 He runs like two miles if he could get this in one story. All right, I should watch New York. Okay, so one more talk before we do the complicated stuff. It's my pleasure to introduce Andrea Schaefer, who I tried to look through everything that all the various transitions he's done, and I can't do it. I'll tell you kind of the categorization, matter of the discriminations you will. So he did his undergraduate diploma in physics in... Is it already a handbrake? Yeah. And then also stayed there to do his PhD in Dr. Schachmann's super route. And then after that went to London to try Margris at that time with UCL to do a post-op. And then stayed on there as some fancy fellow which came in an independent position. The point is it was an independent position there. Before he moved off to the marks point, he moved in Heidelberg and then became a professor at the University of Heidelberg. I think he probably left the camp back to London. I have trouble keeping him here. Someone will see something first, which then became latest Francis Crick Institute, which is where he will be for the next year or so. Thanks, thank you for the somewhat worryingly detailed introduction. It's just that I've moved back between Germany and between Heidelberg and London a few times. And now I'm happily, happily in Brexit Britain. Well, thanks a lot for this amazing conference and giving me the opportunity to talk here. In particular because of a couple of people in the audience that was very excited to have here, particularly Antonia Massimo, as you will see later on. So for the last day we've been talking a lot about chemicals, chemicals like these that evoke activity in the brain, how they are received, perceived, encoded and so on. Maybe not surprising in an olfaction meeting. But I want to now focus on a slightly different aspect of olfaction that maybe sounded through in Tim's talk yesterday. Namely, looking at this sort of turbulent, let's call it an older plume, or that's actually a smoke plume, there's just a lot of structure and a lot of dynamic information way beyond just the chemical composition. I find that particularly exciting avenue to go down to try to understand how the olfactory system deals with. And a lot of people around the world seem to agree with how exciting that kind of project might become. So what does this sort of spatial, turbulent cloud of odor, how does that reflect into the point source that maybe a nose would be in here? Well, if you stick a detector, a fast detector in here, you see these sort of rapid temporal fluctuations that happen, I guess it's safe to say across all kinds of times, scale up to several tens or hundreds of hertz. So there's a very rich temporal structure in an older plume, and the question that I'm trying to get a little bit into in the next hour or so is what they could be used for and how they might be processed. So I'm certainly not the first to look at that. So studying the dynamics of plumes, this has been done for probably for centuries, this is one of these sort of maybe earlier meteorological papers where this group looked at a large smoke plume, I think from a sort of to study forest fire smoke as well, and looked at different distances. And you already see that at 80 meters, so very different length scales than we usually deal with, it's a very different temporal structure that something approaching half a kilometer, but you still see even half a kilometer away from the source, you see dynamics in this case again, smoke plume. More towards neuroscience, a lot of groups, in particular working in marine organisms, have looked at plumes for a long time. I think some aspect may be because marine organisms allow you to study plumes in water and as Tim has already said, it's much easier to visualize something in a liquid water environment than in air where a lot of the normally fluorescent molecules just become suppressed, I think, largely with oxygen nitrogen. So this is one of the again earlier studies from these groups I've never met personally actually so far, but there's pioneers in studying turbulence in liquids and using saline traces to identify different features of such an order of plumes or measuring what typical neurophysiologists would measure, peaks and these d's and lengths and gaps and so forth. Maybe the most well-known from a neuroscience, neurocircuits perspective along the lines of turbulence and the role of turbulence certainly comes from the insect world. This is part of a famous paper by Hildebrand's group also quite a while ago, this point showing different positions along the plume you do see very different, just qualitatively very different kind of temporal structures and this is a very much more recent, I think probably the best attempt to actually have a quantitative understanding of prediction about turbulence and concentration variants at different positions from, well, Tonyo and Massimo. So still I don't want to exactly talk about this, talk about turbulence and distance and direction. I think that's a very, very fascinating, very interesting topic but I want to now kind of look at a different aspect that turbulence and order plumes might be good for and that goes back for me to a paper that has been quite influential for just my small personal experience. This is the paper that got me into affection, I was in summer school in 1998 as a student and randomly got assigned to read some paper, a different one, not