat map is our decoder. The results of the query can see a bit of errors, particularly the in the middle of the track at 3:15 and then then that decoding improves with age and we show that here. It goes from 11 centimeter error. When I want me to track this time to about 4 but all of these errors are much, much smaller than when we scrambled a identity of cells and run again this so this is a this is proof that the network even as Earl SB15.
• 55:42Despite not individual cells, not being very well tuned to individual locations in on the track.
• 55:48Is able to perform special coding?
• 55:52So the first thing we wanted to know is whether during the run.
• 55:56So when during the run the network, the ability of the network to relation rebind past current and future location. Within a data cycle, which we and others think is important for plasticity.
• 56:09Emerges so this is an example from the adult this is position. This is time phase and we can see this look ahead and look back type of phenomena. The network does by which past current and future looking are bound within a data cycle.
• 56:25An interesting quite very interesting, Lee only around P 2324, which is the same age where infantile amnesia ends and ability of of rodents to form what looks like episodic memories lost again to the output.
• 56:42This emerges at 2324 but we do not see experiment. We do not see any clear sign of data sequences during.
• 56:50Before that.
• 56:51This is quantified here we have an algorithm to add this quadrant with this squadron. One and 3 and then subtract 2 and 4 and then divided by the sum so it is an index of how look ahead and look back. This is binding occurs and we find that at P 2324, there is much higher than 95% of chance.
• 57:13And about like the values in the adult but not at 2122.
• 57:20So we decided to compare and test whether the experience repeated experience within the day so in half of the animals that showed up before, but in half of the animals. We let them run again after the second sleep. So there are 2 running sessions and we were wondering that at the same age if that recent experience was sufficient to.
• 57:40Speed up this type of phenomenon and so we, we primarily looked at P 2122 second round versus first run and we find that there is not different whereas B 2324 versus B 2122, so the age difference, dage effect was much longer.
• 57:57So age was experience has been.
• 58:00At least attempted to clarify here.
• 58:04So this happens during the run and we wanted to know.
• 58:10When does briefly emergent when does experience there occurs during the round.
• 58:14Induce plasticity such that we can see replayed representing better the experiment and then pray so this role will be devoted to replay replay across ages. This is in the adult this is just one example of of each particular replaying we find it.
• 58:31At 3:15 sixteen the network is able to decode and these are at least 6 neurons recorded simelton not single is able to record individual locations. This is position in this is time across time, but no sequence as we saw, here and this is not only the ends we could be somewhere in the middle.
• 58:51So the network does a instead of abcd?
• 58:54Right later on, and gradually.
• 58:59Longer trajectory depicting frames are being detected so it be 1718. We see some signs of sequential activity, both in the sleep before and after.
• 59:10And up until 2122 replay is not stronger than playing.
• 59:15And finally in what we called Stage 3. We see adult like phenomenon in which both reply reply are present and replace stronger than so experience induced changes that over lasted experience into sleep?
• 59:31This is summarised here in stage one both replaying replayer below chance, then they both above chance.
• 59:38In Stage 2, but they're not different and then finally replaced on that replay.
• 59:43These individual location depictions are not that fewer that random, they occur in about 20% of cases at this age and then there.
• 59:54The Lord Chancellor in the adult so they've never been reported for saying about.
• 60:00Alright so final slide.
• 01:00:04So with summarize the age dependent stages in the development of compressed, temporal sequences particular those 3 stages.
• 01:00:12In the first one. We think the network performs representation of individual locations. Wichita logically could be quite meaningful that they're not actually been moving much.
• 01:00:24Later, the animals start exploring more and more and then representations of increasingly longer trajectory's occurs.
• 01:00:32Network reconfiguration emerges and develops into more longer sequence is probably more complex too. And then we don't detect any experience dependent temporal sequence plasticity.
• 01:00:43And finally in Stage 3, which coincides with the emergence of episodic like memory's.
• 01:00:50Describer Adalah brothers.
• 01:00:52We see coordinated emergence of data sequences that are binding past current and future locations. And then experience dependent plasticity in temporal sequences.
• 01:01:03So with that I'd like to thank you for the attention and then the lab woodsman in the development cafe in the predictive coding and tablets and the address of the lab and this is the funding starting with seed funding. From here and then growing up a little bit into NIH and hopefully more to come.
• 01:01:23Thank you.