 Indeed, I enjoyed the past few days here. I learned a lot of things and met new people. It was extremely instructive. The place is beautiful. So I think the organizers, I hope you can organize it again and invite maybe some people again. Please stop me at any point in time and ask questions if you have questions. And I hope I'll be able to answer them. So I'm interested also in how the brain extracts information about the environment in terms of smell. And what we're going to discuss today is a few of these folks. I need them. Can I click on them? Oh, here. So smells come in nature in all kinds in plumes. They appear and disappear, as Andreas has talked to us yesterday, and they carry a lot of information about the identity, the intensity, the time, the position, and the balance of all kind of objects in the environment. The brain has evolved such a way that it extracts all this kind of information and it parses it in such a way that it's organized according to the needs of the moment. And sometimes you might need some information and ignore some other kind of information. And that may change as a function of context. So I'm interested in understanding how the brain is extracting some information and ignoring some other kind of information as a function of context and experience. And we think that the brain is using a mix of feed forward and feedback signaling. And one attempt is to observe the brain in action by monitoring and manipulating the activity of inputs and outputs and feedback loops as animals are engaged in behaviors and hope that if you collect a large number of such information, then you'd expect actually from some prediction and obtain some understanding of the brain. And the question here is really how the brain builds an adaptive model of the environment as the animal is interacting with the environment in open loop or in closed loop, and how it's possible to actually use mice that are engaged in a partial task to study that. Today I'm going to pose two questions. And I hope by the end of the talk I'll answer them. One is if different types of feedback projections from the cortex, from different parts of the cortex into the olfactory bulb, are carrying different kinds of information. And second, if different types of outputs from the bulb, the mice are in tough cells also carry different kinds of information. And this is going to be all done first in a wake and help fix animals that are naive, are not behaving. And if there's time in the second half of the talk, I'll introduce two different tasks in the lab related to reversal, learning, and closed loop olfaction. So as we heard over the past few days, and as we are all aware of, olfaction starts in the epithelium in the nose, as ligands are binding the receptor molecules. And the receptor neurons form these beautiful balls of synapses in the olfactory bulb. And here, the information is transferred on the dendrites of the mice that are in tough cells, which differ with respect to each other in terms of intrinsic properties and connectivity in the olfactory bulb, as well as in terms of operation targets. There is a zoo of local interneurons that is transforming the inputs from the receptor neurons onto outputs in the mice that are in tough cells and normalizes and decorrelates the inputs and makes them into outputs. And then the information is carried by the mice that are in tough cells to at least maybe 10, maybe maybe more areas in the brain. And most of these areas send back very strong and rich feedback that impinges on to all kind of inhibitory interneurons in the olfactory bulb. So as I was saying, the olfactory bulb is sending information across the whole brain. And these are the names of the largest areas, the anterior apoconucleus, MPD for cortex and tubercula and so on and so forth. And a few decades of research has attached some labels onto these areas in the sense that perhaps some computations would happen here, and this would range from localization of the stimulus, extraction of identity, say spatial navigation, and so on and so forth. It remains unclear if these computations emerged locally in these areas, or information is already sorted out at the level of olfactory bulb output as a result of interplay between the feedback and the feedforward signals that are sent in between the bulb and these areas. So the first question is, do different type of feedback signals from the cortex carry actually different kinds of computations? And there's a lot known, or some is known about feedforward signaling in olfaction, just like in other areas in sensory systems. But much less is known about the roles of top-down feedback. And just like in vision and in addition, a lot of ideas have been proposed about what is the role of feedback that is ranging from the idea of gain control to predict some prediction about the stimulus that is sent in the olfactory bulb and perhaps the level of mitral and tough cells, the information from the incoming stimulus and the prediction are mixed together and an error signal is sent back on the cortex and so on and so forth. But in terms of real data, there's actually a small number of experiments. So what we started to do is to use information from the wiring of the brain and focus on two areas, on the piriform cortex and on the anterior olfactory nucleus and use to our advantage the fact that perhaps the projections of the mitral and tough cells are not completely unbiased. That you do have, if anything, the prediction bias of the tough cells to the more anterior olfactory streams and therefore, we think that or have been shown that tough cells are projecting onto the anterior olfactory nucleus and the tubercula. Whereas at the same time, the mitral cells send information pretty much across on the whole brain with as Charlie has reminded us a few days ago, as a function of birth and a location on the surface of the olfactory bulb. So at the same time, what is also known from the wiring is that feedback that is coming from the piriform cortex or from the anterior olfactory nucleus is biased as well. So if you're to look at the feedback fibers that are coming into the bulb from the anterior olfactory nucleus and the piriform cortex, you'll see that the piriform cortex fibers innervate the deeper layers of the olfactory bulb and the input from the RNA is more spread out across all layers of the bulb. So what we started to do a few years ago is to try