>> Eric Horvitz: It's an honor to have Paxon Frady here with us today. Paxon did his undergrad work at Cal Tech working in Christof Koch's lab where he worked on the kinds of things that Christof Koch is known to be interested in, visual tension and deep neural networks, and then went off to join UCSD for his Ph.D. work with William Kristan, one of our long-term colleagues here we've been talking to for years about the kinds of things that Paxon will talk about today as being possible. The phrase that Bill and I came up with years ago was the whole idea of pursuing a computational microscope by analogy and by transition from light microscopy to electromicroscopy, electromicroscopy to computational microscopy where you have potentially noisy data coming in in high-dimensional spaces and you want to sort of process it and visualize this with a new kind of sets of views that collapse, clarify, compare, contrast, cluster, and that these kinds of new microscopes would be really critical one day for looking at neurobiological data especially multi-neuron data, among other kinds of complex biological data. And he's been working in Bill's lab for several years now. Just finished his Ph.D. work there, computational neuroscience. I think Paxon was probably our first and only since then neurobiology intern. I guess second because we had one right before that actually who went on to finish his Ph.D. work. And so it was fun having a summer of neurobiology with him here building tools and tool kits that might one day be the basis for computational microscopy and I think that looking at his work and his recent efforts what happened over four years I'm proud to think that some of his internship work was actually formative. We'll be hearing about that today. Paxon. >> Paxon Frady: Thanks, Eric. Hi everybody. I'm Paxon. I'm in Bill Kristan's lab at UCC San Diego. Like Eric said, I'm going to be telling but this concept we call computational microscopy. And it's really about using machine learning and developing tools in an interactive way so that we can really try to understand complicated systems. The brain is just the key example of a complicated system. We really need a lot of progress in this technology to move forward in understanding the brain. So I just want to start with this simple question, why is the brain so hard to understand. Why have we spent the last hundred plus years trying to figure this thing out and yet we still don't have a sound theory about how the brain works at all. And I'm going to just blame this thing we call reductionism and so we know the brain is very collection. It exists over a large number of scales. We think all the way down to the molecular nanometer size scale or synapsis and how these are kind of the foundational units of the brain up into neurons and small neural circuits up into these complex hierarchies of different brain areas into the entire brain itself. So bridging across these scales is a huge technical challenge. And using reductionism is also a philosophical problem. So reductionism we're taking this complex system and we're breaking it apart. And this has taken us a long way to understanding how the brain works. We now know that there are Synapses, there are neurons and these are all the pieces. We've learned, we've used reductionism to learn the pieces. But with the brain and with other complex systems you have to know more than the pieces. It's really also about how the pieces are put together and how they interact so these types of questions are outside of the scope of reductionism. So that's why we need to develop these new technologies and experimental ideas to get past this and understand how complexity emerges from organization. So this has kind of been the foundation of the brain initiative. And so this paper which is called the brain activity map project came out a few years ago just before the brain initiative kind of pointing out this huge problem and the very first quote that it starts off with is just talking about this phenomenon of emergence and how complexity and new properties can appear when lots of pieces are put together. And then along these types of problems, we have this book, the forth paradigm that came out when I first started here. This is centered around using all this data, this new revolution in big data and how to harness that. Part of the problem in big data is the ability to understand these complicated systems. When we were here at Microsoft, Eric wrote in this book this idea of a computational microscope and how to use these machine learning tools to make sense of these big data problems. The brain initiative is also another push in this direction. In neuroscience, we're kind of going through this big data resolution. We have all these types of new techniques that enable us to record from nervous systems at unprecedented scale and resolution. I'm going to tell you about voltage sensitive dye imaging but there's other imaging modalities like calcium imaging and multiunit electrode arrays. There's even new technology like light sheet microscopy that's giving us even more neurons. In the classical regime, we're doing electrophysiology. We can record from a couple of neurons. Maybe if you're bold you can record from a half a dozen neurons. And now, at the turn of the 21st Century, we've kind of entered into the next scale. So we're now at the dozens to hundreds of neurons scale with imaging. And then in the future, we're going to have to keep pushing. And then to get to the scale of the brain, the brain has a hundred billion neurons so we're still a long way away. But just like broaching this first scale is a huge issue. And it needs a lot of work. So I think over the last few decades we've started to realize something fundamentally important about how brains work and how neurons work and the insight is that in high-dimensional space, you can do a lot of powerful things in regards to computation. We have these chaotic neural networks now where if you just take a bunch of neurons and you randomly wire them together, you can use their chaotic Dynamics to learn arbitrary output patterns in really the power of these simple things is just that these neural networks are so high-dimensional that you can find a projection of them that gives you whatever pattern you want. So this is kind of illustrated in this little GIF where we have information of these words, words are these two-dimensional projections of information but if the they exist in a three dimensional space, we can look at them from different perspectives and the information is entirely different. So the word no and the word yes, it's there always in this three dimensional space but the information just depends on which way you look at it. So as we go up into higher and higher dimensions to really understand these types of systems, we have to look at them from all sorts of different angles and that's what the computational microscope is supposed to do. So how do we explore these high-dimensional systems? So we need lots of data. I'm going to tell but the voltage sensitive dye recordings been doing that give us large scale recording of a nervous system. We want to have multiple behavioral outputs so we want to see kind of a diverse array of complexity. And then we want to be able to consolidate it cross multiple experiments. So then we need the algorithms also to make sense of this data and two of the key things are going to be visualization so that we can see what we're looking at and then