 Talk to everyone. So I want to talk about the work of I have to stand over here to see them all I want to talk today about the work of David Greenberg, particularly Jürgen Sevinski maybe at the end of a good time Steffi and Damien Wallace and half of Giuseppe Nattoro So what we're primarily interested in our lab is how animals use their vision to make decisions when they're freely moving In order to get to that point Has been a very long road, which I'm going to drag you along a little bit on the way But I think one pertinent question we really want to understand I think as neuroscientists at this point in 2012 is What do we want? What is actually that we need to know? How do we know when we know it? And I think really a lot of this is going to come from of course building very very nice models Building theories testable hypotheses which generate experiments which can actually test these hypotheses now This might not be good news to a lot of computational neuroscientists, but to experimentalists. This is this is somehow still not within their range They're still dedicated to technology in a lot of ways Our primary focus has been on the rat and the rat is a great model system because they're very clever I think they're a lot clever than some of the people in my lab sometimes Because we've actually had rats that have got out of the cage and somehow the postdoc has always ended in the cage But they're very well equipped. They're very well equipped on their head here They're very well equipped to survive in extreme different circumstances. They're generalists So really to understand how these guys operate you really have to understand the behavior in a lot of respects and they've sort of equipped with all this equipment on their head to make sense of the environment and They've got the eyes the ears and the whiskers and the snout and you've seen some beautiful work button Cal break before Which is detailed there. So what do we want to know? What would be nice is if we knew something about the difference between two cells Which are located in the middle of a column versus two cells Which are the same distance part at the edge of the column versus two cells which are separated by some distance in space If we could somehow make a prediction of the firing rates the response rates given what we know about their anatomy We'd be in good shape. We've been very good shape. So traditionally speaking as you've seen earlier today Exa-cellular recording and electrical recordings have been the mainstay and given us the vast majority The absolute vast majority of our knowledge about the brain has come from exa-cellular recordings But also more recently in the last 40 years 30 years from intercellular according to sharps But also with now wholesale recording and they all have their advantages and different disadvantages But I don't want to talk about that One of the main advantages of all these techniques is that you can actually then have your macro space So you can find out where in the brain you are operating from Particular area for this example here. I've used something from a review of Mikhail Brecht and Dan Feldman And what you see here is the barrels so you can tell which whisker here innovates this area roughly But the nice thing about it is you can get the anatomy you can get the input What the cell receives the dendrites or little knobs synapses, but you also get the output the wiring if you like and The output of course is very important because this is what really communicates of other cells Now this might seem like it's basic knowledge, but it's be surprising So if you think about it if action potentials are the currency of the brain These are the things which they communicate in a fast-time course, and of course plasticity must be the the bank Is a joke Okay, now the problem with exa-cellular will also recording electricity recording is that really you've got to do the same thing Many many many times give the same input to the cell many many many times to understand on average What it's interested in and of course we don't operate like that. We operate in real time We don't do many things exactly the same way animals certainly don't do the same thing every single time and I can show you evidence for that direct data So really this is one of the problems about recording one set of the time is a is that cells don't operate the brain doesn't operate one Set of the time as you just saw before Christoph's talk But ready to understand what the cell likes about the outside environment You got to do the same thing over and over and over again. You get an average and say on average the cell likes this Now, of course one way around this of course is to put space back into it as you can do you can use these recordings from Border sensitive dies is the work of Damian Wallace out of my lab when he's with Zuckman and now What a sense of dies brilliant because it reports sodium And so there's very fast you get very fast signals It's very fast time course gives you vast areas vast tracks of the brain the problem is you don't have any cellular resolution Now why do you want cellular resolution because these are most likely the computational units of the brain These are the guys that generate the actual potentials which is the currency between the neurons So you really want to better get single cell Single cell fidelity so step in two-photon imaging now two-photon imaging of course has a lot of baggage like most techniques When they first start and I want to talk now for the next sort of 20 minutes or so about Two-photon imaging and the absolute brilliance of it But also the absolute horror horror nightmare in hell that you can face of it and the reasons why and what we're doing To solve this problem what you saw there, which I'll just play it again is a a Stack just a stack through the cortex some out of sensory cortex from where layer one down to deep layer three And you can see here's the cells labeled in green Astrocytes and yellow and the blood vessels moving around now the key point here is that you can see all the cells in this local