to label these steps of feedback fibers and ask how does the feedback impact the activity of mitral and taxi cells. And two great individuals in the lab, Honggu Che and Gonzalo Otazu, have taken this job. And they first asked how specific is cortical feedback? And they tried to test two different scenarios. In one scenario, the feedback is extremely unspecific and dense. It functions as a form of gain control. So in other words, as activity is ramping up in the bulb, you send some feedback into the olfactory bulb. That is impinging excitatory on the inhibitory interneurons and pulls down activity in the bulb to avoid saturation. And a different scenario that is extremely different is that the feedback is extremely specific and sparse. And signals particular predictions about its particular features of the stimuli that are sent back into the olfactory bulb. And their information is captured about the incoming stimulus. So to begin with, what Gonzalo and Honggu have done was to actually label these fibers by injecting a virus in the p-reform cortex. And they expressed this case g-cam 5 in the p-reform cortex. And they tried to then image activity in the olfactory bulb using a photomicroscope in the wick and hex mouse. So you can see here some nice labeling. And as you zoom in, you can take a set of optical planes in different layers of the olfactory bulb and image through the olfactory bulb. This is what we have done. And we have to show a Z-stack through the surface of olfactory bulb. You'll see a large number of buttons as you go through the Z-stack from the surface into the deep layers. Suffice to say that actually in terms of numbers of fibers, there are many more fibers coming into the bulb from the cortex and other areas than inputs that are coming from the novice in the epithelium. So what Gonzalo and Honggu wanted to do is to image the activity of these buttons in awake and naive animals simply by shining a light and presenting a large number of osmars. And what they found is, first of all, most of the buttons is wanted to know stimulus that you have sampled. It's still a small sort of stimuli that was chosen somewhat arbitrarily, but about 65% of the buttons do not respond to any of these stimuli. And those that did respond, respond is so very sparsely. So if anything, they respond to maybe one to two different odorants. And that's not simply because there is something funky happening with the indicator, because if you stick an electrode in the pediform cortex and stimulate the pediform cortex electrically and image the activity of the buttons in the bulb, you actually can drive them. So the buttons have the ability to respond to odorants. It's so happened that within the range of stimuli we have represented and of intensities, the responses are extremely sparse and all specific. And if you were to take a comparison, there are about 10 times sparser than the activity of the output neurons of the mitral and tufted cells. Second, they simply observe two kinds of responses which you'd expect normally an enhancement of activity with respect to baseline and the suppression of activity. And in any circuit of neurons, you'd expect that to have a mix of excitation and inhibition. But the striking part came when it seems that there are two independent channels of information. Some buttons, although they respond very sparsely to odorants, if they respond at all to any odorants, they respond with the same sign. So if a given button shown here, each row corresponds to a particular single button. And here is the odorants. So if a given button is responding by enhancing its activity with respect to baseline to particular stimulus, if it responds at all to any other stimulus, it responds also by enhancing its activity. And that's true not only for the sparsely responding buttons, but also across the whole population. And the same is true if you look at the suppressed buttons, if they're told to respond to odorants, if they respond by suppression, they will respond by suppression across all odorants. So the traction of buttons that respond by enhancement and suppression to different odorants is very small. We don't really understand why this is happening in terms of its substrate, if it arises in the cortex, or if it's a mix of that and local modulation in the archibald. But we speculate that perhaps these two channels are used, indeed, as ways to make a prediction about the stimulus. And the kinetics of the enhanced responses and the suppressed responses are different. The enhanced responses are faster and follow the stimulus while the suppressed responses are actually quite slow. So we speculate that perhaps the enhanced responses represent the similarity between the incoming stimulus or the prediction that is made in the cortex about incoming stimulus and a previously stored information about a particular stimulus. So you could imagine that in the cortex is a dictionary of elements, of objects, or partial objects. And as you receive an incoming stimulus and odor, you compare this incoming stimulus against your set of objects and inform and extract how similar this particular stimulus is. The second signal, which is actually a suppressed signal, may affect the uncertainty. How certain or uncertain you are about this particular stimulus. And to begin with, the uncertainty is high. And then as you sniff the stimulus again and again, the uncertainty is going to be low. And it will stay low for a while as used for the environment. This is all speculation. We have no evidence for this. I'm starting to do some experiments to actually test it. But the focus of the site, to begin with, was to understand what is the impact of the preformed cortex and partial nucleus and the level of the mitral and the entire cells. And to begin with, what we did was to do a very coarse manipulation, simply inject a drug in the cortex, so most of all, that will actually silence activity of the cortex, and image the activity of the mitral and the active cells, again, in the wake, and naive animals that are sampling these odorants. This is an example of the experiment. This is a field of view of mitral cells in the wake and have this animal. And each of these is a cell body of a mitral cell who presented about, say, 50 odorants across your magnitude. And each row here corresponds to the response of a particular cell to a particular stimulus and a particular intensity. You can see that