what I call synthesis where we take new things and we put them together with old things and finds the canonical relationships. We also want it to be scaleable because we're just now broaching the beginning of what needs to be a long increase in scale and as we keep going higher and higher in scale, we're going to need machine learning and algorithms more and more. And the other really important part of it is we need realtime and guided experiments. We need to be able to have these tools, take in some data, give us some visualizations and then use the tool to tell us what to do next or what's the target or what questions should we try to figure out with our experiments in realtime. I work on the leech. It's kind of a weird organism and before I started grad school, I didn't even know that was a thing people studied. But it's actually a pretty amazing beautiful organism. It has kind of its unique advantages over all the other model organisms out there in neuroscience. So this picture is just from a course I teach called neuro systems and behavior. And in this course, the students, they get to run through every organism that kind of the main organisms they use in neuroscience. We have mice and electric fish, zebra fish, fruit flies, the crab STG and leeches and I think there's a sea elegance somewhere. Maybe that's a sea elegance. I think that was too old school. For the leech, I'll get a little bit more into its nervous system in just a second but essentially it has these kinds of advantages. It's extremely accessible. We can cut the brain out and get access to the certification system. The nervous system is relatively simple. It's only the hundreds of neurons scale. We can get access to it while it's still performing behavior so the leech can an array of pretty interesting behavior while we can monitor the nervous system. It has defined cells. So cross neuron, across brains, across leech brains, it actually has the same cells individually cell by cell. This is pretty unique for almost compared to any other organism. Another important thing is we can do electrophysiology. So we can actually stick electrodes into cells and record voltage whereas like in sea elegance, there's another simple organism you actually can't do electrophysiology. They're just now getting electrophysiology and [indiscernible]. [Indiscernible] neurons are so tiny that if you try to stick an electrode in them, they essentially explode. But if you're really talented, you can do it. We also have these new voltage sensitive dyes which gives us a very unique perspective. In the imaging world, almost everyone images with calcium and this is kind of the second messenger of voltage and everyone really wants voltage but in the leech, we now have this new technology that allows us to image voltage directly. So this is what a leech looks like. It's just a segmented worm. Anterior, posterior end. If you stimulate it in different behaviors, you can get -- if you stimulate it in different places, you can get different behaviors so if you kind of shock it or poke it and the anterior end it will perform this behavior called shortening where it kind of withdraws. If you shock it in the back, it will try to swim away. So this is just like the sign wave muscle oscillation it uses to propel itself through the water. And then if you kind of apply a light stimulus somewhere to the middle of the body, it does this aversive reflex called local bending. So the way we've studied these behaviors over the years is what you do is first you take a full leech and then you just watch the behavior and you kind of maybe attach beads to the leech and look at the beads and get the kinematics of its movements. Then you can do these cool things where this is what my advisor got kind of famous on where you can actually cut out part of the animal to expose the nervous system and then you can do recordings from the nervous system while front and back half of the animal are still fully in tact. So you can see this leech like get shocked and then the front and back halves of trying to swim away while the middle of it is sitting there spinning out actual potentials. So this is kind of how we've been able to relate the behaviors to the motor neuron outputs. And for more complicated things like imaging, you can't have the animal flopping around or everything is going to be ruined so you can essentially take out the entire nervous system, take away all the muscles and the animal still performs these behaviors. So we can still shock the nervous system and get swimming and at the same time image one of the ganglion with voltage dyes. So the leech's nervous system is made out of a head brain and a tail brain and then this repeating structure we call a ganglion. There's 21 of them in the mid body. So all in all, there's about 10,000 neurons in the whole nervous system. But what's really cool is that each of these ganglia appears to be the same structure of 400 neurons just repeated. And even the head and tail brain appear to be ganglia just fused together over the course of evolution. And so we think that the leech's brain is made up of this one core structure made up of only 400 neurons. And so we're trying to just map this one core structure. So to do this, I'm going to image from this particular ganglion, ganglion ten. And then I'm going to be recording throughout the animal different extra cellular nerves that go -- these nerves go up to the muscle. So this is behavioral readout so I can record what the motor neurons are doing as well as input. So I can stimulate these nerves to kind of stimulate the animal in different places. Actually the different behaviors. So then this is what just a single ganglion looks like. So this is the ventral face of one of these ganglion. There's also a dorsal face. And then so there's cell bodying on the ventral face cell bodies on the dorsal face, and in between the two layers of cell body is where the leech nerve system makes all its synaptic connections. So in this picture, all these colored cells are the cells we've known about that over the last 50 or 60 years of leech nerve physiology, we've uncovered these colored cells mostly the sensory and some motor neurons. But then the vast majority of the rest of the cells are white because we don't really know anything about them. And actually, the white cells are highly under represented in this diagram. So we only know about a third of the cells so far. So to activate the different behaviors, we can just shock the nervous system so get shortening you can shock either one of these anterior nerves to get the local bend response actually and target this sensory P cell and activate the sensory P cell and we know this activates the local bend reflex and then to get swimming we stimulate these nerves back here posteriorly. So the whole nervous system will generate all these behaviors without any muscles attached. So we call these fictive behaviors. There's only about a third of the neurons that have been identified. So some of them, you can identify just by looking at them. So these two R cells, they're by far the biggest cell in the ganglion. They're always sitting right here. So you can just see by eye, you can know that they're the [indiscernible] cells. But then I kind of grayed out the sensory neurons over here so there's this line of sensory neurons but you can't necessarily distinguish them by eye. So you can guess that like these pretty large-ish shells over here, some of them are sensory neurons, like