 So first I would like to thank the organizers for inviting me, a bit less for giving me the pleasure to be the last speaker, but I'm happy to see that there are still people in the room. I hope I will not be disappointing you too much. But so today I will try to basically discuss and try to convince Dima if he's still in the round that some part of the SNF can be useful in case of complex other discrimination. So I will basically try to see what kind of neural coding is taking place in the olfactory bulb, the computation and the transformation that is done inside the olfactory bulb to transform the glomerular map into ensemble recordings, ensemble representation that will be delivered to the cortex. And I will try to also give you some data that analyze the stability versus the plasticity of representation when we learn to discriminate others. So the first thing that we are going to look at is basically classical examples that we know in humans that everyone knows that if we drink wine as very normal and regular human beings, we usually have hard time to discriminate a Bordeaux 95 from a Bordeaux 59 or a Bordeaux from a Chardonnay. But these things can be learned. We know that sommelier can basically be trained to recognize. And the question is how these representation that they have in the brain can be influenced by training. So as you may know, a couple of weeks ago, we elected our new French presidents. And these are the two last candidates. And what you may not know is that basically we choose our president not based on their program but based on wine testing. I mean, that's the usual tradition in France. So here you see Emmanuel Macron that is now our president and every woman that we had to expel from the presidency. And you can see that in the wine testing paradigm, she was already knowing that basically she would not make it. But we had to basically train those people. And eventually, after the vote, we biased the program, the contest. And we helped Emmanuel to become our president. So the question is, if he had the time to basically train himself, how the representation would evolve inside the brain? And what could be the stability versus the plasticity issue inside the olfactory system? So we are not studying humans because they are very weird and complex social interaction. As you can see here during the Independence Day, force of July, with the beauty and the beast as we call it. And you can see this very strange dance that we have, you know, I let this for 30 seconds and you will see that it engaged even other people. Only the American president can do that. Okay, so in Switzerland and as a French citizen, we are more concerned about the stimulus we are using. And we are not using humans, as I said, but we are using rodents. And rodents do enjoy not only cheese and wine, as you can see in this video. So you know, if we go to a brewery, you would see that mice can basically learn, you know, to appreciate wine. And eventually, at the end, you know, we will not have the fancy type of analysis of behavior that Andreas is doing, but you will see that at the end you have a discrimination task that can be set up for mice. Okay, so we are going to just focus today the representation inside the olfactory bulb. I'm not going to go into a lot of details about, you know, the coalescence of receptors that you have heard extensively over the last four days, but I will concentrate on, you know, the olfactory circuit. We will concentrate our recordings in the mitral and top-to-cell population and see what kind of computation is done in the circuit by interaction with this population of interneurons, the granule cell, leger interneurons, and eventually also the pereglominal leger interneurons. Okay, so the first thing that we know is that if we try to image other representation in the olfactory bulb, we have seen some of these recordings in Dimas talk, for example, in other talk, that if you use, for example, natural odors, here we would have a banana, a fresh banana smell, and a fresh kiwi smell, that means that we press the two fruits and we see the other representation evoked in the dorsal olfactory bulb of the same awake head restrain mice. Basically what you see here with intrinsic optical signal imaging that the representation activates a large number of glomeruli, and that the combinatorial code is very clear. You have tons of glomeruli that are activated by the two odorants, but that the patterns are extremely different. So for these complex blends that evoke very different percepts, you know, the representation at the combinatorial level would be sufficient to discriminate those representations. So we are not going to use today the natural compound, but we will use the monomolecular compound that will use the similar kind of percept, ethyl butyrate and amylacetate, and we will compare a single monomolecular representation versus binary mixture of these two compounds with different ratios. And we know from work that we initiated when I was a postdoc in the Sackman lab working also with Andreas, that if you mix these two monomolecular compounds, like 60% banana plus 40% kiwi, and the opposite ratio for the other mixture, if you monitor here in calcium imaging but you would have the same, you know, the type of imaging monitoring, the representation at the input level, you see that the representation is extremely similar. So the question that we addressed at that time and, you know, already the answer is that whether this representation that we see here and these two mixtures would be discriminated by the mouse. So we use different go-no-go assays and so we have different throughout this presentation there would be two different type of assays, one that would be freely moving and that was initiated while being in