as the stimulus is presented here, there's an increase in fluorescence. Now, as you monitor the activity in this small field of view and do some manipulation in the cortex, you can assess how many cells are responding to this particular stimulus before you suppress the cortex and after you suppress the cortex by introducing muscimol. And if anything, you would realize that after suppressing the activity in the cortex, both the number of responsive neurons, as well as, if you were to, let's say, pick a few neurons here and show the response vectors across all these odorants, the sparseness of the responses is decreasing. In other words, they respond to many more odorants, which is, in a way, it's not surprising, because the feedback is excitatory and impinges onto inhibitory interneurons. If you remove the feedback, you have less inhibition. And therefore, you expect to see some increase in activity of the outputs. However, the picture is very different if you are to look at the activity of the top cells. That is the second class of output neurons. And that's shown here is a few such example cells that are shown for a particular smell before and after muscimol. And as you can see, both in terms of the numbers of activated cells and how many odorants are given cell responding to, that number is not actually changing much. So somehow, at least in terms of this course metric of response amplitude, the mitral cells are affected much more by suppressing the cortex than the activity in the top cells. Yes. How sorry if it's too difficult? Yes. The tuberculosis doesn't send back feedback, but the AON does, and the second part of the talk is exactly about that. So you're on the right track. So we wondered also in terms of representations of odorants at the level of the mitral and top cells, let's say a particular odorant would activate a set of mitral and top cells in a particular set of spatial and top activation pattern. And if you were to be an observer that is reading information from these mitral cells or top cells, you could compare the representations of different odorants and assess how similar or how different these representations are, the level of either mitral cells or top cells, simply on the first pass by looking at the overlap in the responses across all these cells in the same field of view. So you can compare, in a way, the cell response vectors for different odorants and assess how similar or how different the representations of different odorants are. And to begin with, the similarity between all pairs of odorants that you have tried and the level of mitral cells is quite low or slower compared to the similarity of pairs of odorants level of top cells, once you block activity in the peripheral cortex, this similarity shifts to the right, which means that in the absence of the cortex, the odorants appear to be more similar to each other from the point of view of mitral cells. In other words, if you were to be an observer that is trying to decode information from these mitral cells, it would be harder. Whereas in the case of top cells, there's no much difference. So again, suggesting that the effect of feedback is specific on the mitral cells somewhat and not on the top cells. A different way to put the same observation is if you're to look at how many dimensions, if you're to explain the activity of the neurons in terms of variance, of how many different dimensions you need to actually capture these variants, as you block activity in the cortex, the number of dimensions that you need to explain activity of mitral cells is lower. In other words, the response has become less complex, whereas the activity of top cells actually is unchanged. So I think we have uncovered a way of saying that the feedback from the peripheral cortex aids specifically the separation of odorants only at the level of, or mostly, at the level of mitral cells and not so much level of top cells. And clearly, we have wondered what is happening with the activity of these guys. Since in these, there is a bias in projection, the mitral cells project strongly to the partial cortex and less so to the anterior partial nucleus, whereas the top cells project more to the anterior partial nucleus and less so on the peripheral cortex. We wondered if there is a similar thing happening from the point of view of top cells. So we asked, how does the feedback from the AON impact the activity of top cells? And we did exactly the same experiment, the image activity of top cells. In this case, this is an example of a control experiment where you inject the saline in the, yes, yes, yes. So on average, that's one idea in the time that I should test it. The projections from the peripheral cortex seem to innovate the deep layers of the bulb, more so than projections from the anterior partial nucleus. And indeed, they may actually synapse on the deep internal cells, whereas the projections from AON may synapse across all layers of the cortex. That is our thinking. Yes, and a few groups have shown that already, including the gentleman. We have not yet looked into that yet, exactly. We like to be able, ideally, to look in different colors at the feedback fibers and the activity of the ground cells. And more than that, do manipulations and suppress activity or the feedback fibers in a particular sub-regional of a cell. But that's for the future product. So we started extremely coarse manipulations. And in this case, looking at the effect of injecting mousse in the AON on the top cells, and indeed, compared to the cell in control, there's an increase in activity. It's a first activity of the AON at the level of tufted cells. So similar to what we've seen in the case of mitre cells for the peripheral cortex, if you suppress the AON, the firing of tufted cells is increased. And in this effect, it's much stronger. So here, what is plotted is the distance from the regression line with respect to saline for tufted cells. So each dot here is one cell order pair. Of course, this is the order we have used. And this is shown from the point of view of tufted cells as you block activity of AON or of opyriform cortex. And you can clearly see that the central effect is much higher for the suppression of AON. And complementary, if you are to look from the point of view of either tufted or mitre cells if you suppress activity in the AON, the effect seems to be much stronger on the tufted cells compared to the mitre