this might be an N or might be a P, or a T, but you can't tell just by I. Unfortunately what you can do is stick an electrode in and record their action potentials. And so a lot of cells have these unique action potential shapes so like the T cells are kind of bursty. The N cells have these really long after hyperpolarizations. They have unique defining characteristics. And that allows people to study them experimentally because you can reliably identify them across animals. There's a handful of other cells that have kind of these unique S potentials like the S cell. But the vast majority of the other cells, there's nothing really that characteristic about them. You see these kind of little dinky action potentials but there's nothing really obvious. And the techniques we used to identify the rest of the neurons rely on observing what they're doing during these different behaviors. Or some other kind of complex way of distinguishing them from other neurons. But this is hard to kind of do at scale and we can sues voltage sensitive dye imaging to actually kind of solve this problem and so we can record from almost all of these cells all at once while they're each doing individual behaviors and use that to distinguish the cells. So we use a voltage sensitive dye to record from these neurons. I'm not going to like do the mechanism of it but essentially, the dye gives us these amazing recordings so here, these are optical traces so we actually can see individual action potentials in these optical traces. We can see these large bursts. These are little 5-millivolt spikes in this particular cell and then we can even see oscillation. So down here in black is an intracellular voltage recording and then in green is an optical recording and so even the cell that's oscillating about 3 or 5 millivolts during swimming even this little oscillation you can pick up on the voltage dye or like little subthreshold synaptic potentials. So we have this new voltage chart a gives us an unprecedented resolution. >>: [Indiscernible] that's not actually potential [indiscernible]. >> Paxon Frady: >>: Yeah. So a lot of the cells -- [Indiscernible] basically. >> Paxon Frady: Right. Unlike most neurons, the leech neurons are all Monday know pore. So the soma is like way out here and they send the little process down into the neuro pill. And then they make all their synaptic connections and generate the axons out here. I mean generate axon potentials out here. So what you record at the soma is a very filtered axon potential. So a lot of times you can't even see them and you see just like the slow membrane oscillation. >>: [Indiscernible]. >> Paxon Frady: Right. The way I kind of envision this, the kind of idea of the computational microscope is to go from here where we have just an experiment we collect some raw imaging data, we transform that into cell by cell data, and then we're going transform that into like characteristic features and then we're going consolidate this into some kind of canonical framework or like the homologs of the same cells. And I kinds of like to think of this as like a deep neural network kind of where we have like these are just different layers and so this is like 10,000 dimensional feature space this is like a hundred dimensional feature space and then this is like ten dimensional feature space and then we do some clustering here to come up with some canonical formation. And this kind of process is what I call the imaging computational microscope. The first step is to taker a raw movie. So just like pixels and an image and a bunch of frames and turn that into individual cellular components. Here's an old screen shot of what in mal app GUI looks like. So essentially like here's just some raw data over here I've selected this particular ROI. So up here is exactly what the ROI is over time. So like the average within the ROI over all frames. So you can see this cell spiking. And down here is actually the automatically extracted component that corresponds to this ROI. So all these ROIs are actually generated automatically. So I'm going to tell you how that works now. So the first thing do you is take your image ask we're going to try to reduce the dimensionality down and we're just going to use PCA. The other thing that I've already done is I've taken all of the data. All these datas I have multiple trials so I have multiple trials of shortening, swimming. Each trial is a ten second long movie. So what I've done is I've used an image registration algorithm to actually align all the trials up and pretend like everything is just like a single time series. So I just had this one big movie. So then you do PCA on the individual pixels of this big movie and then you get two things out. So you get a map and you get sources. So here since I have several trial in this one movie, the map is going to be the same for all the trials, but then these traces can be just broken back up into the individual trials. So for a single animal then, you'll do PCA and then each component will have a map and then it will have multiple traces and then the number of traces are just the same number of trials. So we're mainly going to use PCA to reduce conventionality. PCA tell us a lot about what's in the data. Usually if you just look at these maps, what you see is these mixtures of cells and that's just because PCA is just looking at volume and mixtures of cells, louder than individual cells. But there's this algorithm called ICA which looks for independence. And this will pull out signals that we're interested in and so most of the time what you'll see something like this where you have this kind of localized spatial region and then you have these traces corresponding to the activity at that region so this is like pulling out an individual cell. So here's just another cell and so here this is actually two different trials so one trial is swimming and this cell oscillates and this trial is shortening and it just does this early part. But ICA, these algorithms are just completely naïve to what you're telling it and they don't know the difference between a cell and an artifact but what's really nice about ICA is it actually will extract and separate cellular components and artifactual components and then in the GUI you can go in, manipulate the parameters, all these kind of magic numbers like the dimensionality of your PCA and toy with it until you have something that you like and then you also can then manually kind of remove these components that you don't want. So then after you've removed the artifacts, there's a segmentation stage essentially just applying a threshold to these ICA maps to kind of get a binary mask and then occasionally the ICA will actually cluster cells together into a single component. So in this case, there are these two bilateral cells and they essentially are firing spike for spike because they're strongly electrically coupled to each other and so ICA just thinks they're the same cell because their activity is the same. By using spatial segmentation just to break them apart again and then that generates a bunch of ROIs. Finally, sometimes it does the opposite thing of this where it breaks apart individual cells. And again, you can just use the GUI to put them back together. So that's like how we get that data out. And next we're going to try to describe the individual cell data with a handful of features. But first let me just show you kind of what the data looks like. So here is shortening. So this is kind of false