this one, but someone else was presenting this paper and I was very excited by this work in P&S from 1991 from John Hopfield, where he says that animals that are dependent on affection must obtain a description of the spatial location and individual order quality, so again in the direction what Tim was saying, but also order quality, they must get this description through order of affection alone, which I think is a strong argument for nocturnal or crepuscular animals like the lab mammals we work with like mice or rats, they must get a description of their spatial environment through affection and he argues in this paper that the variable nature of the turbulent airflow actually makes such a problem more tractable and makes it computationally feasible and even argues that the olfactory bulk would be a very good implementation of the computational algorithms he proposes in this paper. So what does John Hopfield mean with this problem? Well, let's take this sort of slightly clouded picture where you see, if you imagine somewhere a mouse or something sitting in here or here on the other screen, then what are you going to be challenged with? Are you going to be challenged with a lot of different odors from all over the environment? Certainly there will be some chemicals that on their own will give you a lot of important information about a specific mate or so, but there will be a lot of information, a lot of chemicals that indicate, let's say, that this flower is closed and other chemicals that would indicate that this sort of animal is deer will be present in the scenery. And what you have to do to kind of figure out which of those chemicals belong to the flower, which of those chemicals that might be partially overlapping belong to other flowers, belong to this specific deer. If you want to be able to allocate objects in the scenery just using your sense of smell. Maybe a more natural environment for a mouse would be this Victoria tube stop where a mouse tends to live in here and wants to maybe understand whether the cheese sandwich here is actually a cheese sandwich although you might get a lot of cheesy smells from all over those people as well. So that's the kind of challenge, I think is one of the key ones that so far has not been addressed very much in effect. And how do you separate your environment? How do you segment your scene, figure out which chemicals belong to a specific odorsource, the cheese sandwich, or the shoes, or the armpits, or whatever you find in Victoria tube stop? So the idea that Hopfield put forward as well, there is rich temporal structure in those natural plumes in normal turbulent environments and one of the key features would be, so he argued that chemicals that come from the same source, let's say the cheese sandwich here, they would all fluctuate a lot, but they would fluctuate in a very highly correlated manner. Because again, as we sort of discussed a little bit earlier today, diffusion over these scales generally plays a much less prominent role than sort of turbulent airflow in terms of carrying chemicals. So whether these different chemicals have different diffusional constants doesn't matter so much as the fact that they might be carried by the same kind of little air turbulences. Whereas odors that belong to a different source, like these again are highly dynamic, but they will be correlated within but not correlated so much maybe with the other group of odors. So that's the idea that would allow an animal to segment a scene to figure out which chemicals belong to one, which chemicals that again might be overlapping belong to a different source, just sort of figuring out whether the green and the square and the red and the triangle belong together, so whether it's a green square and the red triangle or a green triangle and the red square and so on. So this is sort of the general idea of scene segmentation or an auditory system as all of you know very well, it's called the auditory cocktail party effect, the idea being that if you communicate with one person you need to understand there all the components that make up the sounds they make as being separate sources and all the background activity. So this scene segmentation or cocktail party quality effect or binding or whatever you want to call it I think it's very prominent in all faction and some people it was mentioned that specific even a specific combination of smells a very different epilogical meaning even for insects than the components in itself it would be a very good example if I could remember who exactly said that and what it was. So I couldn't resist putting up this beautiful picture that's the cover of Nature Neuroscience from paper from Dan, Dan and Venki where they tackle a very, very similar problem sort of figure ground separation and show that individual odors can be identified in the context of varying of very highly different background odors. So they look essentially at the static way of separate identifying individual odors in the context of a large array of different background odors and I think Dan might be talking about that to some extent. So in tomorrow I think. So to visualize a little bit to give you a bit more intuition again this is still kind of a very lengthy introduction I try to not do much introduction to visualize what the problem I'm trying to address here these are just two little blobs of dry ice and image from top and simply using a color