cells. There is actually a small effect on the mitre cells as well. But there is a strong preference towards boosting activity of tufted cells as you suppress activity of AON. So it looks like there are two different channels. And in terms of representations at the level of neurons of odorants, again, you can compare different odorants in terms of the overlap in their responses of mitre and tufted cells. And as you'd expect, as you block activity in the AON, the similarity of odorants from the point of view of tufted cells is increasing. In other words, if you are to be an observer and try to extract information from the tufted cells about the similarity of odorants, it would be harder in the absence of input from the AON. And this effect is much stronger for the tufted cells compared to mitre cells if there is activity in the AON. So it looks like perhaps there are two different streams of information both feedforward as well as feedback that will help some matching of the biases in targeting of feedforward projections as well as in specificity of feedback, in such that mitre cells project most or strongly to the pediform cortex and, in turn, the pediform cortex through a set of interneurons that we are still trying to study is suppressing activity of mitre cells and a similar activity of tufted cells is reflected in AON and then back suppressing the activity of tufted cells. So if this is true, then you do need a substrate and we're trying to understand if this is actually happening at the level of the ground cells in B. And the idea is that both AON and pediform cortex project and synapse onto ground cells, maybe the pediform cortex in the deep layers of the bulb and the AON across a few different layers and, in turn, the interneurons inhibit activity of tufted and mitre cells. So we started to actually image the activity of the ground cells in a wake and eat animals, ask how the activity is impinged upon if you suppose activity in the pediform cortex or in the AON and see if this idea is actually true or not. So perhaps there are actually different subsets and these cells are known to be many. So actually if you look at the bulb, about maybe 90% of all cells in the bulb are ground cells and they're also known to be very silent if you record in anesthetized animal. This is not true in a wake, even a fixed and naive animal. There's a lot of spontaneous activity and upon blocking activity of the cortex, the pediform cortex, we noticed that we see both a decreasing drive, a spontaneous drive, as you'd expect, but also in some cases, you're actually seeing enhancement of activity, that as you move activity of the pediform cortex, actually in a subset of cases, you actually boost the spontaneous activity of ground cells. And that's a bit surprising because the expectation is that the feedback is excitatory and impinges on these ground cells, but I'll maybe tell you why we see that as a case. This, the ground cells are labeled using CST, Cree. CST labels all kinds of neurons in the body of the bulb because all there is pretty much, we simply image deep in the bulb. But the same kind of data you can get if you simply inject deep in the bulb a virus and image, you will see very similar activity. Yes, yes. And also on top of that, you can find ways to label in different color the deep short-acting cells, for example, and compare the activity of the short-acting cells and the ground cells. Yes, question? Yes, that is the problem. That although we claim we are good at imaging, we are able to only image the superficial layers. So, superficial means as deep as about 500 microns, but this is still somewhat superficial because if you go to the structure of the bulb, there are going to be layers of ground cells, layers after layers after layers under pretty much 800, 900 microns deep. And the dendrites of these deep cells probably are connected to the lateral dendrites of mitral cells whereas the most superficial ones may contact the dendrites of octavid cells and mitral cells. So, we are limited to some degree by how deep we can image and we are trying to actually go around it as I'll talk about in a second. Also, if you are to look at other induced responses for particular cells, cell A and cell B, as you remove activity in the cortex, the response is abolished. In some cases actually you boost the activity of the cell. And interestingly, this effect appears to be cell specific and not odor specific. In other words, if at all a given cell is responding to any of this arbitrary odorance that you are presenting, if you see an effect because in about half of the cases you don't see any kind of effect of blocking activity in either the cortex or A1. If it's an effect, it seems to be a specific sign. So, the polarity of the effect is either enhancement for all cell odor pairs at this particular cell response. So, for all odorants at this particular cell response too, all these are enhanced and similarly it happens for this cell and this cell and this cell whereas the responses of these cells are all suppressed. And perhaps we still don't understand actually how this is happening but it suggests that you have very specific connectivity of the feedback fibers at a level of particular cells and perhaps these two channels of enhanced and suppressed boutons may be biased in a way. Yes, question. So, in our hands, it could be simply our hands. In our hands about half of the cells we do not see any effect of the drug. And the rest half about a quarter were enhanced and about three quarters were suppressed. So, as you'd expect, the more frequent event is suppression. You remove the excitatory drive but in about a quarter of the cases you actually see boosting activity. And perhaps one explanation for this which is again a bit hand-wavy at this point is that feedback doesn't go only onto those on the other cells but also on many other interneurons and maybe even also on the open neurons. And one class of such interneurons has been described as deep short axon cells although they actually have very long axons. And these particular cells get strong feedback from the cortex and in turn they actually inhibit the activity of ground cells. You could imagine ways in which by removing the excitatory drive you could actually induce both lack of activity as well as some disinhibition of activity of ground cells. So, I told you that we started to actually look for are there any classes, any