volume every single cell There are no unlabeled cells in this volume with this calcium diet But you also need to know as well that there's a lot of other cells except for neurons in the brain And there's lots of different sub types of neurons in the brain So you need to better sort those out somehow and with genetics you can do that very nicely And of course you can generate Very nice side views by the one that shot past there It is a side view we can get space now about now you want time So how do you get time out of this because you've got this beautiful view of all the neurons now you want time So you use a calcium sensitive diet. These are the best ones There's the mainstay of imaging in the in vivo and these are the ones which give us The best hope at this stage although it would be nice to get some vulva sensitive diets in vivo that work so The nice thing about the calcium diet is that it's very robust it reports out action tendrils if you recall from the soma almost exclusively Because of the the sub threshold depolarizations that generate the calcium a tube that's too much in the noise to better pick out But the other nice thing about it is that you can build genetically encoded calcium indicators Which is what this guy Lauren Luger did here. This is some work We did with him to try and extend the depth limit of two-photon imaging And what we saw was that if you label these cells in somatosensory cortex and layer 5 a and b This is just an overblown shot the show the axons Then you can actually record for these neurons. Here's a close-up of the post-mortem tissue You see layer 1 there layer 2 there 3 layer 4 layer 5 a and 5 b in the end of the brain into the cortex and Here's a nice image We talk a nice stack showing that we can now image down to a millimeter below the cortical surface by using these extremely bright G camp 3. This is a this is a gfp based Molecule Which is sensitive the calcium changes and calcium you see all the dendrites here And you can just follow your favorite dendrite and just follow it down to the soma And what you can see here is that you can follow down now neurons Plus their branch points all the way down to somas and layer 5 a and also layer 5 b The reason why we couldn't get layer 6 was because we couldn't label it So we're working on that moment label layer 6 is the layer 5 a cells And you'll see it breaks a little bit in the layer 5 b cells of much larger somas, which is expected It goes down to a millimeter so Really this has got a good muscle of a hope that there's actually hope that we can also get not only The calcium transients out of the neurons and upper layers But also all the whole cortical column which of course would be a real dream if we actually label and record from every Single neuron in the cortical column in some in some reasonable fashion then then we can really say a lot about the cortex And so here's here's what the study basically did we recorded from somatosensory cortex layer 5 and layer 5 b neurons And we moved the whisker and we could show that that we could get calcium transients out And we got roughly the same firing rates as you'd expect from exocellular and intercellular recording done by Krista Koch and others Okay, let's go through some problems Calcium imaging is difficult because it's a very general signal calcium calcium goes up and down for Multitudes of reasons and it's not just spiking it wasn't put there just for our spiking. It's a very general signal It's extremely insensitive way of determining fast sodium spikes sodium spikes are very fast to happen in the order of Millisecond or so calcium is very slow sort of runs over the course of 100 milliseconds or more It's not so it's very slow. So what are we going to do to solve these problems? to give you an overview of how this how this sort of calcium signaling works is that if you have one action potential here And you get the result in calcium trans this is all in vivo work to none of this is in vitro Just want to make that very clear which makes it doubly worse because of the noise which I'll explain later We have two action potentials here These are the differences what you know here what you see here very nicely that they have this exponential decay See this very fast rise exponential decay which goes on back into the noise, but this is where the problem lies The nice thing is it's slow Which means you don't have to be an exacting to the right spot at the right time when it fires But the slow that so you can come back to it and you have a dwell time and come back to a cycle time of say 100 milliseconds you can still pick it up before it heads back into the noise because it's decays here But the problem is if you don't scan fast enough if you don't actually get around and scan enough over these cells Fast enough you're not going to be able to detect singles and double action potentials it heads off into the noise And so here's an example why so here's your frame at at the top Here's a idealistic so so matter calcium signal and here's if your frames are Point one second frames you can see you can pick it up main most of it But if you had say one second frames you would have just sort of the square the sort of rectangular shape here And here's your here's your spike sitting here But the most important thing is you said well, okay, but I can pick that up Sure, I could just threshold but no because remember. There's a lot of noise. There's a lot of noise on the either side This is ideal noise is sort of wiggling around like this around the thing And so the problem is if you take too long it actually just decays into the noise So if you sample resolution here is your single and double this is the normalized transients So these are two normalized together and as you move through time is hidden to the noise This is the noise of a single. This is a noise of double. So you've got to go fast enough to be able to pick these out Now what do you want singles and doubles what why do you care because maybe you just want burst Maybe you just care about these big explosions in the brain But as you saw in