 The organizers were inviting me and also whoever canceled the last minute and gave me a spot. And I should say before I begin, everything I'm going to tell today is work done in Venki's lab with other people involved. I'll mention them later. And what we're interested in is figuring out how mice extract information about things they may care about in a cluttered environment. And the idea is that we all assume that mice depend on the sense of smell for many daily functions and we believe that the objects that they care about are never presented alone. So somehow to be able to identify the cheese here, this mouse will have to segregate the odor of the cheese from other odors in the environment. And the way we think about this is in a factory version of the cocktail party problem, the auditory cocktail party problem, in the sense that signals from different objects activate overlapping sensory neurons in the sensory organ. So the brain now needs to take this mixed signal and kind of recover the original sources that gave rise to this mixed signal. So before going into the neuron mechanisms, we wanted to have an idea of how well we understand the behavior itself. And by understanding behavior, often we take this approach, the psychophysical approach, that basically says if you can predict the performance of an animal based on some parameter of the stimulus, then you have some good understanding of behavioral ability. This is what we wanted to do here. And to do this, you need two different things. You need to be able to measure performance and you need to be able to define and vary some parameter of the stimulus and see how that relates with performance. So let me start by describing this arm here. So to do this, we devise a task in which mice are required to report detecting a target odor. And the target odor is presented with random sets of background mixtures, pseudo-random. In more detail a little bit, the world for these mice was composed of 16 monomolecular odors, odorants, of these two for each mouse were selected as the targets. We varied the targets for different mice. And by targets, I mean if you can detect these odors, you can get a reward. The other 14 odors are distractors in the sense that they may be present, they carry no information about a possible reward. And then in every trial, we present a mixture of these odors. And these mixtures are binary in the sense that each odor is a binary variable. So odors are either present or not present. We do not play with concentrations. And the job for the mouse is to just tell us whether one of these two is present within this mixture. And they do so by licking a water spout, if they do it correctly, they get a water reward. So we trained a set of mice on this task using this kind of training scheme in which we start with, so what I'm showing here is the distribution of the number of components in the mixture. So we start with sessions in which most trials actually had just one single odor. But some trials had more than one, still low numbers. And then as mice learn the task, we progress until we reach this flat distribution in which the number of component, the chance of getting a seven component mixture is the same as six or 40. Yeah, so we go through two different distributions. These are four defined distributions. This is one mouse. This is, I think, the fastest learning mouse. And this is already the flat distribution. So one of the nice features of this task, because I mean, everyone here uses olfaction and many people use mice, but one of the really nice things is that this is really quick. So in about 800 trials, they get this distribution. And of course, they mostly need to learn the test rules. Everything is arbitrary to begin with. And then as you increase the complexity of the task by having larger mixtures, the drop is not huge in performance. All the data I'm gonna show you is from sessions in which we use flat distributions. So one more nice feature of this task is that just because of the combinatorics of 16 different odors, you can produce a huge number of different mixtures, way more than the number of trials that any mouse ever does. So that most of the trials that a mouse encounters are novel mixtures he's never seen before. So there's no way to make a lookup table and just remember combinations. You have to perform the actual test. So we trained 13 mice collected a little over 30,000 trials. And when you look at the performance, first thing to ask is whether it's even sensitive to the background. And you can see that increasing the number of, I'm gonna move this. Increasing the number of components in the mixture reduces the performance of the animals, which is kind of expected. Another thing you can see here is that because the task is asymmetric in the reward, there's not much information in the go trials. They basically never miss the target. And the way we interpret this is that when mice have any doubt that the mixture may be there, sorry, that the target may be there, they just go for it. So going doesn't tell us that they knew that the target is there. But when they decide not to go, then that really tells us they knew that the target was not present. So if we want to understand what makes them fail, which is really what the psychophysical function is about, we really need to analyze the no-go trials and that's what I'm gonna show. So, I mean, you could think of this as some sort of psychophysical function. You have a parameter of the stimulus, that's the number of components. And that relates nicely to performance. But it's not extremely pleasing because you don't think that all ten component mixtures are equally difficult, right? So there needs to be something that depends on the specific target and background molecules that will make a certain ten mixture component more difficult than another. And we wanted to figure out what this is. I'll just take a small digression. We tried to do this with humans and we did not complete the project. So I'm not gonna make any extremely strong claims. But one of the reasons we could not complete the project is that we basically could not get humans to perform at any reasonable