two distinct classes of ground cells that are specifically affected by manipulations of AON or PD from cortex? And currently we don't have an answer yet, simply because we are not able to image as deep as we like to do so. And what we try to do currently is to do some damage to the bulb, implant a small prism in the bulb and let the bulb recover. Actually the prism is a small mirror in a way that looks like a prism. So you shine light from the side and image across all layers of the bulb after a few weeks after the animal has recovered with a big gadget you make a hole in the circuit. Alternatively you are trained to use a different forms of microscopy, three photo microscopy and image deeper and deeper in the bulb but that will take some time. So it looks like if the ground cells are the substrate of this feed forward feedback interplay, we still don't know if indeed we have not found that it is true that there are two classes of ground cells but in the process they discovered that perhaps the suppression is of feedback has violation effects of the level of ground cells and the feedback seems to be very specific by in terms of cells and not necessarily of odorants at the level of the effect on the cells. So I told you in the beginning that, yes, question. I'm not very sure at all. Can you imagine that? Yes. Yes. Yes, so there is the dirty secret of any kind of a muslim experiment is that you don't know for sure how much the spread is. What you do is in each experiment we inject into different locations at different sites in each location and then after the experiment we quantify roughly how big the spread is by using at the same, so you inject muslim oil that is fluorescent and you can see how much is spread although you can spread quite fast. And but the answer to your question is that you don't know for sure and therefore this is just one first attempt and what we currently are doing is to replace muslim oil which actually in the first place is a bad tool to use because you block activity of the cortex and that's not what you want to do with tools that allow you to shine light in the bulb so you express a virus in the cortex and then instead of suppressing activity in the cortex itself, you try to shunt activity of the fibers and I'll talk about it in a few minutes but it's a good point. It's any other question about this? I don't. You think there are different cells because the responses are so strikingly different? I mean, it's like a surprise. Yeah. That's what the point is always. Yes. I mean if you believe that indeed the activity that we see in the axon is the terminal is a proxy for the firing of the neuron that is true could also be that there is some complicated and not understood mechanism of local modulation in the bulb that I don't really believe but it's true that actually if you record in the cortex people have reported that they can find cells that are mostly enhanced and mostly suppressed so it could be that indeed they are certain in the cortex. Now if this is a genetically identified type of cell obviously arises simply by wiring by what kind of impulse it gets. I mean it's a matter of debate what you define to be a cell type. So I have no idea but I hope that one day this gentleman here may find some markers that would allow us to label them. Yeah, so it's still unclear. Yes. Yes, yes. They are just a no special arrangement. No special arrangement. We lost hope of any kind of special arrangement. And not even loaded a package. Yes. So in the same small neighborhood you can have all kinds of effects. Okay, so I started from the feedback side but clearly there are two streams of feedforward information the mitral and tufty cells and Orco and Honggu and Prade and Lab have wondered to begin with if they carry different kinds of information in the beginning from the point of view of extracting information about the concentration of the stimulus and identity of the stimulus or extracting identity information independent of concentration. And if you're just to image the activity of mitral and tufty cells again in awake and naive animals you would immediately realize the responses of these cells and by no means where the first ones to show this has been shown over and over again they are actually quite different that the responses of mitral cells are quite sparse and not so monotonic as you increase the intensities of the stimulus whereas the responses of tufty cells are quite dense so they spawn a lot and as you increase the intensities of the stimulus you get stronger and stronger responses. Now, if you are to average the response of a particular cell for a particular concentration and then plot the response kind of across different concentrations for this particular cell for both tufty the mitral cells becomes quite obvious that you can't really rely on single cells to tell, to extract information about identity and stimulus intensity because as you change the intensity and the identity of the stimulus the response of the cell is changing in quite an individual fashion. So therefore individual cells are poor encoders of both identity and intensity and then the next best regime or the next simplest regime will be a linear regime indeed if you look at a set of neurons and ask yourself if the behavior of mitral and tufty cells is amenable to linear decoding or not and I would argue that if you plot the presentations in this schematic of two other ends in terms of the firing rate of N neurons in this space if across concentrations you behave monotonically or these neurons actually are responding in fashion then you send us a better chance to separate the identity of these two other ones if the responses are monotonous versus if not. So our prediction was that tufty cells ensembles would be superior to mitral cells at linear decoding of identity of the stimulus in a concentration invariant manner. It's unclear if the brain does linear decoding in any way but this is just an exercise to begin with. So we try to train a classifier and formally say for a set of five other ones across four different concentrations they say train five hypothetical classifier neurons to extract information from the experimental data from the mitral and tufty cells with different weights and the train them using sparse linear sparse logistic regression and impose some constraints in such a way that these weights are sparse so you have a small number of non-zero