Christos talk just before very it's very important point was that this patient was sitting there watching the screen and in between these samples of Images it was almost silent. You didn't hear me spice. It wasn't blasting along at 20 Hertz What we used to think it was it was very quiet very precise. So the chances are I don't know that the code is very precise Very precise. We don't know this. So what do you want singles and doubles? The cells do different things. Here's some beautiful work by to stick some parame We're done and so the recording of sharks and extra recording and this thing's blasting along at quite a high rate These cells you can see if the two seconds scan two second window there on the other hand This is the work from a cow Brex lab This is a freely moving animal and they're recording from a neuron in a freely moving animal in the motor cortex I think I remember and you see here This is the velocity of the animal and speed look at this. These are single spikes Yeah, this is 30 seconds And we've had a very similar experience as well that when we've done recordings exercise it both in the wake both both in Anesthetize we have extremely low firing rates, especially in the upper parts of the court especially in layer 2 3 the firing rates are very low Very sparse but very precise So that's why we want to get singles Because obviously the models you build are going to depend on your rates They're going to they're going to produce two different models of how things work Okay, back to calcium So how do we do that? How do we get how do we get all the spikes? How do we pull these guys out? So here's an sort of a normal recording. You'd see a whole lot of gray values here This is gray values here. This is time this way grayscale values Here's some cells there you can see here that they're all different brightnesses So you get different levels of calcium the nice thing about calcium is That in the ranges of the cell the calcium dyes that we use in the ranges that the cells Use the castle the concentration calcium the cell is reasonably linear. It's not completely linear It's not the saturation the dyes have been sort of made to fit around this problem So you can sort of approximate quite nicely Delta F over F. You've probably seen all the same as Delta F over F. It's based on this formula here We subtract the background and do divisions and subtract things But here's the problem And you can end up with these nice flat baselines here with the calcium transit sitting there and these peaks are all relative to each other The problem is that in vivo. This is all very nice in vitro, but in vivo all hell breaks loose The things drift you've got things moving up in slow wave forms fast wave forms You've got all types of signals at different periodicities in your signal And this is this is like baseline drift This is not from bleaching this happens all the time They sort of move up and down you have the spikes sitting here on top now trying to estimate that is Quite difficult. It's more difficult than you think because the noise isn't sort of no noise per se You can't model the noise. It's a combination of many different things It's part biological, which I don't have time to go into because it depends on the on the properties of your microscope The properties of the wavelength you use it depends on the depth depends on your actual resolution of your microscope basically But trying to actually estimate this is quite difficult because you can imagine this if you had lots of activity here Your F zero will rise if you just did some sort of moving average It would rise up when she wouldn't get the true of zero Why is this important because you're divining by it? Which means it's going to change the outcome of your of your peaks and where the peaks are So our first attempt at doing this is we took the knowledge that this is back in 2005 when I was in Fritchoff-Helmholtz's lab We took a just a basically a linear match filter and what that allowed, of course is us just to pass Over here look at the correlation between these two and then we just run a threshold through it And it was quite successful finding spikes now this we thought this was great The problem was it fell apart when we did the next paper because we did this in the motor cortex And the motor cortex and the necrotized animal you could say is reasonably sparse. It's not doing anything So the spiking was incredibly low, which means all we had to do to get rid of the noise is extend this This decay out and it just killed out all the noise, which means you could just search for these points It was easy and we're high-fiving in the lab thought it was awesome And then we did the next study and found it was useless So we first generated this paper here, which we showed you can detect successfully very easily single-action potentials And pretty much all of the cells that you record from this is from cell attached recording So we have ground truth which I'll talk about in a minute So the nice thing about calcium imaging you can put the electrode in sit it up to the outside of the cell generate your ground truth Which means you know when the cells spiked put that away in a mystery envelope If you like build your algorithm run it over and test it to see whether it actually is produced in the right result This is how this first papers in 2005 and it worked great Until we did something in the barrel cortex will be sort of moving to whisk and adding more activity And we couldn't detect the spikes using this method So One thing that we realized after we've been sort of using these this is the work of David Greenbeard By the way if you've got any questions about the intricacies of the math and this he's a mathematician in my lab Feel free to email him his email addresses at the end What David realized was that there's a lot more information than just the shape of the of the transient And what he did was just a very simple thing And just the amplitude here says amplitude there and transient area and these are all singles These are all doubles and these are all triples is a quadruple