colored. Effectively, you hear the stimulus come on and then you'll see a bunch of cells turn green and red. Green just means it's depolarizing and red means that are hyperpolarizing. If you look really closely, right after the stimulus, you see this kind of flash of green and this slower flash of red. This is realtime. A lot of stuff is happening, right? >>: [Indiscernible]. >>: [Indiscernible] after the actual signal's input to the system. you're specifically looking at change. >> Paxon Frady: >>: Right. Yes. So So all of the optical -- [Indiscernible] color all the time, right? >> Paxon Frady: Yeah. But the optical recordings are always relative. you can't get the absolute voltage values. >>: You're conditioning on this background, stabilize -- >> Paxon Frady: >>: So And then it changes. It changes. >> Paxon Frady: Right. So for shortening, you see a bunch of different types of response. Some cells go up, some cells kind of go down. Some cells go up and then down. Some cells have this late phase. Some cells very early. But really, all these responses can be pretty well characterized by these two factors. You can think about these as the principal components. So essentially there's factor one which is this slower component which corresponds to cells kind of getting depolarized or hyperpolarize in conjunction with the actual motor output and then there's this factor two which is this fast component which is cells mainly getting depolarized due to the stimulus. So then you fit these factors on to all of these traces and you can get coefficients and then we can plot all of the individual cells into these two dimensions. And then I can give the cells colors just based on where these coefficients are and then use the ICA maps color the maps and create this activity map. And so with the activity map we can kinds of get a visual sense of Wells Fargo what the cells are doing because now we can plot four dimensions here, two per space and color representing these two shortening factors. So now we're kind of characterizing the cells with this slow dimensional representation. So here's swimming. So swimming is this repetitive motor burst so that leeches doing this sign wave oscillation into the water. So what you can see is a bunch of cells that oscillate with this motor burst so this is like the easiest most obvious one. All the circles oscillate so just going red to green red to green red to green and then if you look at a different circle, if you look at this yellow circle you'll see it oscillating too but it's actually at a different phase. So to characterize swimming we're going to do what we call a coherence analysis so here's this motor, this repetitive motor neuron output burst and then here a bunch of different cells and the voltage is oscillating with this motor neuron burst and so we do the correlation and foray space which tells us a correlation magnitude and phase so this is just giving us now another two dimensions to characterize these oscillations and so you can see a bunch of different cells are oscillating in different phases and then give these phases a color and again you can make an activity map and get a visualization of what the cells are doing during swimming. And then for local bending we're going to do the same analysis so the local bending experiment is actually just activating this century P cell at a repetitive input so this is just like a half a hertz stimulus input and then I'm measuring the coherence between this stimulus and the rest of the voltage so effectively it pulls out cells in this pink phase which are getting depolarized by this stimulus and cells in this scion phase which are getting hyperpolarized and again, we can make an activity map and get visualization. So the important thing to make sure that's true is to make sure that like the ganglion across trials, so individual animal performing shortening, recognize that the neurons do something consistent or at least some of them are consistent, right? So it wouldn't be very useful if it's neurons were doing different things on every trial shortening because you couldn't really use that to identify them. So here I'm just going to show you like some visualizations to get a sense of the variability across animals and across trials. So each of these four boxes is from a different animal and then for each animal there's two trials of shortening being shown. Okay. So these are just two separate trials. And you can see like the activity maps so you can see across individual trials they're really similar. And then over here are just the same coefficients that I described to you just a second ago. So from one trial, I fitted those factors and to every trace and that gives me one end of each of these line segments. And then for another trial, I fit the factors again and other end so each line segment is the activity of a single cell across two trials. So the start and end indicate what that single cell is doing across these two different trials. And so then the size and the kind of isolation of the line segment is an indicator of how consistent that cell was across trials and how different it was from other cells so small isolated line segments mean that we have a lot of identifying information. And then visually you can kind of look across animals and get an appreciation like the ganglion during shortening is effectively doing something pretty similar. And so that's good because we need across animals for the neurons to be doing similar things during these behaviors. >>: How would it change the picture if you had same animal performing different actions, swimming versus some other thing? Would you see a very different picture? >> Paxon Frady: So you mean like if you applied this -- >>: This is kind of qualitative, right? It's not quantitative. And I'm trying to -- so you make the claim that this is reproducible, so if we would have had one image in which it was swimming and the other image in which it was performing some other action, we expect to see that we'll have long lines, right? >> Paxon Frady: >>: Right. Very long line, stuff like that. Do you have something like that? >> Paxon Frady: I've never even really thought of trying do that. So here's swimming. Comparing it across? I mean, what it would look like is that in one trial, you would have something -- if you did the swim analysis like a swim, it looks like this, right? You have a bunch of neurons that are significantly coherent. If you try to do this analysis on a shortening like nothing is oscillating, everything is going to just be clumped in the middle, like this, so all the lines will be lines out here pointing in to the middle like this. Radiate like that. >>: Let me ask a question. I'm curious if you showed this to Bill, your committee and your advisors or other neurobiologists, be interesting to know what kind of insights just a question that they held for a long time might come to their mind for like helping the traditional people in this field who have standing questions. First step sort of visualizing that it's beautiful, say lots of things about these animals, for example, but I wonder if we assume the question Bill's about circuits for example and circuitry could be answered directly by these visualizations. >> Paxon Frady: >>: Yeah. Obviously this is towards that, right? >> Paxon Frady: This is a step towards that. >>: Visualizations, one reaction is they're beautiful and I see lots of things going on and I can learn many things including maybe the personality of