weights and you minimize the difference between the predictions of the network and the reality such as to match the presentation, the sparses of presentations that you observe experimentally. And for example, the job of order one neuron would be to respond to any concentration of stimulus one but only to stimulus one, the job of order two neuron would be to respond to any instantiation of stimulus two and so on and so forth. So if you do that and you train the classifier and impose the constraints and challenge the classifier with our set of data and some validation, then you actually see that indeed the performance of the classifier is much higher for the activity of tufty cells compared to the activity of mitral cells. So if the brain would be to do linear decoding of identity in terms of invariance of concentration the tufty cells would actually be superior and that's perhaps becoming more interesting as you go back and think about the production patterns of tufty cells. If it's true that tufty cells project mostly to the anterior partial nucleus and the tubercula then perhaps these areas also have access to or are also in a position to compute information about the identity of the stimulus and it's not only the peripheral cortex that has been highlighted over the past few decades that is exactly this information. I mean obviously the information from the AON is going to the cortex and vice versa so there's a lot of crosstalk but this suggests that there is information that leaves the bulb that does not need necessarily to pass through the partial cortex you can go to the AON and the tubercula and these areas may be in a position to compute the identity of the stimulus and by currently doing experiments and recording from AON and tubercula and the peripheral cortex and see if this is actually true. More than that, what we have shown you so far was pretty much an exercise of futility because all these animals are awake and head fixed. They're not doing anything they're just experiencing these odorants. We like to understand if these loops of feed forward and feed back signaling are engaged in the way that we predict through these humble experiments as mice are actually engaged in concentration reporting and identification of odorants. The same animal is going to be trained in reporting either changing concentration or in a different session of a change in identity of the stimulus and what we like to do is to suppress activity of the feedback from the peripheral cortex or feedback from the AON and image activity of mytolyntalpha cells and test if these manipulations have differential effect in terms of reporting one feature or the other or perhaps these two features of concentration identity are actually not what these particular areas are important for but something different. Okay, since we got very interested in the specificity of projections, we were a bit obsessed about the idea whether far too bad projections to different areas are organized by the information content or you can rephrase a question. Do select sets of mytolyntalpha cells carry similar information? If they do carry similar information do they project to similar areas? And perhaps to begin to answer this question we first posed a question about the anatomical projections from the mytolyntalpha cells to the rest of the brain. Are there any biases at the level of single cells? Single mytolyntal cells or single sets of... At the level of single cells can you see any biases in terms of what areas the particular cell is projecting to? And clearly this question has been posed over the past few decades and the classic approach is to inject the dye very sparsely and label one or two neurons and reconstruct the whole axon of the neuron across the whole brain and see if any biases. And we are extremely lucky that we are taking advantage of new developments spearheaded by our neighbors in Kostpihara by the Zeder lab. So we are using DNA barcoding of neurons and sequencing to answer this question. And the way that I'm doing it is the following. So these people are doing the work is Justus and Shaoyeen in Tony's lab and Yushu and Pedro in my lab. And the idea is I would say elegant and simple and robust. It's Tony's idea and he, my mind had quite a brilliant idea because it's very simple. All what this method entails is to express a large diversity of barcodes, virally, and do it so in such a way that a given barcode is going to infect one cell at a time. And then if you are interested in the statistics of projections of neurons from one area to n other areas across the brain, all you have to do is to extract the barcodes from the injected area and the areas of interest and sequence them and determine what barcodes are present in area one and two or one and two and three or two and three and so on and so forth. So in one experiment you can actually, because you can express a large diversity of barcodes, a few hundred million by now, you can be quite sure that you can uniquely label on the order of a few thousand cells per brain. So as you go from having one or two or three cells labeled per brain and basically actually slicing the brain and constructing it, now in one go if you are not really interested in the high spatial resolution, simply understanding what are the chances of cells from the Focci-Balt project to the tubercula and the AON, but not to the amygdala or to the amygdala and the pediform and so on and so forth, then using a chance to actually do that. Now we done a professional experiment with this idea by labeling a small nucleus LC that contains about two thousand neurons and the field was a debate whether these neurons which are projecting across the whole neocortex and the opalchic bulb, are they ramifying? So a given neuron, does it send branches across the whole neocortex and the bulb or particular sets of neurons project to particular areas? And I'll say the classical model, although it's been a debate for a long time, was that the neurons are projecting discriminantly across all areas. Using this idea, this map-sick idea and injecting the LC, you actually see a very different picture. So here what is plotted is the identity of the neuron which is the identity of a particular barcode against different slices that are taken on the anterior posterior axis of the brain. And the color or the intensity reflects relative frequency of a particular barcode in a particular area