These are just spontaneous activity which we've got cell attached recording So we know the ground truth and you can just see you can easily separate them out roughly by two just very two simple parameters So we thought well actually maybe if we moved beyond a linear type filter We could actually do much better because ultimately what we wanted to do is get to the awake case What we wanted to do ultimately actually was to get to the wake freely moving case Which I'll talk about right near the end using this approach So We did a whole host of cell attached recordings We generated a lot of data just going up to the outside of the cell of an electrode listening to it while it fire You've produced both the calcium transients from the soma as well as the electrical recording You have your ground truth and then David built an algorithm and the algorithm the first thing we noticed is that this is the signal to noise between a Single action potential and the calcium transits is from about 36 different singles just to show you the rough noise signal It's about ratio of two which is doable What David built? Was a very nice approach where he used like a lookup if you like of all of our cell attached recordings He has a lookup of all the possibilities all the different combinations of spikes because spikes are additive They sort of interact with each other with calcium And he built this library he split the data in half He had testing data and data we actually reported out and the testing data was used to build this algorithm Which searches through for the most likely hypothesis. So it's a hypothesis-based tree search as he calls it And what you'll see here that runs through Here's some data the blue is the actual calcium data and the blue at the end is the actual fit and these are the Hypothesis that this thing's testing through a tree and then it comes up with a final solution if you like And then what we could do is head one free parameter, which we could then tune We use that one free parameter and we set it so we test it on test data set it and left it then the statistics we reported We use the other data we've never seen so we didn't overfit was the main point there How good was it this approach? Well, if you want to now turn this here, which is 13 overlaid calcium transients from 30 different cells In a wake animal, this is simple, right? We can all get this one We know where the calcium transient is and the algorithm does very well on these they can pull them out But this is more like what you see in vivo in the awake animal when it's running around you see these calcium transients here There's no way we can tell where they are by eye the algorithm goes through very nicely It's in green and this is the ground truth underneath that we looked at later And of course you could easily just do this by saying well I'm just going to say everything's a calcium transient I don't care and you'll be 100% right the problem is you'll be 100% wrong at the same time because you'll have false positives You'll pick up a lot of spurious spikes, and you'll start getting high firing rates You'll start reporting the wrong thing, which is not what it's not very desirable So it has to be very very good at rejecting noise But it has to be able to cope with many spikes at once very close together How good was it? So the accuracy that we report here is twofold the first accuracy we report here It's just if you want to detect spikes You don't care how many spikes you want in the bin you just care about spikes and you get about the sorry the green one here It's about 97% accurate But if you want to have a one-to-one relationship you want to say I want to know exactly how many spikes In each of these bins so if you want to know it's two three four five six seven nine ten or just one Then a one-to-one relationship You get about 93% accurate over all cells tested the average it's an average of course There's a lot there's a variance the lowest was 85% Detection and most of them 100% using this approach Okay, and then you can turn this into a raster here We can say how many spikes are per bin, but this is still the problem You can't talk about what those how close those spikes are within a bin You can only say that there were two in that bin, which is of course of issue still But I get to that in the end how we're solving that. Oh, yeah, and this is the other important bit It's false negatives. You want to keep your false positive rate Extremely racist hurts here. You want to keep it extremely low as low as you can get it So we tuned it our false positive rate to be lower than the lowest fine cell would ever got the cell attached. I Think that was a reasonable assumption Okay, then enter in the genetic encoded indicators now genetic encoded indicators are fabulous for so many reasons that I'm sure we're all aware of Because you can leave them in the tissue and they'll keep reporting back things for eternity Well in theory right until the animal dies of old age in the old age Pension or home of animals that we send them and what we can see here is the first sort of attempts at this is the work by Master Simon is a Winfrey Danck's lab and Damien Wallace out of my lab did all the cell attached for this Was he used a ratio metric die? We did a lot of cell attached report recording from mice and Move the whiskers and stuff like this It was very nice. We could detect single-action potentials very well. Here's a here's a cell attached recording. Here's a resulting R over R You can see here the ratio metric recording and what you see here is very nice You can see singles and doubles, but it wasn't the case for all cells and this is one of the problems We face the genetic indicators. It's not the case of all cells So what you see here is all the cells we recorded for the study and there's the average of the single-action potential Traces plus the variants so all the individual trials in gray Cells like this. You have no chance of detecting single-action potentials. So it's like this. You can't get them either There's no way you can pull it out in the noise so