a cell and the match of a cell among animals. But I'm just curious if any other questions came up over time that people say, wow, I just had this one functionality here where I'm seeing something about these two cells that makes me think they're interesting way beyond ninety degrees. >> Paxon Frady: Yeah. So one of the things that stands out long these lines, maybe a little different than what you're saying, but does the same cell across behaviors do the same correlated things. So if you're in this phase of swimming, does that imply you're over here during shortening or are these things independent? Right? And that goes a long way in talking about how neurons encode information about what they're doing and what they're sensing and I think a big thing in the field now is that you have these neurons that are multifunctional that really are kind of representing these joint features and for whatever -- I mean, there's computational theoretical reasons like every neuron trying to be as independent as possible something to go for. So then observing the same neurons doing a bunch of behaviors, what you see is like there's only a handful of cases where like they're really correlated across behaviors and actually I think pulling out those correlations tells you some of the structure that you're talking about. And then I think when I get to the preparatory network too, that will also be like some illumination. >>: So when I look at these images, there are differences between animals. So in each cell, the left and right image are more similar [indiscernible] same animal. >> Paxon Frady: Yes. These two are the same animal. >>: Do you know or speculate about the differences between animals? Does it stem from structure differences between animals are just random formation differences or if these are more like the way they initiate the swimming patterns are different between animal? Do you know where the variation comes from? >> Paxon Frady: Yeah. So there's certainly variation. But really, it's remarkably consistent and so it's consistency can be expressed in that set of cells that we've already identified and so if you open up any leech ganglion, you can immediately find a bunch of cells, P cells, a bunch of motor neurons. And so for about a third of the cells that we've identified so far, they're always there and they're always in the same function. So there's definitely slight variability in exactly how the ganglion puts together and that's actually why I actually didn't mention this but like the reason you can't just like solely rely on where the cells are is because there is like this anatomical variability. Sometimes the cells get pushed around. And even sometimes by doing this dissection I stab it or something. I have to cut off this little tiny capsule and so the cells can move around and the capsule is actually kind of compressing holding cells and so when I cut it off, they kind of like expand out so there's slight anatomical variability. And so that's what makes it hard to identify the cells too. But as far as there being functional variability or there being learning or something, environment dependent, it's not entirely clear. Everything we studied so far points no, but we're kinds of biased because we have mainly only studied the sensory and motor neuron and perhaps there's this huge array of interneurons and that kind of space is flexible. That's not entirely clear but from this where it seems like all these cells seem to be very consistent as well, so it seems like the animal has this pretty well defined functional homology. So swimming, so again, it's the same as before. There's two trials of swimming. And then each cell has a line segment indicating each of these two trials. And so as you can see, swimming is remarkable. There's a bunch of line segments. They're really well spread out around the outside edges. They're very short. So swimming is very promising feature for identifying cells. And then here's the same thing with local bending. And so mainly what you see across local bending is essentially this one access so there's these pink cells which are getting depolarized and these kind of scion or green cells which are getting hyperpolarized. But again, at least at these extremes, there seem to be a handful of lines or so that are pretty well isolated. So then the final step is taking all these animals and putting them into a single canonical representation. And so here's kind of the gist of the strategy. So if we just zoom in on these five cells, so based on just their position and their size, you couldn't really distinguish them from each other so they're too close to the other. There's nothing really too characteristic about them. But then if you do shortening and you get these shortening features, then they start so get separated so now you can tell like one from five and five from two and three. But you can't really distinguish between two and three. So they're still -- they're doing the same thing and they're in the same place so you can't really distinguish them. And then if you look at swimming, get the same gist. So seeing swimming helps you kind of distinguish some of them but maybe not all of them from each other. So now we can distinguish between two and four and three and five because they're in the same place, doing the same thing but by like combining all these features, we're creating this higher dimensional feature space and then this allows separation of all the cells individually. So this is our strategy so we're going to take all -- we're going to make a ten-dimensional feature space so we're going to have two for the position of the cells, two just indicating the size of the cells, and then two more for each of these three behaviors. So that's going to be our canonical feature space. So then with this, we kind of developed these ideas a long time ago. >>: Back in the 60s. Rebels. [Laughter] >> Paxon Frady: Yeah. So again, there's another interface that kind of lets you explore, compare all these feature maps and then explore this feature space and then we have a couple of machine learning algorithms that manipulate the feature space to kind of accent certain things. So there's an iterative kind of strategy. So essentially, the idea is like we're comparing these two animals. This is animal H, this is animal C and then here just like the activity maps so if you want to reference that. So then animal C I'm going to select three cells manually and just indicate it by these three ROIs and then the algorithm is going to use like the Hungarian algorithm for instance in this particular case to guess which cells are homologous in this other animal. So the computer has told me that it thinks this cell is the green cell, this cell is a blue cell, corresponding blue cell, and this red cell is a corresponding red cell and then it also shows me in this color heat map the kind of relationships so the brighter green indicates that these other cells are more like this so this is like the distance in this ten-dimensional feature space that all these cells are away from this particular green cell. And so we can essentially visualize this high-dimensional space by picking a cell and then you can see the cells that are similar to it and that gives you an idea of like what are the possibles. So then we can then tell the computer that oh, I'm very confident that these cells are matches and I'm very confident that these cells are matches. And based on these matches, we develop some learning algorithms