across all these slices. So you can see that a large number of neurons have very specific projections to the opalchic bulb and then across the whole neocortex. So it appears that perhaps these neurons actually have a bias in their projections, it doesn't mean that they are doing actually different things, but at least anatomically, they are projecting in differential fashion. So we wonder if by any chance, a similar or a biased view we merge for the projection patterns of mitre and tufty cells. And we have been trying to inject the virus in the opalchic bulb and slice the rest of the brain and extracting tissue to begin with from six different areas and determine the statistics of projections. This industry has the effect of all cells. Yes, so on the first class, we don't know, we don't know if they are on the social or tufty cells, but what we are doing currently is to actually slice the opalchic bulb itself and pick cells from the dorsal surface, the ventral surface, so the dorsal arterial aspect in the ventral aspect and also separate between the mitre cells and tufty cells. But in the first class, no. And yes, yes. No, the conclusions are actually consistent with each other. There's just a difference in methods. So the vision room lab has not had the ability to have single-cell resolution. So if you're to actually look at bundles of cells, you can, or if you're to pull together cells, you will see that indeed you get projections across the whole, exactly. Yes. But in any case, I agree with you that you need a separate metric to actually test these ideas because you are limited by how much RNA you can extract and how much volume you can extract and so on and so forth. So what I'm trying to do is to compare this with what you'd get if you do imaging as well. Now, if you take this approach to the opalchic bulb and this is, again, very preliminary data, we only have two brains analyzed and these are the first two brains that usually the person who is doing it has actually, has ever touched. So perhaps you should keep that in mind. It looks like, perhaps you do have some biases that these are the barcodes extracted from AON, the pyriform cortex anterior and posterior, the tubercula, midline cortex and so on and so forth. So perhaps you do have some biases in some cells projecting mostly to the AON but not to the pyriform cortex which will actually be consistent with the idea of getting the mitre and tuxed cells separated, but more than that, some new biases will actually emerge that perhaps some cells are biased because the tubercula and not so much the AON and some to the AON and so on and so forth. So this is all very preliminary, I would caution you. Now we're trying to do this properly and really slice the brain at fine resolution and from each slice extract all these areas of interest and in a way for each barcode create a vector of projections. And then what we like to do, which should be probably a very tedious process, is to compare the functional responses of the mitre and tuxed cells in vivo that you obtain by imaging with the odorants and ideally not with arbitrary odorants but odorants actually mean something to the animal in amounts that is actually engaged in a particular task. And then use the beauty of in situ sequencing. So there is a very fashionable technique for instance in situ sequencing that is extremely inefficient but still it gives you the ability to sequence in situ which means that you can identify the same cells with image in vivo and you've played your odorants across in the impact brain in the slice and then sequence the barcode from this particular cells and then determines through MAPSIG the projection patterns of these cells. So you'd actually go from the functional tuning of the cell to the location of the cell in the farchibald and the position patterns across the whole brain. It could be again that you will not learn much from it, there's no guarantee, but it could also be that in an extreme other scenario that actually perhaps you'll learn something about some encoding of particular stimulus features as you compare the tuning of these cells to odorants and the position patterns across the brain. So initially I posed two questions. One was the different fit for the say mitre and toxic cells projections can have different kind of information and second the different feedback projections implement different kinds of computations. And I would say I'd show you some evidence that perhaps the ensembles of toxic cells are superior to mitre cells in terms of linear decoding of identity of odorants independent of concentration and the feedback tends to be specific. The feedback from the pediform cortex mostly affects the activity of mitre cells and therefore my own most efficient field of toxic cells although we don't really yet know what mechanism. And now I'm not sure how much time I have. Five minutes. Do you have any questions about the things that I talked about so far? Yes, Tim. Yes. Yes. Exactly. Yes, yes. If to be more... We are trying to actually look into that and I hope I have an answer soon, but not yet. Yes, question. Can I go with one? Yes, that is true. So in this case, what I've shown you was acquired at... So in the beginning we started with Gcam 3 which was extremely slow and also with slower scanning microscopes was acquired about five hertz. These days we are imaging at 50 to 100 hertz so the Gcam 6F is still... It's calcium imaging so it's slow but still you have the ability now to actually bin and by recording respiration as well so you're able to look at single SNFs and separate the activity of the first SNF and second SNF and so on and so forth. And indeed you can see that the response of toxic cells predates the activity of mitre cells. But I think it's more than that because you can actually increase the integration bin and include both activity of toxic cells and the mitre cells and on top of the fact that they spawn early you can actually separate them more easily in the space of. It's a speculation. It's a speculation. We have no recordings to show that. It's just that the projections are going there. No, no, so... Yes. Yes, yes. So I'm not saying in any way that the responses to concentration are flat. As you increase the concentration in this case for the toxic cells the responses are monotonically increasing but the fact that they follow a simple function like this they simply are going up as