what we found was that although some cells were very Sensitive that reported out the sensitivity of the single-action potentials a lot of cells You couldn't it was very variable across the population. So you couldn't rely on it You couldn't then take an average sort of approach like we did with the OGB one the organic dies Which I showed you before we couldn't do that So we waited we waited for a better die and the major problem of the ratio metric die that That Mars who made was that it was good at detecting singles was but above three action potentials it is saturated So you couldn't distinguish between larger bursts of activity, which was Not desirable So enter in the brilliant Lauren Luger. He's a very good collaborator of mine And he came up G camp 3 which is a modification of G camp 2 which is a modification G camp 1 and we'll get to G camp 5 by the end That's true So again, they had the same problem the other Equal but opposite problem that was brilliant at detecting these large bursts when it came down to single-action potentials This is Exocellent recording that it failed again and actually in the end it turns out the variability was also in the second Two-action potentials so this thing was a little bit insensitive to the low end But still it was a revolutionary approach because it wasn't ratio metric It just used a single wavelength which means there's various reasons why that is of advantage Enter G camp 5 skip G camp 4. I don't know what happened to that, but G camp 5 G camp 5 was what they basically did was they know they had a DC shift if you like they just shifted the thing down and I Can explain that later why that's why that's a good approach And again, you can see here that it was beautiful reporting the burst and here's some cell attached recording again Here's a tuned cell the problem is that it's not picking up these single spikes here whatsoever again So it was very insensitive again to singles and doubles it's beautifully expressed Hundreds and hundreds of cells of respect that was very stable over time It wasn't toxic like you can't three and it seemed to seem to be a really good approach And you can see here again with the noise associated with single-action potentials and single trials So there's great hope that there's a hopefully G camp 6 7 or 8 or whatever's going to come out next So the G camp 6 will be the greatest thing. We're all hoping I mean it's a basic engineering problem at this point We know the problem they discuss their engineer when they solve it It's going to be brilliant because then you can tag specific cell types and so on and so forth So it's worth waiting for okay, so what's happening with the with the detection now What what are we doing about the detection? How are we going to make this better for everybody because ultimately what you want is you don't want to end up the Same problem of exocellular recording which we've talked about a few months ago in Stockholm You don't want to end up like you said in the beginning everyone has their own algorithm And they won't they would rather share that what their toothbrush you said or hair comb or something Share their their dog or something. I don't know but they'll rather do that than actually use each other's algorithms And we're going to try and avoid that because I've been working now with the incf. Luckily Thank God for the incf. We're going to try and solve this problem We're going to try and make a Universally applicable algorithm which does the dishes and that washes your clothes now Which is able to be used better to be trained on Genetic indicators organic indicators and worms flies mice rats monkeys so on and so forth. That's what we want Right, this is what we want everyone to use the same package have a website of open source We're everything you can upload your data or you can use test data off the website Whatever else it's gonna be beautiful. I'm sure So Josh Volkstein sort of made the first inroads into this and his his algorithm I don't fully understand. I'm not a mathematician, but I'll sort of stagger my way through it He uses a lot of parameters which you choose the parameters to biologically relevant Settings and this is a statistically based algorithm which will go through until you the probability of spikes over time and in bins Now it was brilliant because it works one vitro expresses uncertainty and learn can learn its own parameters The problems are it fell apart in vivo. It just collapsed and You only get one spike per bin so he made a new one which is he's working on now with David Greenberg out of my lab They're working furiously on this at the moment So what we're doing is we're using our baselining approach all the all the sort of empirical ways that we designed it for getting around All the in vivo problems, and we're combining with Josh's program plus making it much more efficient And already this is the test data that we've been running on Different people's data for people who've been very generous with their cell attached data from different animals and different cell types And they're giving us all this data here, and I think they mentioned at the end his G camp 5 L and V1 and frambo cells son zebrafish And into neurons people been very generous So we're hoping that the cell of all everyone who has cell attached if you have any cell attached in your bag out there Give it to me And we're going to try and make it university applicable and this should go online around about December And we're going to have it supported by NCF again, and the idea is we'll make it open source We're going to upload all of our cell attached data from our lab there So you can download it into your lab and play around with it and use it as ground truth for your own stuff and tune your own data That's the aim now as far as speed goes What are we doing about speed how are we going to how are we going to solve this problem speed? The good news is that physicists are smart and they've done everything before us We just got to realize that and use it for biology, and so there's this thing called AODs which you can use