that warp the future space to kind of accent the important features. So just as this is the very simplest version of the algorithm, but effectively what you can see is so here's just like the normal regular feature space for these two animals, so this plot over here is indicating the position of the cells. This is indicating the shortening factors. This is indicating the local bend coherence, and this is indicating the swim coherence. So this is just what it looks like when you don't do anything. You just plug in the features so then after warping so I've assigned -- so there's really 8 different animals in this data set and then I've gone through and I've assigned matches across all these different 8 animals and then in this iterative procedure, once I assigned matches, this algorithm can show me this warped distance space and that emphasizes the features that are important for the matches. So it kind of -- and it ends up showing you kinds of what our visual intuitions that led us to believe earlier. So position gets kind of stretched out. Swimming gets a little stretched out. And so those feature dimensions are emphasized. Whereas shortening kinds of gets squashed because it's not quite as distinguishing. And then local bending gets squished long this one access. And as we saw, it was essentially this one dimension of local bending that was informative. So then this gives us ways to use machine learning to kind of warp the future space and pull out the things that are more interesting and relevant. And this is how we kind of explore this high-dimensional space. Here's some gifts. So this is just to like kind of show you give you some visual sense of what's happening so this is like the principal components of the ten dimension feature space and then every single cell that I've matched across 8 animals or shown, okay, and then so each cell that's in a single cluster is colored the same and connected by a little convex hall. The point of this is really just so you can see that after doing this warping algorithm that the clusters come out much more nicely and that you can start to tease apart the cluster cells a little bit better. Really, there's a lot of corruption in feature space, like the position of the ganglion like some ganglions are oriented like this a little bit relative to others so just kinds of compensating for those minor changes can take you a long way in cleaning up this high-dimensional space. And then also from this, you can really get a sense of like which cells you're going to get confused so the nut and the AE cells are kind of like in the same cluster but it's hard to understand how to separate them because I'm not just solely relying on this to do the clustering. I'm also going in and visually assigning these matches so this is like this algorithm and my own refinement. So this gives us all these cells and now I'm going to describe the canonical set of cells. This is an entry I'm going to show you for a single cell so this particular cell is called 151, just numbered based on this standard number system that the leech community has developed. And then again we have the position shortening local bend and swimming features plotted here. >>: [Indiscernible]. >> Paxon Frady: Yeah. That's exactly what I was about to say. So these eight circles correspond to the eight animals that are in this experimental data set. And then the fill-in just indicates which experiments this particular set was seen. So cross experiments, we're not going to necessarily see all the cells. We're going to see some subset of the cells just because there's anatomical difference sometimes it's just due to the dye not giving enough signal. So this just tells us which experiments this cell has founds and then these circles tell you where the cells were. This tells are what it does during shortening. So this cell has a big fast component and a pretty large depolarization with a motor neuron and this cell almost always gets excited by P cell stimulus. Got inhibited one time. And then it typically oscillates in this red phase during swimming. So that's kinds of how you read this table. So just another cell. So this cell down here kind of posterior lateral. So this cell gets hyperpolarized during shortening. It has both kind of responses during local bending and then it oscillates in this kind of green scion face. So here's the whole table of cells that we were able to identify across animals. So almost all of them are swim oscillators. The only ones that aren't are indicated by these asterisks. And then I have kind of outlined in yellow these particular cells which share a kind of interesting set of features. So like we were kind of saying earlier how these features, there's computational reasons why the nervous system would want to make its featured representation independent but then sometimes the features come out to be correlated and I think this points out like some intrinsic structure. So this yellow network is what I call the preparatory network and you can kind of see it has this characteristic two things. So one is that during local bending it's excited by the P cell stimulus so directly excited by sensory input and then during shortening it has this big fast component. So this AP cell is kind of the mother of this network so you see this big fast component so it has this rapid depolarization during shortening and it has direct sensory input. So then this is just the canonical maps of these networks so I went in and looked at the preparatory network much more closely so in this case, if you remember, I'm always imaging from ganglion ten. But then I'm going to be stimulating different ganglion to elicit these different behaviors so just depending on how far away the ganglia I stimulate, it will take longer for the cells to depolarize. So here in the bottom row, it makes a little more sense. So the ganglia are just sorted by how far away they are from ganglion ten. Then I would go in and you can find these inflection points where the cells depolarize and you can mark the beginning of these depolarizations. And that tells you the timing of the response. So remember that like the 14 and 17 and tail are eliciting swimming and then 7 and three are eliciting shortening and so you can see these neurons, they don't really look different based on the ganglion. The timing is slightly slower, but it's effectively always depolarizing. So then, here's just like three examples that are in the preparatory network and then just here's some other cells that are not in the preparatory network in the same kind of gist. >>: [Indiscernible]. >> Paxon Frady: >>: No. This is not in the literature. So what do you mean? >> Paxon Frady: This is why we call it the preparatory network. The first thing is that it's rapid. It's receiving rapid input from every behavior. This is showing us this rapid not only during shortening which we got from the factor two, but also, it's rapid now from swimming as well. And it's also receiving sensory input directly. >>: Solution between [indiscernible]? >> Paxon Frady: Yeah. Because they shared this common feature dimension. >>: [Indiscernible] jumping up a level, when it comes to single cells, how likely to richer categorizations of sets of cells that might be same phase, shared attributes and then once you have [indiscernible], does this set have -- I guess my question would be calling it network versus set of cells. So the question is what gives you the sense to call it a network? Everything is connected