scale and you see it's a very dynamic airflow that's here tracked by the sort of dry ice plume that you can imagine that if you're here you see these bouts of odor or dry ice coming from this source you see them coming from this source and they seem to be sufficiently complex that you might be able to attribute different chemicals from one source and distinguish them from the chemicals coming from a different source. So how can we address this problem properly? Well what we want to do is we want to compare a mixture of odorants and compare two odor sources that contain exactly the same chemicals whether there is a difference in correlation if you measure it let's say half a meter or so away to stay in the length scale that's relevant for mice the way we do that is using photoionization detectors 95% of you probably use them themselves or know them for those 5% that don't the idea is a very simple one you suck in and these are commercial devices you suck in odorized air into a structure where you apply so if you relight with a specific wavelength that's going to ionize everything it's more ionizable than this specific specific energy then you apply a high voltage to separate those ions and then simply measure the current sort of sensitive amplifier measure the current that is proportional for large concentration ranges to the specific concentration these devices can be very fast and detect everything below a specific ionization energy and can be very fast with a bandwidth of several hundreds of hertz so what we have to do now if you want to measure these mixtures or individual components we have to actually have PIDs for two different kind of odors to simultaneously measure these temporal correlations problem there is that commercial PIDs don't tend to not have that but if we could get to several different lamp energies we could have one that only ionizes one specific component of one specific odorant so it would be sensitive only to this bluish one and others that would be sensitive to a large group of different odorants and then taking both measurements together we could piece a piece apart which odor is present so what we actually took is this device that many of you are using Aurora Mini PID with a bit of help from Aurora we managed to find other UV lamps that actually fit in there unfortunately most of them seem to be not produced anymore then you need to modify the PID a little bit so that it's resonant circuitry to allow it to operate with this different lamp and then you need to find odors that have sufficiently different ionization energy that are not too toxic for your postdocs and PhD students so the two odorants we picked is Ethobutray there are many many others that have a relatively high ionization energy of close to 10 electron volts but there are not so many volatile chemicals that are not toxic that have a relatively low one of a taping to opinion is one we are using so these are now two model odors we can use this one being being sensed by both PIDs with two different lamps and this one only being sensed by the higher energy one so very simple take those two odorants if they are mixed in some sort of dynamic way we have one high energy one that detects all odorants the low energy one that detects only one of the odorants and then essentially we can take the difference of the two to get to get to one one last sentence they have slightly different filter properties there's some scaling factors but in essence so the ones we looked at the temple bandwidth would not be high enough to I mean we think you know hundreds hundreds of hertz probably the olfactory system is not faster than that so we want to see hundreds of hertz but the ones I saw were kind of in the more hertz or few hertz which would been too slow obviously if you go to water you could use fluorescence and look at fluorescent dye but for air it seems if anyone has a better idea I would be enthusiastic to try something else so what's the outcome of these experiments again as I said we have two sources approximately 50 centimeters apart in a turbulent environment we measure odor one which are at the butrate and then after processing a little bit odor the both are a little bit processing that's the concentration of odor one that's the concentration of odor two two different sources both look kind of dynamic but they don't seem to be very highly correlated we go to the mixture however again same distance we see highly correlated temporal fluctuations this sort of example you clearly if things are correlated you know it was the mixture if things are less well correlated you know two independent sources and that is very robust it's truly if you do that tens of times you see always highly correlated mixture always barely correlated or uncorrelated to individual components so in principle the information seems to be there it's actually relatively gradual if you go tens of centimeters apart it's somewhere weakly correlated this is all done indoors we don't have contaminant odorants and we create turbulence by fans with little cylinders in front we're not yet at a stage where we could actually calculate the turbulence to see whether we understand everything but if we go outdoors on to my my slightly oversized and windy and rainy balcony with all kinds of background odorants we do see the same strong significant high correlation despite all the background going on for a mixture and weak to no correlation for 