Heidelberg with Andreas, and there is a version that is a head fix that basically gives the same results that we see in the mouse. So what we know is that if you train animal to discriminate simple orders, monomolecular orders, they learn very quickly to discriminate and make this association in the go-no-go task. And if you train later on the mixture to be discriminated, we see that the learning is also very quick. And what Andreas has mentioned already is that we know that these mixtures take longer to be discriminated, about 100 milliseconds on average that we need to add for the proper discrimination. So we see here that this is a common theme in the field now that a single sniff is sufficient to basically recognize these orders. So the first thing that we wanted to know is basically the imaging I showed you before was done on a naive animal. So whether the map representation here after learning is different and that would explain why the animal can actually discriminate, maybe at the beginning it's hard to discriminate but they learn to discriminate and this would change their representation. So we tested that with different type of orders and what we observed here, we compared trained mice versus a naive mice, that means mice that have never been doing any behavior to mice that were trained but without any rewards or just exposed to the same number of tribes for the same orders and what we see is that over a range of concentration the amplitude of the order evoked response here monitored by intrinsic signal imaging is increased in the trained animal. So that means that the strength of the response to a particular order at a given concentration is higher. But if we correlate the representation between two orders, so that means if we see whether the similarity has changed between two orders, what we observed for the monomolecular orders the patterns are decorrelated so that they are separated for acetate and ethylbutyrate are very different over a large scale of concentration and the mixtures are highly correlated over a large scale of concentration. And the training you see here is basically not altering the representation so that means that although there is a plasticity induced by associative learning so you improve the strengths of the representation, you are not changing in fact the representation similarity. So if you have two orders that evoke input patterns that are nearly identical it's not changing by associative learning. And so the question is how the mouse is actually making the difference between those two representation if the learning is not altering dramatically the input mouse. So to address that and considering that we have to nail down the mechanism at the single sniff level we initiated a recording instead of calcium imaging that would be relatively slow to monitor a mitral and tufted cell response. So we implemented tetral recordings in a wake and anesthetize a preparation and we basically monitor the response of large cell assemblies to other response. So you see here one of these cell assemblies so you see here four example neurons this would be the inspiration in gray and the expiration in white different respiratory cycle this is the first respiratory cycle after order onset and every spike is basically a bar. You see the response to the same order across different cells and what you can easily see is that the response is quite variable from cell to cell but it's reliable from try to try. So the trials here are evoking the same response. So based on the literature in the motor system and also worked by Gilles Laurent we decided to analyze the other evoked response in cell assemblies using population vector analysis. This means that we look at all the cells not by averaging their response all together but seeing the activity in the population describing this activity over time. So population vector are nothing else that basically a vector where each line is the rate of a given neuron and what we do is basically a time series where over time we see the rate of a particular temporal window that is fed into each line of this vector. So it depends I mean between couple of cells to up to 30 or 40 cells yeah all by SNF yeah all the all the all the all the recording are done with a SNF sensor for DMA the other application is temporarily locked to the expiration so everything is time locked and so basically we can realign everything from animal to animal. We looked at the finding inside a given animal and we find the same as across animal so it does not really does not really matter. So what we can do basically here we have as you have heard already the different talk so we have in this case 101 dimensions so basically with the population of cell that we record gives a space of 101 dimension and what we see with this population vector is basically a trajectory of activity that is describing this multi-dimensional space and because many people have introduced the PCA reduction of dimensionality I don't have to go too much into detail and so basically what we are going to look at is the evolution of the population activity evoked by an order in this multi-dimensional space but reduced to the first three principal components and I have to say that more than 90% of the variance is captured in these three principal components. So what you see here is basically the activity in the cell assembly that we have recorded before another application and to our surprise what was not very obvious at the single cell level turned out to be a very clear phenomenon at the population so that we have a resting state activity that is describing basically an inspiration-expiration cycle that is over and over again coming back in this activity. And so as soon as we have an order that is applied you see that the representation