eventually. >> Paxon Frady: Now that we have this hypothesis, the preparatory network and the ability to use this computational microscope in realtime, I can go into new animal, I can identify the preparatory network, and then I can target it and do more experiments so that's why I'm heading with this. I'm trying to establish why calling a preparatory network. You can just fit these response curves based on distances and so then you can see why you see it's like you plot the response times of all the preparatory cells, they're all faster in this blue clump which are just another handful of cells that I've chosen out of cell identified. So the preparatory network is consistently fast and it has nothing to do with behavior and essentially looks the same across animals too. So the reason ->>: [Indiscernible]. >> Paxon Frady: ten. >>: That's just like how far away the stimulus from ganglion [Indiscernible]. >> Paxon Frady: Yeah. Well, I mean, so this is like something that Bill kind of hypothesized a long time ago that the sequence is more like there's a stimulus, do something, he called it the do-something network, just like get ready, like prime the muscles. And then there's like the decision and then there's the actual execution of the behavior. >>: Just the do something. >> Paxon Frady: That's what I think. There's no real behavioral defined thing. It's all about the stimulus and it looks the same regardless of the behavior. Even just activating center cells activates this similar network. >>: It's consistent. What do you mean by consistent? >> Paxon Frady: I mean it's the same cells regardless of the behavior. So this is all one animal and these are cross behaviors. So these are just the on set times during these behaviors of these cells. >>: You cannot distinguish but you can see them piling up every time. >> Paxon Frady: Right. Yeah. You can't tell a behavior is going happen from just the preparatory network. >>: But you could just suggest [indiscernible] differences are in the decision. >> Paxon Frady: Yeah. >>: If you know that some prep going on and you hypothesize that that was some sort of a queueing up of something. You think something about the prep's end, show you little bit more about the do something, what it is do. >> Paxon Frady: >>: Yeah. We don't know. Queuing frame to look for the differences. >> Paxon Frady: Right. >>: Look over here now as the -- actually compare more deeply the difference between the prep activities for different behaviors downstream. Maybe show earlier in the prep level. >> Paxon Frady: >>: It's possible. Prep is really a decision being made. >> Paxon Frady: Maybe. So then the really important part of the computation microscope is to close the loop. >>: It's like this is our dream, BOI dream. >> Paxon Frady: So this is like kind of what we call computational guided experiments or computational guided electrophysiology. And so there's just like a verification study. So I like find a new animal and then I generate this swim activity map and that allows me to identify a handful of swim oscillators and then I can go in and target them and so I'm just trying to verify that these cells are really doing what the volt dyes say they're doing and turns out essentially yes. So I can now come and have a new animal. I can image a trial of swimming, do this analysis, produce canonical relationships and then target these cell specifically for further experiments. So there's just three ->>: So new animal, got your old data set, old analysis, and then it's going to image it and the analysis [indiscernible]? >> Paxon Frady: Yeah. I think that this component is closest to in the future space to this cell that we see already, like 152. Right. And then the computer can tell me what it thinks the cells are and then I can go in and target them. >>: The interesting thing do would be to [indiscernible]. >> Paxon Frady: >>: Yeah. Hasn't gotten that far yet. You're close. >> Paxon Frady: I'm close. >>: You could basically say, hey, listen, I don't need to waste my time sticking the needle in 152. Where there's uncertainty right now is 201. >> Paxon Frady: >>: Right. And that's where I put my electrodes. >>: That's where [indiscernible] directly. [indiscernible]. >>: Not waste time. The same that we're using Correspondence. >> Paxon Frady: Yeah. So I mean, still building towards that but this is the first. So then I can tell you that the S cell network and so we know about this for a long time and so one of the cells that are part of this table is part of the preparatory network is this cell called the S cell. And it's really interesting because it's one of the few cell there are not bilaterally paired so most cells have bilateral pairs. But it also forms this like electrically coupled network that goes up and down the entire length of the animal. So every S cell has this huge axon that's electrically coupled to every other S cell up and down the length of the animal. So it's been characterized as like a giant fiber system. So the hypothesis is that the S cell is mediating the preparatory network so that the distal stimuli are coming in through this giant fiber system activating the preparatory network to get the animal ready to produce the behaviors. So then we can use the voltage dyes to kind of like answer this question. And so fortunately the S cell you can identify it just with electrophysiology. It has these really big, really skinny action potentials. And in this particular experiment, I ran a swim, I identified a couple cells and then I found the S cell and then I targeted the S cell with an electrode and I just excited it while also recording the rest of the ganglion with the voltage dyes. So then I'm just passing this square wave of current into the S cell so that's the optical recording of the S cell. And then you can see like a bunch of cell getting depolarized and a bunch of cells getting hyperpolarized in response to this S cell stimulus. And it turns out that these three cells are also part of the preparatory network. So this is just exploring for the S cell connections and it seems like the S cell is likely connected to a handful of other preparatory cells and then we can verify this by again -- so here's a different way of identifying preparatory cells so if you remember like a lot of the preparatory cells had this pink phase during this local bin behavior. So this is just like one of the easiest behaviors to do to identify it. It's like the simplest experiment so then with this, we can identify AP 511 and 153 which are all part of the preparatory network and then I can then go in with electrophysiology and actually do the intra cellular recording to verify that this synapsis are real. And so you can see like the S cell to the AP cell where the AP cell is just kind of the mother of the preparatory network you can see like this one for one UPSP for SAP and then the S cell is also connected to 155. Another preparatory neuron. It's weakly connected to 153. And then you can see that it has no connections to these other two cells in the same area with that or not part of the preparatory network. So then here's kind of the summary so their cell bodying kind of spread out like this through the ganglion. We have the S cell which is going to be forming this electrical giant fibro system with other ganglion animal. And then sensory evoked stimuli are going to come into either anterior or posterior through this S cell network, activate the S cell which in turn activates the rest of the network