450 centimeters apart so this is without doing anything except placing vials of odorants onto our balcony and measure with the two PIDs 50 centimeters apart so there's flowers, there's hopefully not rats but all kinds of other smells around that do reduce this correlation as well as increase this correlation so it's a quite a robust phenomenon it's so robust that if you go to different distances so much closer where you see stronger temple variants or further away where you see weaker ones you still see this massive difference in correlation between mixture and two components and obviously we tried another odor pair we could find and again it's the same kind of behavior so the PID is between 20 and 60 centimeters and the two odors are between mixed and 50 centimeters so we tried to be in a length scale that we then could do for example navigation or experiments in the lab and it seemed not too small not massively too small in a length scale for our mice so there's a lot of temple structure and this temple structure there's a lot of information in there about things like distance and direction and so on but in principle those two should be the same because it's a physical mixture the individual odorant is still present exactly the same distance we have started looking at quantifying these things but we are way not as qualified as some other people in the room to do it properly it's certainly not a trivial problem what kind of the measures are that you would have at a specific distance but I think very interesting one we see so these sort of bulk correlations are done with many repeats even independent days so the correlations say the same the structure itself varies it's different again it's essentially a similar answer to Matt we might not understand the temple structure but what we do understand is that they are weakly correlated and highly correlated in those two cases so we use for simple technical reasons of finding odorants with a low enough ionization energy we did these two odor pairs where we find the same correlated versus uncorrelated I think they are a sort of different temple structure you would start getting a situation where diffusion becomes a relevant transport mechanism so I think we discussed that paper even already kind of a few months ago so what we're doing we've been trying those four odorants very simply have a little robot opening lids because we don't want to push air out very similar to what Tim was saying we don't want to push air somewhere we just open up and then wait for 10 seconds or initially we just did it manually we haven't systematically gone through many odorants and just the temple structure and not several odorants then it would be very feasible to do but that's essentially I would think what that paper did I don't think there's a good quantitative theory behind it but just because it's very difficult to do but maybe we can discuss that this is afterwards it's well possible because what we're doing is we're actually lifting two lids we try to not lift them too fast to not create but we will have correlation through the onset so the measurements we actually do for this correlation coefficient are only for the period of odor presentations obviously if we correlate the entire trace there would be a strong correlation simply because of the onset effect I think that's more a technical flaw in the way we can present and can get lots of odorants and conditions because the more natural situation as flowers and animals sit around and maybe there's an additional additional effect from an animal moving so that everything gets emitted at the same time so generally I think from these experiments that are somewhat intuitive from what one knows about the turbulent transport what we would conclude is that the dynamics of odors indeed following hop fields that allow for sort of seen segmentation or object identification or if you want to use it maybe not any more overused term for binding of different components so the question we now want to of course ask is can mammals use this kind of information so in particular if you sort of look at some of the spectrum, I should have put a spectrum here I guess, there's information up into the tens of hertz or even hundreds of hertz where you find very tight correlation so it might be a stretch to figure out whether a mouse that usually I don't know sniffs at 5 to 12 hertz or so maybe that mice would be and with slow receptor neurons that they would be using this kind or be able to use this kind of information. There's several reasons but I don't think it's too far a stretch to try one obviously comes from insects where a lot of work in particular by Paul somewhere in the room up here by Paul showing that bees are very good in detecting instantaneous or correlated onset or sort of simultaneous or slightly shifted onset down to time scales of several milliseconds so there is definitely a factory system out there that would be able to at least detect onset maybe not correlations but onsets in activity and again this old work from maybe we can try to discuss that in the end so a simple answer to the question that Dima will ask is we haven't yet measured sniffing during the tasks that we'll show but we are about to do that so we don't know exactly and we don't know mechanistically very well. What are the patient on the previous slide? I think it's volatility of the two odorants. Is that how you deal with the