goes away from this resting state and enter an absurdocyclic trajectory that eventually set up after a couple of breathing cycles into a kind of steady state that we see here. So if we see this representation in a non-unintended way so basically we leave the first inspiration after order on set would be here and then you have the expiration and then you move progressively into this progressive respiratory cycle and you change the representation. So what it shows here and what I should mention that each point here is basically a snapshot of every 40 millisecond. So what we see is that the representation every 40 millisecond is described in a different part of the trajectory. And this is basically meaning that a sub-population of mitral and tufted cell is encoding the representation of the order at every 40 millisecond. And every 40 millisecond it's somehow because the representation is changing in the space is a different cell assembly that is representing the order. So what we also notice is that if you compare orders together so basically you have here the green order versus the magenta order that are unrelated that would be the ethyl butyrate versus amyl acetate, very different representation. You see that the representation and the trajectory of recording are nearly orthogonal in the space. So this is not very difficult to discriminate those kind of representation in this multi-dimensional space. Now if you compare these two orders that would be much more related to each other that could be concentration, small difference in concentration of the same order or binary mixture you see that the representation is actually much more similar between those two representations. It's similar but it's not identical. So if you unfold the space here you see that there is a difference between the two representations. And so we wanted to know is basically the mixture representation that we see here could be discriminated by a process that is called pattern separation. So those representations are nearly identical, they are not exactly the same. But people have speculated over nearly 40 or 50 years that they are in the brain mechanism that basically take a correlated input and as a product of the local network processing would give less correlated output from the network. And among the different circuit that have been proposed the olfactory bulb is supposed to be one of this network that could do pattern separation. Although no one has really demonstrated that this is true in mice and it's actually useful for behavior. So we wanted to test these two questions so can correlated inputs be separated into distinct output patterns by the olfactory bulb network and does this pattern separation if it exists would help really animal to discriminate orders. So in other words if you have these two trajectory that look very close to each other is there inside your olfactory bulb network a mechanism that tend to increase the distance between those representation in order to make the mouse discriminating more easily. So the first thing that we did we created a stimulus space that would be basically varying into similarity so we have mixtures that would be very easy to recognize with nearly no overlap in the glomeruli until orders that would be very similar in the input pattern. So for that we use binary mixtures I will show you in a moment and we quantified with the calcium imaging in single trials the dynamics and the spatial maps that would be evoked by different mixtures. So you see here for example of these mixtures so in fact there would be 16 mixtures. We have two category mixtures so one mixtures one type of mixtures contain the amyl acetate and anti-butyrate components and we vary different ratios in different mixtures and we have a different set of mixtures in which we have anti-butyrate and exanone so anti-butyrate would be a common monomolecular order to the two mixtures but the other component would be different. So what we basically did we had 16 different mixtures of 8 of this pair, 8 of this pair just changing the relative ratios and then we looked at the order response in all the glomeruli created a vector of response in this glomeruli population and compared across orders the order evoked similarity of the pattern so if you take for example this pattern versus this pattern you would see that you have a high correlation that would be shown by these reddish colors. If you compare these two patterns you see that you have a strong correlation also across the different orders that share anti-butyrate and exanone but if you compare these two mixtures they vary in similarity so some of them are very low correlated and you have everything in between high correlation and low correlation. So like that we could create a stimulus space that would basically span within the entire space of possibility of similarity in this experiment. So this is the input space we created it for on purpose. The question is what will happen at the level of the output? So for that we recorded the population of mitral and tufted cell and basically the firing activity that will be recorded from the mitral and tufted cell will be the information transferred to downstream cortices. So you see here the example of some of the recordings basically we recorded 100 cells roughly and we created again a population vector correlation analysis where we computed all possible correlation for all the order evoked response by these 16 different mixtures so like that we can see are similar all these output representation can be. So what I have not mentioned so far that this activity is the average rate or the average calcium response on the first respiratory cycle. And so what you can see here that the representation has somehow changed so if you don't remember how the input