level. So maybe we didn't do it. We did it basically the same way as we did with mice, same designable spectrometer, just less odors. And we didn't just ate in the spectrometer, we just couldn't do it. We tried to train ourselves to do it. It still didn't work. If any human or faction person is interested in figuring out whether we can actually get humans to do that, I think that would be fantastic. Yeah, so the number is very low. So in Lang's papers, they basically did a very similar task. And there were two things that I looked at that I thought may be the difference. One of them is that the performance on a single component is kind of low for the humans. So I thought, get good performance at least on the single component before you say anything about the mixtures. And the second was that they don't train the subjects and we train our mice. So what if we train people? And we try training ourselves, but it didn't seem to work. Yeah, right, they tried to make the task as easy as possible by letting people rate familiarity and similarity and choosing the ones that are not similar and familiar. So and we didn't do that. So I'm not saying that humans can never do this. That's not a claim. It's always cinnamon that people use for. So I don't want to make any strong claims that I think one claim is that at least humans are far away from mouse behavior in this respect. So we use the same, less odors, kind of the same odors. Most of them are the same. We train mice easily. We don't train humans easily. So that's the only real claim I have here. Okay, going back to mice. Yeah, so we can debate about the synthetic maybe later. But I think in my view, once you've formed an object in your memory, it may not be synthetic anymore when this object is presented with other things. I agree that if you, and in a way, maybe these targets become an object for the mouse, right? I don't know. But if you just mix unknown molecules and smell this, you won't know that it's a mixture, that I agree. But I think we can have this debate later. Okay, so we want to move beyond the number of components and figure out some other parameter, right? That will describe the relationship maybe between target odors and background odors. And of course, the difficulty that was discussed here by multiple people is this issue of chemical space. If this is the target, these are two different background odors. How would I say which one is more similar? How do I define parameters that relate them? So one way would be to take one of the spaces that people here work on. And other ways would be, you know, you can take it in just a complete intuitive view. And let me explain what I mean by this. So these are the 16 odors that we used. And half of them were to Glake assets, okay? This structure here within the red frame. And the others were not. So obviously, I can't really say that these two are more similar than these two, but intuitively maybe they are. So the first thing we wanted to ask is whether, if you are searching for a TIG-Late target, do you care about whether the background has TIG-Lates or not? And when you do this, let me walk you through this in this plot. Each pixel is a specific number of combination of the number of TIG-Lates and the number of non-TIG-Lates, okay? So the column is the number of TIG-Lates, the rows number of non-TIG-Lates in the mixture, and then the color codes for the performance. Okay, yellow is good performance blue, not good performance. And what you can see here is that really the color depends on the left-right, right? So just on the column. So if you have a small number of TIG-Lates, these are mice searching for TIG-Late targets. If you have a small number of TIG-Lates in the mixture, you do well. If you have a large number of TIG-Lates in the mixture, you don't do well. And it doesn't depend much on the rows. So the number of non-TIG-Lates in the mixture doesn't matter that much. Not in a way. So I guess, depending on the, they all have some common functional groups, right? So you could find, I'll get to that in a second, but you could probably find a different group that relates somehow. These were not chosen to be related. But maybe relating to your question is you could ask whether this is specific to TIG-Lates or could you find a better group? So to do this, you need to kind of quantify the performance of this group. And the way we do it is by fitting just linear fits to each row, okay? Which tell you what the decrease in performance is when increasing the number of TIG-Lates, but holding the number of non-TIG-Lates constant, okay? So you get one fit for each row and these are all the fits. And then just take the slope, the mean slope of these fits. That's the overall TIG-Late effect in this test. And then, yeah, I didn't have it here. So what we then did is you basically split this group of 16 molecules to all possible eight and eight and ask what is the slope that we get. And we found that TIG-Lates are among the highest few percent. And if you look at the other groups, it's seven TIG-Lates and one replaced. So within this group of molecules that we chose, TIG-Lates have a very strong effect. Now, I still think this is kind of striking, but on the other hand, it's not clear how you can generalize this to some other functional group. It probably depends on the size of the functional group on which receptors relate to this function group, and we don't know. So it would be nice if we could find some other parameter that is maybe less intuitive and more easy to generalize. So the way we wanted to do this is by thinking, so it doesn't matter really whether a chemical structure is similar or not. If we can tell whether they activate receptors in a similar way, then that would be a good measure of similarity. So basically use receptor space. And the idea is that you can, so basically to do this, we imaged OMP G-Camp-3 mice, these we got from the Dulac lab. The OESA guy made them, so we can image glomerular activity shown here. And then look at the ROYs for a different glomeruli and get up for each molecule, you can get a vector that is just a list of activation by different glomeruli. Okay? And you