BID? Yeah, that's a problem. That is all stuff that is actually in the error bars of our measurements here. The reason that it's not one correlation or in particular is the fact that we do have noise in particular the probably maybe a couple of people are interested in this. Technically these beautiful mini-PIDs if you tune them for a lamp that is outside the normal range becomes somewhat unstable and actually you do get noise and then finding chemicals that have a substance that are below that lower ionization threshold they tend to be relatively close to the ionization threshold because it's very inefficient to actually so the calibration I should have said and actually the debundant has done that warned me again today this is a very well calibrated measurement this is much less easily calibrateable because of exactly those noise and PID issues. But for the sake of the correlations because we're comparing two conditions that are the same problem I don't think it matters too much. So in the insect world definitely there's indication that also sort of projection neurons follow fluctuations of auto-concentrations and work by many labs including I don't know Rachel Wilson's lab and Marcus Meissinger-Rohan I think about collaboration in this paper. In crustaceans there's maybe indirect evidence that crabs navigating seem to make use of turbulence or it is difficult to explain without that. So there's maybe indirect evidence that they can perceive those kind of turbulences. In mammals I think evidence is maybe a different issue and this I think the only slide that's published data from my lab, from mine and Dimas and Zack and many other labs work. We know that at least mammalian affection is not incredibly slow so we can do simple detection and discrimination tasks can happen much slower than people thought maybe 20 years ago and most importantly in this case from Dimas work we know that if you give very precise inputs optogenetically mice are very good in discriminating on the order of tens of milliseconds or even 10-20 milliseconds so there is the potential of the system to deal with high-temporal temporary structured stimuli and the system itself can actually respond in a quite precise manner. So what do we have to do if you want to figure out whether mice are really able to use those turbulent information? Well the first thing we have to do is to be able to reproduce temporal dynamics in a reproducible manner. So we have to get away from what we're trying to do to create sort of square order pulses to have it defined onto an offset we need to go somewhere that we can reproduce those kind of fluctuating order stimuli and that is a little bit of finding and engineering effort we ended up using these micro-dispens valves that can operate it up to 1200 hertz with very small volumes and if you use them and some manifolds do some sort of electronic control of them to create the right kind of pulses you actually have quite a small dead volume and can create pulses that should be at least in the sort of tens of hertz and if we are using again a PID to characterize that, these are pulses with increasing frequency we're giving and ultimately the one showing here is sort of 50 hertz stimuli and we still see a very strong modulation on a 50 hertz on a 50 hertz time scale. So, importantly this means that we can now take a recorded order so again in principle replicate those kind of signal structures from day to day and no we don't have to rely on a good generative model of what's happening what can actually give precise and sort of simplified, highly simplified simplified order pulse strains. Importantly with no change in overall flow because we have several valves so we can compensate flow quite well. So if the first question we need to ask then can mice deal with this kind of temporally structured information so for that we need to train them and it's expected that this training might be a little bit of an effort because as I think Leslie put very well you can't ask and talk to flies but unfortunately it's despite expressing Fox P2 you might not be able to talk to mice either so it's just you can't tell them pay attention to the order source you need to train them in a lengthy conditioning process so what we've been working with for a while as we built this sort of cage where we have a group house dozens or 20 mice or so each chip tag they get free food but have to walk and work for getting water so which allows us to have sort of train them on simple go-no-go tasks for quite lengthy period of time this is sort of the system as evolved over the years thanks to our workshop so there's a bunch of mice sitting in here in quite a happy environment that the Home Office, the regulatory authority in Britain likes a lot because they're actually they're not really water deprived they can go to water anytime they want they have free access to food, social environment and importantly very little interaction with any human so if they want water they just have to walk up these little stairs go into a little separate tunnel where they do a where they essentially it's a little operant conditioning box attached to it and that allows us to measure many many trials over periods of weeks or months importantly they perform a lot during the night and not so much during the day as you would as you would expect and with each animal the way we've set it up does approximately 300-400 trials per 24-hour period up