was so you can see the input was organized like that and the output representation is extremely different. So some others tend to be decorrelated so separated more than it was at the input level and other others are actually even more correlated at the output than they were at the input. So overall we observe that some others can be separated by the olfactory network not all so the overall tendency is to decrease the correlation that we see at the input level but it's not a linear process so it's more reformatting of the representation and we cannot we don't know right now why some others are separated and not other others. So what is important to know is what is the time scale of this pattern separation process because we know again as I mentioned in the introduction for a long time that it takes couple of hundred of millisecond for the mouse or the rat to discriminate complex orders. So what we did was to basically analyze the data now in the first sniff but just taking the information every 20 millisecond roughly 21 millisecond here and just see how fast the decorrelation process would take. And what you see here is that we started at that time point that is the maximum of the input correlation so that basically it takes on average about 40 to 80 millisecond to decorrelate representation across all the matrix here. Some others are decorrelated more rapidly than others but on average that's what it takes. It means that not only we the your factory but can do pattern separation but it does it at a time scale that is relevant for rodent behavior because for many for many years we had intense discussion with Reiner Friedrich for example that demonstrated pattern separation or decorrelation in the official factory system but that would take much more time than it would be allowing for the behavior in rodent. So this at least is consistent with the behavior of data that we have a mechanism potentially at the relevant time scale. If you have any question if it's not clear just interrupt me any. Yes. Yes. So here I mean we can discuss about it. We can correct for the change in the firing due to the respiration. We know that when we breathe we have more firing after the inspiration and it doesn't change anything. Yeah yeah yeah sure yeah yeah you know I completely completely okay so because we have done you know this large array of comparison you know we have 16 by 16 comparison I mean in fact divided by 2 so we can select pairs of orders that would be either decorrelated or still correlated at the output level of the olfactory bulb and ask the mouse is it really useful to have this decorrelation for you can you discriminate more easily the orders that are separated by the olfactory bulb or is it really you know an epiphenomenon if it's an epiphenomenon then you know we move on and we do something else. Okay so we selected 11 pairs of orders that I said would vary in similarity or correlation of the output of the olfactory bulb from ranging for orders that would be nearly completely decorrelated or orders that would be still highly correlated and you see here some of these pairs so we have very simple orders that you know would show the tendency to learn very quickly to discriminate these other pairs and you see here many of these mixtures you don't have to look at the name it does not really matter some of them are learned fairly quickly like these ones some are intermediate and others are much harder to discriminate. So basically we took the accuracy over this initial period of learning as an indication of the ability of the mouse to discriminate more or less easily the two orders. And we plotted this discriminability index as a function of either the input correlation at the glomerulus level or the output of the olfactory bulb. So what you see here is that this is the input so that means that the maps of the glomeruli can basically not predict very well the behavior of the mouse so for some orders that are having a very high correlation for example at that level some of them are easily discriminated and others are not so this is basically what we knew from from from the past so that the input map is not sufficient to predict behavior but if we look at the output of the olfactory bulb if you look at the mean output correlation in the first respiratory cycle or the minimum output correlation in the first respiratory cycle we see that we have a much better correlation of the activity to the behavior. So that means that the more you separate representation in the olfactory bulb network the easier you can predict the behavior and the easier it is for the mouse to discriminate when output correlation is low. Okay so this is this is this was very encouraging but we know that this is not a causal relationship so we aimed at trying to demonstrate that basically this pattern separation process would actually be really useful for behavior and that's not just again of NEPI phenomenon. So I will show you two different type of modulation that we did and we tried to alter so this pattern separation process. So the first one of the first that we tested was basically a collaboration with Thomas Hiansher from Berlin where we decided to alter all the inhibitory load that the mitral cell are receiving by altering the chloride gradient. So we selectively inactivated the potassium chloride co-transporter specifically in mitral topsoil using cell type specific recombination in mitral cell. So we use PCDH21 CreLine crossed with a flux KCC2 knockout. So what what end-ups that we mostly remove all the KCC2 from mitral cell and we still leave some KCC2 from cells that are not mitral cell like pereglomerules. So when you do that you basically change the chloride gradient and it's stuck. Sorry for that. That's not the chloride gradient. Sorry Microsoft