do this for different molecules and now you have the representations of these different molecules and these we can relate in some mathematical way that does not rely on our intuition. And once you have these, you can ask, so what is the parameter that you want to define, what relationship? And obviously one of them would be similarity. You expect that if two molecules activate very similar sets of receptors, then probably one is a difficult background for the other. But that maybe doesn't have to be the case. And let me walk you through two different parameters that we looked at. So, assume that this is your target. This is the activation pattern of the receptors that you're searching for, okay? And now you're presented with a mixture that looks like this. So it's not exactly the same, but you've been trained a lot and you know that there's some noise in the system and you think this could be just one version of the target. So, if I were a mouse, I would probably lick for this, I want to get the reward. But another option you could think of is a mixture that looks more like this. This is, clearly you don't think that this is the target, but this is not what you were asked, right? The task is to figure out whether the target is included in this mixture. So potentially this mixture also includes the target. So, we call this kind of a mixture, a masking mixture for the target. Now, what we would want to do is be able to, for each trial, okay, for each mixture that we ever presented, be able to quantify how similar it is to the target and how much it's masking the target. And then ask which one of these parameters correlates with performance. And so for similarity, we just use the Pearson correlation coefficient. We're masking what we did is threshold the target response. So, we only take the mirror light that are activated by the target. And then basically take the ratio between the mixture and the target and average across the mirror light. And these ratios we bound between zero and one. So, it's basically like asking what percent of the target activity is present in the mixture activity, okay? These are all fixed concentrations, yeah. So, when we look at the performance, it seems much more, much better explained by the masking index than by similarity. And you could think of this as telling you that when there's a background, it doesn't need to be, similarity is a very strict criterion. So you don't need to be similar. It's enough that you activate the right receptors. What you do with other receptors doesn't matter, right, that's the idea. So we now basically think that we have some parameter that we can define that explains performance on a figure background segregation test, okay? So going back to this psychophysical curve, we basically think now we found a parameter that can produce such a curve. Okay, so the next question is, so obviously there's this debate of synthetic and elementary processing. And at least in some people's minds, just the fact that animals can solve the task maybe is difficult. So the question is, should we be surprised? How difficult is this task really? Or in a different way of asking the same question, if you read these receptor activation maps, how easy is it for you to tell whether the target is present or not? And this is work largely done by Alexander Mathis, who is still, I think, in the next slide. So what Alex did is first asked, what if we can just find for each target a single glomerulus that if you just listen to that glomerulus, he's specific enough to the target, then you can solve the test. And the way he did this is by looking at the distribution of the activity level of this glomerulus, so we don't, I should say, go back and say something. We only used the activation maps at this level of individual glomerulus. And we modeled what the mixture actually looks like. I'll get back to this in a second. So what we do here is take each glomerulus for a specific target. Look at the sum of the activity of this glomerulus for different trials. And ask, do the distributions of the activity levels for the trials, go trials, differ from the ones that are no go trials? Okay, can you put a threshold somewhere that will easily tell you whether the target is present or not? And obviously in this case, you'd do pretty bad. So we did this for all glomeruli target odors, pairs, and the accuracy levels that we found were far from mass performing. So it doesn't seem to be, there's always the question whether we, if we recorded all glomeruli, we could have found one. But it doesn't seem to be a good explanation for mass behavior in this case. So we then asked whether if you read combination of glomeruli, not just a single one, can you easily tell whether the target is present or not, and Alex used a linear class part to do this. And first pass, we took the linear sum of these glomerular maps as a mixture representation, and the classifier is just perfect. Not a single mistake ever. We then realized that that's kind of unfair, the linear sum doesn't make much sense, it should saturate at some point. And there's probably some noise that you should add. So we did some experiments to try and come up with a better model of what a mixture is. So these mixtures were again going back to imaging these glomerular responses, repeating each odor multiple times to get some estimate of the variance. And also looking at the mixtures and seeing whether we can come up with a model that explains the mixtures based on the single components. And I'll just do this a little bit more quickly. What we saw is that you could typically fit a kind of saturating curve. So what I'm showing here is the linear prediction. Each dot is one mixture, the whole figure is one glomerulus. Each dot is one specific mixture. The x-axis is what you predict linearly to be the response. The y-axis is the actual response. And it kind of nicely fits some saturating curve. So we plugged into the linear classifier model that takes into account the 10% variability that we measured in the saturating curve. And the linear classifier still does pretty well. And this is with a way smaller number of glomerulite than the mouse