or like these local sensory cells are going to activate the preparatory network directly and also activate the S cell which likely sends this information to the rest of the animal. >>: [Indiscernible] on this last three slides. So can you summarize what about the analysis for virtualization help you identify preparatory network and then do links between AP and the preparatory network, what about visualization helped with that that you couldn't have done without visualizations? >> Paxon Frady: Well, so mostly, I rely on visualization more than the actual algorithms like for the matching. The visualizations are simple enough especially because I'm not really looking at all ten dimensions because like the cells that I've targeted, you can just do like one local bend to I'd 155 instead of having do all the behaviors. So just like getting a visualization, like for me now, I've seen enough swim activity maps that like I can just look at it and now exactly what cells are. So just like the training on the visualizations. >>: [Indiscernible]. >> Paxon Frady: Yeah. >>: That helped you recognize the -- identify formula network and then link it to explore further in terms of the S cell connection with the other cells. >> Paxon Frady: Yeah. >>: Again, nice to know what would it have taken to get to these insights without doing the investment in visualization matching. >> Paxon Frady: Yeah. Like just doing the mapping was important for feigning the preparatory network because then I really was starting to see that a lot of these cells were really similar and that they shared these features and that prompted me to go look at them much more closely. >>: Do you think like zoning [indiscernible] kind of zooming into counter matches helped? For instance, instead of [indiscernible] few candidates. Do you think that was helpful at all? >> Paxon Frady: Yes. I think being able to see the distances. Sort of like I click on a cell and you can see, that helps a lot. A lot of times it will ->>: [Indiscernible]. >> Paxon Frady: Exactly. And a lot of times it will point a cell that you weren't even expecting. There's one cell 232 which it can be almost anywhere and down this central packet and at first it took me a long time to realize it was there but then I started seeing it and then it started highlighting it to me and then I was able to visually confirm it. So here's the overview of the computational microscope and now I plotted to mapping the leech's nervous system. So we have these now automated techniques which can extract signals from raw imaging data without really much user input and then we can make these activity maps or visualization and put them together into some kind of canonical space and then we can map out the function of a bunch of the cell and make these canonical maps for future reference and then we can kind of use these tools to ask new questions and kind of do these experiments which are at a whole another scale that we can never really do before. Then I kind of want to tell you like this other -- where we've been doing this brain initiative and trying to kind of fully characterize this nervous system and so everything I've showed you was on one face of the ganglions really only imaging half of ganglion but our collaborators have now built a new microscope that has essentially an objective on top and on the bottom and I went out there and helped set this up and now we can literally image from every single cell in the ganglion at once and so now record from the entire nervous system. They're in Cincinnati now. [Indiscernible] just started his lab in Cincinnati about two years ago. >>: [Indiscernible]. >> Paxon Frady: >>: I see. What's done about that other half of the sphere? >> Paxon Frady: >>: And that is Daniel's post doc. There's quite a bit. It's kind of like the same thing. Kind of like another duplication. >> Paxon Frady: Yeah. It's like the motor neurons in this. There are more motor neurons on the other side. It's actually kind of like I think like Bill and Anna Free son agreed in the 1960s that one would do the ventral side and one would do the dorsal side. I mean, it's like kind of similar. It's about a third of the cells, mostly the large kind of motor neurons that are known. There aren't quite as many cells on the other side either. You can see this like stripe down the middle. This is actually the connective so like the axons coming in from the other ganglia and going through. >>: [Indiscernible]. >> Paxon Frady: >>: Yeah. Daniel is a post doc in Bill's lab. Speaking of the old days. >> Paxon Frady: Yeah. Yeah. He was still there when I started. He was just leaving. So we have this and then we have my graduate student colleague JSON and Bill's lab. We're working on getting the serial block face EM reconstruction of the leech so if you guys heard of the connect dome, the connect dome stuff, this is like how you do the connect dome. So it's our dream has always been to like image the ganglion, image all these behaviors and then take the same ganglion and get the EM connect dome so that we can relate the function of all the neurons with the structure and anatomy and so the EM is just amazing. You can see these nano meter resolutions. This is leech EM, yeah. Just like look how complicated it is. It's just so incredible. It's like here, just this cell makes like 17 synaptic connections to this other cell. Just one cell. 17 different synapsis. >>: [Indiscernible]. >> Paxon Frady: Yeah. Need more than one. Need like 17. And then we are also working with Larry Abbott on doing in computational modeling so he has these recurrent neural networks that can produce these patterns and so we're trying to consolidate all this data into the single framework. They're at Columbia. So I just wanted to thank everyone. This is the lab and then my advisor Bill. And thank Kevin Brinkman who gave a lot of input at the beginning of this project. Ashish and Eric. And then Roger Chen's lab for the voltage dyes and that's my committee. So do you guys have any questions? [Applause] >>: So those last 3 or 4 slides, do those folks, are they intrigued by the computational microscope and visualization ideas or are they just going off in the directions, those tools? >> Paxon Frady: I think they are. I think it's kind of like another step. Like after we kind of defined the canonical space, the mapping, the next step would be to take the data and make a model from it and kind of incorporate the Dynamics and the models too. >>: [Indiscernible]. It's great. Very exciting to see how far you came from the internship project to real world impact. I have to say, I've seen a bunch of BRAIN acronym initiative studies but even though this is on a simple system, this seems like it's made the most progress in terms of understanding what's going on. What comment here, even leeches, it's pretty clear that the fabric upon which we're perceiving and thinking now is based on some derivative of these earlier systems, earlier evolutionary tree. It's hard to know whether or not the insights that will come out of these tools directly apply but certainly tools will apply to understand the systems. Just one comment that excites me about -- besides [indiscernible] about the existence of doing, the fact that we're listening, hearing, and seeing based on these same kinds of networks to the unseasoned eye or microscope or computational microscope was about the same close up. Any other questions or comments? Thanks a lot. [Applause] >> Paxon Frady: Thanks guys.