crashed. Yes, yes, please, please. Yes, different. They are extremely different. So that was one of our first surprise. So we always thought that this would lead to same kind of mixtures but it turned out that the liquid mixture are way more different than they should be in from the theoretical point of view. So the liquid mixing is giving more different others mixtures than they are in the air mixing conditions. I mean we spend six months trying to do a very precise gas chromatography and try to find a relationship and the logic and we we could not find something. For one other pair for one other mixture there was a logic and it was the opposite logic in the other other mixture. So we discussed with I mean it was not at Fiamminis so with people that know what they do and yeah they say that it would take you know probably a lot of years to to model that and so we just gave up and so we took that as a fact and that was that was it. So yeah definitely there is a difference between the two mixtures, the two air mixing. Okay so I apologize for this intermission. So if we okay so if we if we change the chloride gradient the expectation is that basically you change you know the GABA response so that if you if you look at the reversal potential here you have a more depolarized reversal of GABA currents which basically when you apply GABA would make in a wild type usually an IPSP so a hyperpolarization and in the case of the KCC2 specific mitral cell knockout we would have a shunting inhibition but we would not have a hyperpolarization of the cell. So all the mitral and tufted cell were like that so this was changing the properties of the of the other evoked response I will not enter too much into the detail. I hope it's not going to crash again so I think I will I will stop using the laser the pointer I think this is the problem. Okay so if we if we apply a different different mixture so here they are not the same as in the previous experiment so but we have some orders that are simple orders orders that are binary mixtures and what you see is that in the wild type mouse basically some orders are highly highly correlated and others are more decorrelated and in this mouse in which we we modify the inhibition we basically modify the correlation pattern in this mouse so we strongly increase the the correlation among different orders so that means that we deteriorate the pattern separation process by manipulating the inhibition that is received by the olfactory internals that is received by mitral cell from olfactory above internals and if we use a prediction algorithm on this data set we basically strongly impair our capability to recognize orders but to of course be more direct we ask the mouse to recognize some of these orders and some orders are actually still recognized by by the mice and can be discriminated but the the orders that have that are the higher the that have the highest correlation are actually extremely difficult for the mouse to be discriminated so this was the first hint that manipulating the gabergic network could impair patterns the pattern separation process and this would alter the the behavior of the of the mouse but we wanted to be yes yes yes yes and no if we overall yes but the problems that in these experiments so in in the other experiments everything was the same so the mouse were head fixed the olfactory meters were the same and the behavior was done on the same kind of setups at the recordings in this case the recordings were done head fixed and the behavior was done in freely moving animals so the comparison for the concentration and the relative ratios because as i said before they are not exactly the same i would not make a very strong claim about the direct relationship but overall it was it was the same tendency but was not as a nice regression one okay so we we wanted to test another type of manipulation that we could control in precise timing and so we and also to go into some potential sub class of intangible that could be important so here we targeted based on work done by andreas in the past and mix and abram that you you can inject in the granule cell layer aav and modify genetically a large fraction of the the granule cell population and so we use two different approaches we use one where we force the expression of channel adeptsin so light gated channel that will upon light stimulation will lead to the excitation of the gabbaric intangible of granule cell layer and the second modulation that we we used was basically to force the expression of an inhibitory dread construct so that's a gpcr that has been mutated to be only responding to this inert molecule close up in an oxide and this gpcr is coupled to a gi o signaling so that means that when we stimulate this we cno we basically hyperparalize and reduce the excitability of the internals okay so we have two opposite manipulation increase the excitability of the granule cell layer internals or reduce the excitability of the granule cell layer internals so basically you see here the the staining for the m citrine coupled to the for the dread and so this is not done in a big outcream mouse but basically the the spatial location was sufficient to restrict the expression in them in the granule cell and even when we add cells that we are close to the mitral cell layer the mitral cell were left uninfected on average we across several animals we we added for both manipulation about 30 to 40 percent of the granule cells that were that were affected and all the mice in which we had for example targeted also the entire factory nucleus were were removed okay so the experiment was done for the channel adapting manipulation with an LED so we got inspired from a work done by pia marido so we basically we perform a cranial window on the dorsal