Gestalt devices which can move a beam around very very quickly in tissue normally use galvanic mirrors Which are these mirrors spinning on a on a stem with a little motor, and they're quite slow and cumbersome And if you can just bend it by changing the polarization then you're in you're in good you're in good shape The problem is at this stage you can go very fast speeds you can run at speeds like this and this The problem is of course is that you putting beams Electro laser beams through these blocks makes them non-linear, so you can't actually move them great sort of angles This is working for each of options That is hope there's great hope that we're going to go very fast But I think the major problem there won't be actually the speed of the laser system or the moving it around The major problem is going to be the dwell time which you spend on the cell Dwell time of how much signal you can get out how much time you spend exciting the fluorescence and get in and out That's going to be a major problem Okay, right That's it for the sort of the sort of Talking about the problems. I want to talk about what we can do about it what we're doing with it I should say so as I said in the beginning our lab is primarily interested in How rats use their vision to make decisions or freakin moving and here's our rat again. He's a very pretty fellow And of course We spent a lot of time doing head fixed and head fixes been it's been fantastic in a lot of respects because you get access To the top of the cortex and put electrodes in there. There's a two-photon electron microscope Objective lens you can see it shining IR light here. Here's v1 Popping away that we had to develop hollow movement correction algorithms because you're scanning in a raster So the animal moves the brain moves a little bit and you can have sharing effect halfway in the frame Which you can correct for by putting the pixels back in the right spaces is again work of David Greenberg So what we want to do is take it to the behaving animal and head fix case now these guys are great You can make these guys behave here and they press levers and They can discriminate between these different sort of moving gratings The problem is they don't care. They just couldn't care less They're quite hard to train to do this They're very slow to train to do this and they just don't care about the outcome. Why is that? Why don't they care? Why are they so hard to train in the head fix case? You can do it, but it takes a long time and In addition to that of course the mainstay of testing animals when their head fix is kind of weird It's kind of it's been incredibly instructive We know so much about the cortex due to this, but it's an illusion We don't see this very often in our lives In fact, the weird thing is that they had fixed and you're moving something in front of them But the most of our time We provide the movement which provides the movement across our retinas It's me moving around here. It's not you spinning around and it's a whole host of reasons why this is important Which I'll get to in a minute so we've sort of tested visual cortex using these sort of illusions and sort of Vomitous now, of course Lashley knew this in 1930 Lashley knew all this and he's the pioneer of the visual system in multiple respects and what he did is this very elegant experiment But I think still holds true today, and we try to replicate this He taught rats that they can discriminate between these different shapes that these different shapes actually mean something about the outcome So he placed them on the seat here, and he had a flap here with two instructions on it One told you which one was open to get the food pellet next one told you which one was shut So the random ones would jump through the flap get the food pellet everyone's happy They got it wrong that bounce off the door and hit the net marked in This is the 30s man. You can do that. So these guys were really motivated to get this right Right, they were really motivated They didn't take many trials to learn this because the outcome if they got it wrong was bad for them They didn't like it. We don't like it either. If you think about how we operate we also operate in the same way Part from if your home is Simpson Oh So if you bought this whole idea That you want to have a motivated behaving animal, and you want to have populations of neurons If you buy all this thing then really what you got to do is build a little microscope, which is what we did So the criteria had to fulfill as a few Had to be small enough to be carried on the head had ten minutes. Thank you Had to be small enough to be carried on the head So animals and drag in around because it's going to change the behavior You want to get populations of neurons a single cell resolution and you run a record from the same neurons from extended periods of time No point in your own disappearing in and out of view every so often and you want to have it fast enough that you can Relate it back to literal signal Why the main point is at the end If you think about it the surprise in findings that we've had in the cortex since 1970s the surprise in findings that we've had Have really come from friendly moving animals next to cellular recordings place cells grid cells Head direction cells song and so forth. These are the emergent properties of the network, which you could not have determined by modeling There's no way you could have done that you couldn't guess this You couldn't maybe got it lucky Somehow away with a ball or something This is what the beastie looks like little two-photon microscope has to be two-photon by the way You can't do this one-photon because one-photon just doesn't penetrate the cortex no matter what you're told And here is mounted on the head of the animal And we've also put a whole bunch of eye cameras on the high-speed eye cameras on the head of the animal now the track It's eye movements have now made a three-dimensional model of the eye and we can now track the receptive fields moving through space in real time Okay, so we placed on the visual cortex for obvious