to periods of periods of tens of days so this general system allows for quite efficient and automatic conditioning or sort of cohorts of mice with some advantages that it's minimal human minimal human interference that it happens during their naturally active night and we can do some automatic health control and finally maybe an underestimated point is we don't have to water restrict the animals they can obtain water no matter when they want so they're never dehydrated unlike if you sort of don't you take them off water for 24 hours and then have a brief training episode so this set of advantages now allows us to maybe try to train them on correlated and uncorrelated stimuli so to determine this we train them on a go-no-go conditioning task but not instead of discriminating sort of simple odorants we actually ask them to discriminate the same odors but with different temporal profiles so these are going to be the two extreme stimuli we're starting off with a rewarded stimulus where two odors are highly correlated and an unrewarded where they're actually anti-correlated so the most extreme we can try to do this is an experiment that as many I think of factory behavioral experiment comes with a lot of controls necessary and caveats I will walk you through some of them the first is obviously to calibrate the system such that there's no indication in total flow and no indication in total odor concentration averaged over one of these cycles that is different between the S plus and the S minus trial we actually achieve that partially through carefully calibrating those different valves and I think Paul can probably sing a long song about comparing individual valves in their performance we have to go through similar routes here but also then at some point we add a little bit of noise to make sure that there is additional variability that makes it impossible to distinguish between those two stimuli based on an unintentional parameter importantly also we use several valves to make up the blue odor several valves to make up the red odor and several valves to make up the compensatory air flow this allows us to from with each trial to actually use a different combination of valves so it's very difficult for the animal to pay attention to specific kind of click noise and we can train them for extended periods of time and importantly if we start off with training animals on six out of those eight valves and then we at some point introduce these two new valves that they've never seen before and from everything of the performance of all these other valves there's no way to predict whether there's an S plus or a S minus stimulus coming up so we can use those trials as being a completely novel set of valves, directions or whatever the intricacies of our little manifold manifold might be. I'm going into the length of these control details because we had so many nice control discussions with Tim this morning I have to show them slightly unprimed and then after a few days we can just manually move around and refill fresh bottles and look how that's happening and then switch to these switch controls switch controls again so again we first we first train them for about a week with I don't know a signal versus eugenol or something something unrelated and then we use two new orders for the temple structure so I'm more controlled across frequencies maybe drop that so we can start off also with a relatively simple task we do two hertz correlated versus two hertz anti-correlated then the animals tend to learn that these red group of animals is one where we actually well where valves are clicking in the same way but the odors don't have any prediction any predictability and then we do introduce the switch and here's an example of the first switch we're doing so each row is one animal's performance and you see in the different green colors things they got right in the different red colors things they got wrong and so it said these hits or correct rejections for the lamp misses as people call them and then we have this time that we switch where the first stimulus there's no way for animals to be able to know whether that should have been an S plus or an S minus and you see there's some animals that are perfect before that stay perfect afterwards like these these cases here others make some mistakes but overall there's no difference in the performance before or after so for me this is sort of the tightest kind of control we can do in this behavior because there's no difference in the performance and make this distinction except using using what we intend to put in because it's a completely new system so these are two hertz stimuli so these are 250 milliseconds on 200 milliseconds off so these sorry so this length here would be 500 milliseconds and then we're ramping up the frequency to see where up to what frequency actually detect stimuli two hertz is maybe not so exciting but this is sort of our star animal one example to show we started training on 2, 3 et cetera going up a frequency up to 12 hertz and we do a sort of switch control at 12 hertz go up to 16, 20, 25, 30 and even at 30 to 50 hertz the animals perform at 90, 95% correct so this is 30 hertz fluctuating stimuli every 30 milliseconds there's no difference between those two stimuli and so this animal must have been able to somehow pick up the correlation structure of the stimuli on a 30 millisecond time scale and the population of animals