surface of both olfactory bulb and implanted an LED that would cover the entire dorsal surface of the olfactory bulb and hoping that this would basically stimulate the granule cell roughly a bit everywhere so the protocol is is like that so we have a masking light that is just on the eye of the animal that you will not see the light stimulation coming from the dld we have a baseline period and there is the other the animal is supposed to leak during this other period and the lead stimulation is at 40 hertz it's basically coupled to the other one side that is itself coupled to the respiration and if the animal leaks at the end of the other we'll get a reward and so for adi and raffia that I added this this slide that you could see how this works with the sensor so basically it's not very clear for the moment but with the light you will see in a moment here you have the head of the mouse there is the other port that arrives on the right nostril there is a pressure sensor that is close to the left nostril and we are here basically the LED that is mounted and connected to the power regulator and so when we when we apply when we start a trial there is the masking light then there is the delay and then at some point when the other is coming in you see here the flash that we have increased in this case to see it through the the dental cement the light stimulation arriving on to the button so if you see it on the side the mouse is here there is a leaking port is getting the other and is responding by leaking onto this onto this leaking tube okay so there is one aspect that I want to to to to mention here of course you know if we infect 30 or 40 percent of gabberging heat network and if we stimulate and shine light and synchronize them we may lose percept by a very strong total inhibition of the mitrement up itself and that's what we see in these experiments where basically we can increase the light intensity applied on top of the bulb and in this experiment the mouse are just asked to report whether they smell another so if they smell another they start leaking so the other one set is at zero and when they when they smell that there is something then they leak to report that they have smelled something and what you see is that if you go at zero power they have this leaking property and if you increase progressively you lose the leaking the leaking property which for us we take as a sign that the mouse doesn't smell anything anymore so that we basically shunt the activity in the in the bulb by synchronizing the the granular cell that we have infected what you see here of course this is not the condition we want to have I mean want to have the animal learning to discriminate and not just having no discrimination because we shut down the mitral selectivity so we selectively for each mouse regulated the power of the led until the the minimum intensity that would recover the baseline activity without without light and each mouse has a different profile base you know the accessibility of the the cranial window the number of cells that were infected so each mouse has been regulated so what you would see is basically a non-blast of the olfactory bulb interneurons that we do is more a kind of excitation assisted manipulation that is that is done okay so this is one slide basically summarizing a lot of work so we would see here the granular the mice that have been modulated by the channel adeptsin infection in granules and neurons these are other mice that will be that have been infected with the excited the inhibitory dread and you see here the control of these mice so either light off or before the cno application so these are the other evoked rates recorded in the mitral and tufted cells so we see how many spikes they generate after other response or all the mixtures and when we apply the light as one would expect we inhibit the mitral and tufted cell when by recruiting the interneurons and when we did we inhibit the mitral the interneuron with visibid the mitral and tufted cell so there is nothing really special here the mouse the cells are still responding but we alter their firing pattern and this we did that because in the literature from a various group that model the process of pattern decorrelation in your factory but they suspected that the interaction the interplay between gabergic and glutamatergic cells would regulate the fine tuning of the temporal spike pattern and that would basically lead to decorrelation so our hope was that basically by manipulating here and having a bit more inhibition than usual we would change the level of correlation between similar orders and that's what we saw so when we inhibit the grant when we stimulate the granule cell we alter the pattern separation process and we improve this computation in the old factory in the old factory but and the opposite manipulation is obviously degrading the the pattern separation process so now we add two different ways of manipulating on demand a computation done in cell assemblies and we asked whether the mouse would basically improve or not or or degrade or not their learning capability and what you see here is that if you improve pattern separation you improve in fact learning to discriminate those mixtures and if you degrade pattern separation you degrade basically the ability to discriminate those orders so based on our prediction of the correlation between the amount of pattern separation and behavior we see here and we think we can say that there is a causal relationship so the more the pattern are separated by the old factory but the easier it is for the mouse to discriminate orders what is important that this this process is only important if it's required so that means that this is only valid if the orders are highly correlated so if we have simple