reasons You can use the Google Maps of the cortex, which is the blood vessel pattern To locate an area receptive field So you test for this particular receptive fields and then you can put your microscope of others and load the cells of calcium die And this is what the data looks like So here's the animal is moving around again We've used static greatings and we've let the animal do the movement which pushes pushes the greatings across its retina And you see here this the record is extremely stable These are all five minutes long the reason why we didn't do them longer each chunks because if you lose the data for some stupid reason We've got very old computers in our lab you'd cry so we chunked into five minutes But we recorded for hours like this from the same cells and this just goes on and the animal just runs around Okay, we had the track the animal I won't go into that It's quite a simple solution that David came up with you just have LEDs on the head multiple cameras You can solve for it quite easily. I can tell you about it later And once you have that of course if you know where the receptive field is You know where the LEDs are where the headers and pitch roll and your then you can project the receptive field back into space Once you've rendered your environment in three days, which is what we did And this is what we came up with here You can see here the tracking algorithm running over this and you can see the receptive field of a green box Been projected back into the space here and then you can look at the firing of the cells What do we find we found that there were some cells which are very tuned to particular features in the room very very tuned now We didn't know what feature they were Particularly and I'll get to them in a minute why that was the case Now we tried to turn it into a vector where you look at the amount of firing across the network and project it back Onto an object in the room Okay, the problem was this and this is really the last I'll just give a brief It's got a bit away from calcium imaging isn't it so a view of the eyes is this and All of the all of the imaging that we've done of eyes or recording We've done of eye movements in the rat have been done in the head fixed case and it's come up with the conclusion that these eyes are actually Conjugate they do conjugate I movements like we do like monkeys do and So forth and you can see this and they do faster caves. They seem to flick between two positions Now there was a problem with this whole idea Because we know from the literature there's extremely rich array of studies Which has shown that the VOR vestibulocular reflex in his animals is extremely strong Extremely strong. This would be known for decades So we're not even capturing probably 1% of The actual eye movements and the richness of the eye movements by having them a head fixed So what we did and why is this important is because when you have an animal moving around an environment They constantly move their head. They're never still why is that because they have whiskets They're constantly using their whiskers a sample round They're constantly moving their head pitching role in your the whole time and so the VOI is going to play a big role There's no doubt about this So when you actually build little cameras, we built these little in-house cameras We're super small they run at 60 hertz between 60 and 80 hertz full frame Two cameras very light 1 gram 1.6 gram sorry for both of them together You mount them on your two photo microscope here And this is a glass which you can animal can see through but reflects infrared so we can see the eyes But it can't see the reflective surface Once you do that a freely moving animal you see here that the move eyes move around in unexpected ways They move away around in a very rich way and this guy's just running around. He's not doing anything special He's not scared He's well trained. He's very happy rat and you can see here the eyes move around in a very Very strong way Now we had to track this tracking this was a hell of a problem, but we solved it David Greenberg again Solved this problem by tracking the eye Quite one of the people of the eye and we built a model now which we can now project gaze vectors into space Very accurately. So we're the the movement we can detect is less than 0.4 of a degree In all plants so we can now project that into space Okay, the only other animal we know of of course is the cocky monster which does same movements. It's my little Light relief what we also found was really peculiar What we found was that also they have very strong torsion The eyes talk which means they rotate on their sockets. We do it as well, but not that much So here's the animals eye talking just as it's running around So we had to try and track this and what we did is we peeled away this image Here made it flat and we just tracked this the border between these two structures You can see here at the bottom here tracking the The border between these two structures And what we found was that the eyes can move the same way talk the same way opposite ways the combinations are there of and they're independent They have independent talk as well I could go on for ages. Maybe I'll probably stop now. I mean, maybe I'll make this a wrong last slide slide Okay, so what we see here is we now make it do a task We make a jump a gap this gap is actually Not right in front of it. It's actually below it by 10 centimeters and across by 12 centimeters So the animal can't strike it We know that because we've got a detector on the side here of both tracks It's trying to strike it, but it won't and you see here the eye movements when it's making a A distance estimation because we move the track to different positions all the time So just very briefly just right at the end that we get to the very last slide What we basically the the take-home message from this whole thing Is that they actually keep their binocular region? Sorry, you don't look at these and just trying to get through them They keep their binocular region above them at all times They've got a