actually has, probably not as accurate measurement of glomerular activity as the mouse has, and still relatively easy to solve. So I have no claims that this is how the mouse solves it. I don't believe that's the case. But what this tells me is that the input layer is so rich that you could probably solve way more complex tests. So we shouldn't be too surprised that it's possible to solve this test. Okay, so now what actually drove me to this whole thing is an interest in feedback connections. The reason I started doing this task is because I was trying to think about what feedback may do, and you could think of attention-like mechanisms, and I thought this is the task. So the next thing to do is really to try to figure out whether feedback is involved. Haven't done this yet. What I'm just showing here is, I think something we should remember is that there's about 10 times, we don't know the exact number, but there's about 10 times more feedback projections from cortex to bulb than the other way around. So you probably do something interesting. But I'll just go through what we have done with the feedback projections. And that was to look at, basically ask what does the circuit that receives the feedback information look like? Which neurons in the olfactory bulb receive it, and what's the effect on mitral cell output? And we did this, this is with FIVOS that was in Benki's lab then. We did this looking at the AON. We injected a virus to express general adoption in the AON, and recorded in the olfactory bulb, shine light, and see core from different cell types, see which cells respond. I don't need to go over the bulb circuitry. So the first thing we looked at is where these axons actually reach. This was also described yesterday. And we saw that AON axons don't only go to the granulosa layer, but also reach the glomerular layer. And then in slices, FIVOS recorded from each of these cell types. And we'll start with the mitral cells. So if you look at inhibitory currents, then you see very strong inputs from these AON fibers not directly, but via inhibitory intern neurons. This was kind of expected. What was a little bit surprising is that if you look at excitatory currents, you also see inputs from the AON. And we did some experiments to show that these are direct AON to mitral cell synapses. Now, you can see that the scale bar here is very different. So the excitation is way, way weaker than the inhibition. And it's also a little bit prior to inhibition because it doesn't go through another intern neuron. We then recorded from inhibitory neurons both granule cells and PG cells to ask which ones contribute to the inhibitory input that we see. The short story is that basically every cell type that FIVOS could identify and record from in slice receives direct input from the AON. So they all seem to contribute. And he did some experiments by trying to locally block GABA. And it seems like a substantial amount from each of these involved in this inhibition. So we then asked, what is the effect of this input on mitral cell firing? This is just a conclusion that all cell types seem to receive direct AON. So what is the AON input due to mitral cell firing? We went in vivo to record single units. This is probably the way it's all working rats, the AON stuff. And you can see when you shine light, there's very, very strong inhibition. Basically silences mitral cells. And we asked, it took us a while to figure out why we never see the excitation until David Geyer said, maybe you're not looking at the right temporal resolution. So if you use one millisecond bin, PSDHs, then you start seeing the excitation. I'm seeing this stuff even more. And the excitation has a small effect of producing one spike, with not an extremely high probability. But it's just one spike, and then the inhibition kicks in. So this doesn't tell us exactly what the AON does and how it's involved in figure background segregation. But the one thing that it does tell us is that the input from the cortex is not a weak modulation, at least has the potential to very strongly modulate mitral selectivity, in a way much easier to see strong responses with AON activation than with odors. Okay, so I'll summarize. Yeah, this is going to be much faster. So the first thing is we found that mice can easily detect odors that are embedded in the background mixtures. If you look at the psychophysical curve, we don't reach much far on a curve. We didn't really challenge them with 16 odors. We think that we have a parameter that describes the difficulty of this task, and that's this masking, which is basically the overlap and the representation of the target odors and the background odors. You could solve this task with a simple linear readout. So we shouldn't be surprised that the task is solvable. And feedback is a very powerful or potentially very powerful modulator of olfactory bulb output. So I'll just mention what the burning questions for me are right now. And the first two that are very related is this really an attention task. Can we show that attention is involved by showing that when a mouse is searching for one target, he's going to miss the other target, something of that sort. And very related are top down signals involved. So can we show that they actually do something and figure out what? Another question that I find really interesting is, is there anywhere in the olfactory system where you really see background invariant responses? It's like the IT and visual cortex where the exact features don't matter. It's just the object matters. So we want to search for that. So the question about the masking experiment, so can you- Can I just thank people and then you'll ask? Yeah. Okay. So I- There's a second for the talk now. No, there is no second. So I just want to thank, first of all, Venki, all this work was done in his lab, all these people were involved in it. Kiosk-Bedke was involved in the classifier. And this, I think is, did you draw this, Tim? No, I was a artist, so. So. Maybe if I drew it, you might- I see. Because Andras already showed the cover, so I showed this one. Tim's version of the olfactory cocktail party problem. Yeah, thank you.