orders that are not correlated at all so that means that the input are already very different we see that for the first the pattern separation is not improved by channel adopts in it's actually even slightly deteriorated but not significant and the learning is not affected so we take this as a as an a sign that basically the representation in the glomerular are so different that you can basically do whatever you want the mouse would still discriminate those representation there is no challenge there and this is another this is the same mice another group of animals sets of mixtures and you see that when in the same mice we stimulate for the mixture we have a nice improvement in this in this case so the the conclusion of this part is that basically we think we can say that the other representation is described by cell assemblies in the mitral and tufted cell population of the old factory bulb the more the representation are separated the easier it is for the mouse to disambiguate related compounds so basically pattern separation is in fact a computation that is done and is very behaviorally relevant in in the brain but what we also saw that with the inhibitory dread experiment that in some cases you may not want to necessarily separate representation but if you deteriorate this separation you may group stimuli together instead of discriminating them okay in the last couple of minutes I have I would like to continue further because of course here what I showed you is an external manipulation so I showed you that basically if you know as an experimenter we try to tap in the old factory bulb if we correlate or not the representation we can assist or not the the behavior but this is not showing you that basically during active learning we know that these might these orders can be discriminated takes time but they can be discriminated so if they are not if they are correlated at the beginning of the discrimination are they changing over time as a function of the the learning process and that's where the plasticity of the representation is important so the question is at the beginning of the behavior the representation would be highly correlated is during learning a plasticity an ensemble plasticity process taking place to actually improve naturally in the brain this distance between representation okay so to to address this question unfortunately we cannot do a tetral recording over many days we cannot sample for sure the very same cell so we switch to to photon calcium imaging and I don't have to introduce that because a fluorine did so we basically force the expression of calcium indicator gcam 6s in a mitral and tufted cell and you see here one of the optical plane in awake awake mice so we see here the the cells I think these are tufted cell if I if I remember okay so this basically we we use the pcdh 20 qui rt2 transonic line that was either infected with avs or we also had some experiment with an old gcam 3 mouse but I will not discuss this experiment and by just varying the plane so either we can image the tufted cell or the mitral cell independently and so of course the advantage that while the animal is basically doing behavior we can so here we have a mouse that is responding to s plus order after being trained at the same time we can basically monitor the calcium response that is evoked by by one of the of the mixtures so as I said from the beginning and I will I've always been a bit concerned by calcium imaging to address this question because they have a slow temporal timescale so here basically the the the bar is 1.5 second of order application so you see that the order response is basically over you know the duration of the the order response and it has nothing to do with the temporal dynamics I've showed you at the beginning in couple of tens of milliseconds so this has the limitation but of the calcium the calcium signal but still when we analyzed the response of the order evoked response in cell assemblies of mitral and tufted cell it's the same it's the same finding basically we saw that some cell were excited and we see that as an increase of calcium response and other cells were inhibited as seen by this decrease of the calcium in respect to the baseline that was somehow controversial due to the comiana paper but basically we observed that systematically and on average this is not very visible in this case but we have roughly 50 50 percent of the cell responding it's not all the cell that are responding but the cell that are responding exhibit either an excitatory or an inhibitory response roughly in the same the same fraction and we also observe own response and off-respond that are available seen in the literature yes exactly although it was in anesthetized animal so this is in the way okay so I will not go too much into the detail of all this so basically the the calcium described the what we see in in electrophysiology but we lose the temporal the temporal dynamics the fast temporal dynamics that we have but we gain the ability to track the same cell assembly over time and look at remapping if there is any so I will just go quickly in this in these last slides so basically we see here the order evoked response in the population of cells that were recorded during the behavior this is the mixture mixture one mixture two so one it would be an s plus one the other one would be the s minus and you see here the representation evoked over time so this is the baseline activity and then you have here the other response over time and so what we do is basically we look at the correlation over time so we correlate a vector in time between this order and this order and basically this describes you how the representation is evolving over time so on the baseline there are there is low correlation when we apply the