these are 14 mice and a few hundred thousand trials so we see that performance goes down with frequency plotted here logarithmically from 1 hertz, 10 hertz to 100 hertz up to on the order of tens of hertz we see it seemingly different from chance line trained from doing sort of statistics and comparisons here because we can discuss with people who want to do that in a minute I can in a minute so because we were suspicious this is I think a relatively well relatively extraordinary claim that mice are able to distinguish correlation patterns on the 20 hertz time scale so we repeated the entire set of experiments in the new cohort slightly different training starts with extensive training on low frequencies but the second cohort of 19 mice again a few hundred thousand trials performed same way we have less variance here because we simply have more trials in these frequency bins in the tens of hertz range and then we at that point we also had my very suspicious PhD student wanted to have a control of a group of mice that are not receiving odorants but the same both clicking patterns and those animals didn't perform anywhere near above chance for the entire frequency perspective so there is a kind of realm where animals might be able to discriminate so they are able to discriminate correlations up to tens of hertz so it's I think similar to what we've done for a long time we have bins and they need to licking I think two out of four bins but I'm not 100% sure I should know we simply haven't done that but we thought we ramp up frequency and that would be the easiest way to make this relatively unintuitive task for them maybe work but we haven't done explicit jumps so now then we randomly mix frequency between one and 80 hertz together to get this sort of large data set that maybe we should do I mean I think we try to do that with all the kind of controls we do in terms of making sure you switch to something new you switch to different valves and using this cohort of animals clicking pattern and same directional pattern I was thinking that's with the pharmacological manipulation that Tim was mentioning but that might be I think you would like them to kind of have learned something and start and then kind of train the probe but maybe I'll take the point start with Dan so the difference is whether the two orders are correlated or whether they are anti-correlated and the anti-correlated case half the trials one order starts the other half the other order starts maybe the next slide answers your implicit question that's what we're doing exactly at the moment so it's maybe could lie feet into it so so this is a little bit this experiment what we've done so I don't want to go too much into a mechanism because that would open up an entire new avenue and kind of convince you of the phenomenon first what one of the mechanistic ideas we had maybe and that goes into Dan's direction maybe they detect onset like what Paul has shown very well in bees so what we're doing here is we are training animals so this is a cohort of animals that does very well at sort of frequencies of 20, 30 hertz or above so we took the frequency where they just were able to perform this range and then we keep the stimuli intact except that we alter the onset so the onset might be instant although the pulse is anti-correlated the onset will be correlated all the pulse is correlated but the onset will be anti-correlated so the prediction would be that if they detect onset rather than the sort of correlation structure performance should actually flip so they should mistake the S plus for an S minus and the other way around if the onset is just part of the overall perception it should maybe drop slightly and what we saw is exactly that, that it never flipped but it just dropped slightly so we think they're not detecting the very onset of it but they actually take a little bit more time to integrate and actually use correlation across the entire structure Dima is not convinced maybe so we are we're in a physics institute so I can use the blackboard so what we're doing is we are so this is let's say the correlated stimulus so this is the S plus and the S minus looks like looks like like this and then we are we are introducing probe trials where we are sort of changing this but still the rest of the trial trace stays the same so the onset is now not correlated but anti-correlated whereas we're introducing in these trials we are making the onset the onset simultaneous relative onset between the two odors I thought I kind of managed to manage Dima's SNF question that's why we're measuring SNFing or we're trying to measure SNF at the moment but maybe in the interest of time so we started looking at reaction times as well what we see is that what we're plotting here is performance so perfect performance, chance performance as a function of their reaction time defined by the first time of licking for the S plus trial and you see that those animals that perform well are the ones that respond later and this curve becomes shallower but still for high frequencies those animals that perform around 70-75% at 60 hertz are those that actually refrain from licking for a long period of time so this is contaminated a little bit by the task structure that licking also indicates when their action but it is some indication that there needs to be some more lengthy integration over at least on the 100th millisecond time scale no I think this