you increase the concentration makes it easier if you're to be an observer to apply a boundary and separate the presentation of order A and order B as long as you learn that order A is encoded by some sort of neurons. So it's easier to separate order A and order B in a linear way because the response to the neurons are actually quite simple to describe by a simple function. But you're adding a digital neuron. Yes, you do. Yes, yes. Yes. You have to do that on a slide. So you have to... Yeah, you have to take these stacks. You have to slice the brain. You have to then find the same cells from in vivo to in vitro. It sounds very awful and indeed it is awful but it has been done by a few groups already by Tom Merchage Flugger and algorithm's help actually. It turns out that it's not so hard these days but it is indeed tedious and you'll have a small number of cells to begin with and then there'll be a few iterations until you actually get better. And even on top of that, the efficiency of physics or physics of different versions at this point is quite low. So you will, this is not going to be to begin with a high throughput method but because it's sequencing and because it's a large market, my feeling is that my prediction is that because of that in a few years, this would become standard. So now it's crappy. It's very crappy. Low efficiency and it's tedious but I would say that it has some hope for it at least. Okay, maybe if you allow me for two minutes, I want to tell you something about the behaving brain. Not only some animatis is there, no? One minute. Okay, then I will skip. I will skip all this. All this and I will, okay. I will just give a teaser about something that I'm extremely excited about at this point. It's very early on but to my mind it has the power of answering some interesting questions. That is having an anima that is head fixed which allows you to do manipulations and imaging and so on and so forth. At the same time, having the animal in charge of the stimulus. So there's clearly a huge difference between having animal that is passively exploring is getting all kind of inputs in the environment either in the case of vision if you are static and this is some visual stimulus is coming at you versus if you are actually immersed in the environment and similarly I would expect that if the animal is simply standing there or sitting there for many, many hours and smelling all kinds of things and then it makes sense of the world. The part is very different from an animal that is moving but perhaps you can still have an animal that is head fixed which has some advantages but give the animal the choice to control the stimulus and in my in the lab have done exactly this and this task the animal is head fixed and is learning to operate a joystick, a lever and by doing so is engaging a stipper motor and the stipper motor has a belt attached to it and the belt remounted a manifold that can deliver odorants. The manifold has a central tube that you can support that you're going to push the odor through and a few ports on the left and right side that you use to give air such that the flow is not an issue. So what you have to train the animal to do is to use the feedback from the odor itself and localize the odorant without actually moving. So by moving the sliver they are going to start mapping the position of the smell on this axis so they move the lever like this and the smell is moving on this axis. They could also move the lever like this in principle and the smell can go on this axis but to begin with moving like this and the smell is going on this side and the task of the animal is to bring the odor in front of the snout and keep it there for a while. So the animal is head fixed is initiating the trial by pulling the sliver and holding it there for a bit and then the task is to center the odorant and then if it's sent as if it keeps it there for a few hundred milliseconds, if it does this it gets a reward. The animal is thirsty so it's going to work for water and then it can keep it there for a while, it can release it and then a new trial starts as the animal is pulling the lever again. So we train these mice in such a way that they cannot rely on a single motion, that you don't simply rely on one trajectory but you actually have a large number of target zones. You are not simply learning a motor trajectory and we don't use a single odorant but a few such odorants. So they simply have to rely on the feedback from the stimulus independent of stimulus A or B or C and localize where the stimulus is. All these experiments are done in the dark. Okay, so this is an example of such animal that is behaving and was plotted here. The animal that is behaving, these are its licks. Each bar, vertical bar correspond to the position of in the trial where the animal has got its first reward. These black trays corresponds to the location of the lever itself and these are the target zones. So the animal is set fixed and it's moving its lever and the job of the animal is to take this little mark and put it here in this zone. The animal has actually learned to do that and once they realize they can control the stimulus, they can help some agent on the environment, then they, at least our minds, they seem to be much more engaged in the task and they can learn this task, actually it took us a while to learn ourselves how to train them but nowadays, they can learn within two to three weeks to go up on the road about 80 to 85 percent correct trials and what we try to use this for now is to actually understand how localization of the stimulus is happening and more than that, in the longer term to look at sensory motor transformations and understand the logic of motor feedback signals into sensory areas and even more than that to understand the target and background extraction because you can add a second such manifold that is moving dependent on the first one and in different trials, you can have a target in the background, the animal is in control of one or the other so it can extract information about the particular stimulus independent of it being in control or not. I'll stop here and I'd like to think all the people that have been actually involved, Honggu and Gonzalo have spearheaded a feedback project, Orco and Fred and Honggu might have themselves in Bianca and married the legal task and is our sources of funding and influence about this. Thank you. Thank you.