other these other highly correlated that what you see here and it remains correlated for a couple of hundreds of millisecond past past order so this is the first day when the animal is start starting to be engaging this behavior discrimination so these mixtures are highly correlated and what we what we notice that after six days of training when the animals are reaching 90 percent accuracy almost basically if we take the same two orders the and we calculate the correlation we basically see that the correlation now is becoming much less than it was on day one so here we basically see then this process of pattern separation taking place and so if we plot in the same mouse the behavior and the correlation in the cell assembly that we have imaged it's not all the cell it's just a fraction of the cell in the olfactory bug what we basically see is a direct relationship between the amount of pattern separation during learning and the increase of performance in this in this mouse so it does not really matter where we are imaging apparently as long as we are in a field where the cells are responding basically over days we start to have correlated orders that are hard to discriminate and during these days of learning we improve the decorrelation and we improve the the learn so I don't want to go too much into the detail but so some cells reduce their response other cells new cells are coming in the cell assembly so you have a complete remapping of the representation on average you have a you have a slight decrease of amplitude but it's not that I mean it's significant no no no so we just and this is the last the last slide we wanted to test the specificity of this finding so yeah this is basically mitral mitral cell that were imaged so we see this direct relationship between behavior and decorrelation if we do the same type of imaging in mice that are exposed to the same number of trial the same number of days but that are not engaged in the in the behavioral test so this is a passive training basically we don't see a very strong decorrelation there is a slight change but this is not really significant over the time course and so this would basically say that we have a context dependent remapping so if the mouse is engaged in behavior this remapping is taking place if it's not engaged in behavior in an associative learning this remapping is not taking place and the last piece of data was really was really surprising initially but based on what you heard already you may be not necessarily surprised anymore so because we have the ability to monitor now with optical sectioning either the mitral cell plane or the tufted cell plane we can analyze independently those two cells type that have been for many years considered doing the same but we know now that they project in different territories and if we do the same type of imaging in tufted cell we see that they also have highly correlated representation on day one but after learning these patterns are less much less separated than they are in the mitral and tufted cell to the extent that this if we if we relate the behavior in the mitral in the in the tufted cell to the correlation we see no real regression that was the case in the mitral cell so basically not only we have a context dependent remapping but we also have in the circuit a cell type dependent remapping so this is the the last conclusion of the paper so the inputs converging on both mitral and tufted cells here I showed them in two different glomeruli but obviously they share the same the same glomeruli so we have two different parallel pathway of information and interestingly only this pathway is submitted to be separated in terms of representation so the ensemble plasticity that we we have uncovered is specific to the to the mitral cell and the context so the tufted cell is not affected by the by this by this plasticity and the question is whether you know this would basically I mean why why would the the tufted cell be not affected so if you affect the representation by pattern separation you improve your ability to discriminate two orders that were at the beginning nearly identical but you lose information because they become separated and you don't know that at the beginning they were they were similar but if you maintain this line of information in the tufted cell in fact you have gain information you know that yourself your orders can be discriminated that they are different but you maintain the fact that they were similar at the beginning and so we think that this would be useful for basically learning to discriminate and this would be useful to maintain than the information that they were similar and so we don't know about the the the mechanism explaining this differential this free differential process we know that it cannot come from the input because they are the same for the two population of neurons but it's likely based on on the work of the new one the anyone what we know also in the literature that there are different types of granule cell the different types of cortical feedback impinging on different type of granule cell could potentially alter specifically the representation in the mitral cell assembly and leaving the tufted cell ensemble and unaffected and with this i would like to finish with the acknowledgement so the work on pattern separation was done by olivier schvendon mixon abram and samir lager and the work on the plasticity of the representation with two photon was done by kelly bocorali and yoshi maya yamada and this is a long lasting collaboration with my friend and colleague evan rodriguez and the the kcc2 story was done in collaboration with tomas yench lab and kathleen goddard and this is all the funding agencies that are paying for the release i would like to thank you for your attention