suggests, I think what Zak and Dima and we all saw was that discrimination of two simple odorants even if you mix them together can happen very or does happen very very quickly within one sniff within on the order of 200-300 milliseconds here now we're asking the animal to make a much more possibly much more difficult but maybe also a logically more relevant distinction namely whether two odors are correlated or not and that requires more time no this requires a long time, this is a second so the scale here is one second so the animals that perform at a way above chance need, have a reaction time of approximately approximately 900 milliseconds to a second so significantly longer than what Zak, Dima and we had frustrated so this is so that's maybe then come to my last slide as I said I don't want to go into a mechanistic interpretation but hopefully and if I haven't convinced you that the animals are able to detect temporal structure on the tens of hertz time scale then I failed, I want to give you some notion that we think they are able to do that comes from physiology because as Linda and others have pointed out one should maybe do even more controls to the behavior so one way to confirm that the animals might be able to discriminate those structured stimuli to actually look in the brain whether we can see cells that differentially represent those stimuli so we've started doing a little bit of imaging of the output of the olfactory bulb so as Charlie said I should have a picture of Carl so I think this is the olfactory bulb and these are the cells we are recording from so mitre cells and we find some examples approximately 5% of cells that show significant differences between those two stimuli the correlated and the anti-correlated one but this is very early stages so this is a very preliminary based on a few hundred cells imaged we seem to see most the highest percentage of cells at around 20 hertz I don't know whether that's real we haven't understood obviously what's happening there we haven't tried to directly record from those cells we're now doing a little bit of a little bit of electrophysiology extra cell unit recordings for throughput reasons and there we also do find some cells that respond significantly different between the two types of stimuli so the only thing I want to take away from the physiology is that not only can we do behavior with a lot of controls that animals are capable of discriminating temple structure on the sort of 20-30 hertz level but we also do find some representation in the olfactory valve sort of consistent with that so let me try to try to wrap up so like many others we find that naturally spreading odors have a very rich dynamics we find that chemicals that come inspired by hot fields were chemicals that come from the same source co-flux rate and fluctuate in a highly correlated manner and they're weakly or uncorrelated even for closely sort of 10 centimeter closely spaced sources if we build sort of a fast auto delivery device using these microdispens solenoids that allow us to deliver odors with a bandwidth of 50 hertz or higher and we can then use that to the behavior and show that mice can detect the correlation structure up to frequencies of well beyond tens of hertz and you know confirming that yes the correlation structure is already somewhat represented in the output of the olfactory valve so stepping back my conclusion would be that mammalian olfaction is not only fast but actually also high bandwidth sense that is able to detect and discriminate temporal structure in the environment so by the way I'm not trying to say that the animal knows these odors are correlated or anti-correlated I think it simply learns that they are different we've done some extremely under underable human experiments my PhD student looking at trying to learn to discriminate those it would be fascinating if someone who knows a little bit more about that would actually figure out what the perception of humans is from for those highly correlated and anti-correlated similar whether people report anything that has anything to do with the chemicals or whether it's simply two different smells and they can't even tell whether they are correlated or not so I'm simply saying that the temple structure is something that the animal is implicitly able to able to process and I think that might be sort of the orthogonal dimension to the entire chemical space and allow us to understand how the local circuitry of the factory system extracts this kind of information. Finally I want to as Venky mentioned we've just moved into this beautiful new building Francis Crickin Studio in central London it's an amazing place great great colleagues an amazing core funded institute and we are now hiring every autumn hiring group leaders early careers so directly from postdoc this is core funded for up to 12 years for a group of six people don't have to write grants can use core facilities no mouse costs and the great no teaching is an incredible environment so let me if anyone wants to know more let me know and finally the people that have done all that really was spearheaded by an amazing PhD student Andrew Erskine is just about to finish PhD in the next six months or so and was joined by two experienced postdocs Tobias and Tobanjan and yeah they are the true driving force behind them obviously thanks to the funding sources for that and thanks to all of you for all the questions