massive binocular region which extends 63 degrees behind their head This is what huge showed in the 70s to 20 43 degrees below there below horizontal It's this huge binocular region and it turns out that this VOR what it does is it keeps this binocular region always above the head No matter what so if they tilt their heads like this this VOR the binocular region stays above them like this the whole time It's really amazing when you see it and the reason why that is because their eyes are the eyes are caught up on the side of the head And they're moving around like this and the reason why this is that they've had a huge selection pressure on them to have make this the case So you ask yourself why would they care about what's above them because they're ground-dwelling animals And this is what happens to them If they don't they don't actually monitor what's going on above them. This is the biggest predator Of a rat traditionally is flying birds So this guy's um Had it it's over for him And the reason why is probably because he wasn't monitoring about what's happening above them So it's the only sense they have to spot predators coming from a distance And these are the people who did the work I want to thank especially uh, these two guys here lauren luder and winfrey dank have been very good collaborators of mine They've been very instrumental in a lot of this work. Uh, the guys out of my lab, jogansavinsky, damien and david Did a lot of the damien's biologist, david's a mathematician and jogansay's physicist And winfrey was ever president in the first fruity moving study would do eye tracking was of jezeppine and toro institution sefi ruler grade student in my lab and These two again the deep imaging was done with wolfman mitman Wolfgang mitman yarn herb and andrey sheath and lauren luder and we didn't get to this part And of course, this is all funded by the max blank society and bernstein. Thank you very much So, uh, okay before we start with a panel discussion. Maybe you have some quick questions for jason So beautiful talk. I enjoyed it very much. Um you're able to scan through Pretty significant depth and extract the spiky activity Are you able have you have you looked at the spatiotemporal patterns throughout the layers and and and do you have some examples of that you could share with us or So that's that's that's that's a really good question. So And now I throw the caveats in right that was anesthetized. So the deep what is asked was the deep imaging What what was new when we got down to those depths? What was the new between we had all these cells What was the new what was new between all the correlation structure in the higher-order correlation structures? um It was anesthetized And so surprisingly so we've done a study many years before that we're in the zocton zone 2007 where we had Looked at layer two three in the barrel cortex and we'd moved the whisker and we looked at the spatiotemporal organization of activity And what we found was really interesting. It's very striking was that um The neurons were depending on where they were in the barrel It dictated their response properties and also the correlation between the neurons So down in layer five We found pretty much the same thing but not as strong But then again, I will throw in the caveat that we were using gcamp three Which I said before doesn't really detect singles and doubles And the majority of the firing in layer two three is singles and doubles when you move a whisker That said, um, I think Marcel Oberlander is talking to always has a poster He's got some very nice data which he shows Um That layer fives from christa cox work layer five neurons have depending on the cell type But thick tufted thin tufted will determine the the properties of the um of the output so We didn't do the full study where you would want to look at given the cell type What's the correlation between all the neurons? I expect that will pull up big differences and also of course layer fives the output layer And you can really divide cells up into those that project into basal ganglia and those that go out beyond further And those that go to the other side of the cortex and so on and so forth So I think there's some very rich Possibilities there when we start looking at the anatomy of that. So long story. No So on the final question there So on the calcium bump to spike transformation problem Why not use Two different dyes that emit at different wavelengths one optimized for single and the other one that doesn't Saturate and will report multiple. That's a great idea And so to extend that idea a lot of people again lauren luga have been thinking about trying to do something similar Where what they do is they load this the nucleus with red And the outside Side of plasma with green and you can have one tuned for singles one tuned for doubles um I don't think it worked that well The problem isn't getting single examples of this. It's making it work for all cells And the other problem is that these guys are buffets You're starting to screw around with the calcium concentration right and that's You get a lot of people to say it doesn't make any difference, but come on it must it must, you know, we're not we're not I mean you're packing the thing with gf. My wife's a molecular biologist. She does Fiddle rears cells protein protein interactions and stuff and she just laughs at me daily about The cells that are packed full of gfp sort of bulging out the seams of swollen cheeks Right and she says, you know, you've forced all this stuff into your cell How's it how the protein protein interactions going to work now? You've kind of just scrunched a lot You know something shoved into a train in tokyo downtown russia trying to read or something You know, I don't know. I'm not I don't know much about it But what I would say is that you know, you will change something for sure But just what and what had effect it has we don't know and we need to test that very carefully And that needs to go into the models and we have to be very aware of it So Um good idea, but I might cause problems. I don't know