 It's a life of vision. Well, first of all, thank you very much for having me here. This is a really interesting one, obviously. And I should first of all thank Tom. It could then be a fantastic summary or review of Mike's work for us tonight. So when I would refer to a few of his works during my career, we haven't done a good thing with these ones, which is really great. Now, we have all our memories with Mike. And I think I met Mike for the first time in the mid-90s when there was an ICM conference at Cambridge, I think. And I can't remember from there, but it was in there. Or meet him there for the first time. Later on, when I had moved to Cambridge and was working with Simon, I met him a few times and once it was in 2015 here when I was interviewing for a job. And in the end I got the offer, but for some reason, including some personal reasons, I accepted another offer at the interviewer conference. Next time I saw Mike was in front of the normal degree in Cambridge, where he was present as an external examiner. And when he saw me, he said, there you are, you rat. Well, this might not be the case. So that's not what happened there. And he not coming to Sussex and there would have been. I would have really liked that. It did not spoil our relationship in a way. It sort of evidenced that thought of which Eric Warren put up on the internet. As one he shot the last aspect of conflicts. Now to get actually the first connection to Mike's work and my own work. I use that as an excuse to show some comparative data. We've talked about when we have heard about optic flow or some of the talks, which is related to the fact that whenever an optimal system is moving through the world. And it's some somehow changing its attitude or translating and the whole world will be protected onto the eyes and will result in retinal images. The image shifts can then be depicted as individual local images here at different positions. We will actually project it on to the artist. So if you fly for instance here is rolling around it. The role axis as long as you do money access, then the whole world, so it's rolling in one direction, the whole world will be shifted in the long stretch. This is a very good presentation of the spherical field. And that is a connection between the playing. So we will look at a couple of these vector fields. I just thought I did an idea of what they actually show. So actually each of those local images is probably against a similar to that way. And the angle and are going to play an elevation that's the vertical angle individual space. And it's, I was limited in my view of what you can use vision for to civilization, which is basically something that involves some input that may be related to self motion or some other motion as well. So in this case really it is a motion that is generating optic flow, and this optic flow, whatever comes through the letter, each one of the system is then processed by what's going on later. That's the motion sets to cells that are responding to the patterns of. The information that should be effectively used for behavioral control needs to be passed on to what's in this case, white field descending neurons that basically connect the visual system to the motor system for us. And we can reach them to enable daily solution. Well, that's the motor neurons that are required in the muscle to activate, then there are some systems dynamics which means basically the active activation of the motor systems to the physical environment. And then as a result, you get some sort of little feedback again that is put through that control in that sense. And that's the simplest sort of stabilization. In and lose it. Now, in life that seems not to work. Many, many, many years ago, I was studying some of the neurons in the flight in the system is this kind of error. This is a long line. And it's all that some of these neurons, the logon plate invented cells here, the reflection of the harder than the cells are basically having huge receptive fields, which actually mimic or actually reflect some of the flow feels that are generally doing particular self emotions. And that for two different cell types or cell populations, the ages and the ourselves, and this is here the results of the ESL. And just to give you an idea and these vector maps again now show in which direction. You have to move something together, a strong response in those neurons, the length of in that this turns out how strong that response is. The neural reconstruction of these individual cells. Now that was some an interest there for me and that's actually when I started with us the first papers for my in how the distribution of those local deferred directions that you could call those individual vectors, all they relate to the less orientation in some point. And this is actually taken from a favor that might be with with a primary active I think was his name on the high geometry properties in, in this case it's a female blow from. And what you can see basically is the orientation changes of the so called horizontal and then they are some, according to the terminology that was used by now 1000, you have a vertical row in the I creator, and then you have two of the league roles, and why it makes row that result into something horizontal intermediate orientation as well. Well, here's how the roles are shifting. And this shows you what the density of those material roles is in the yard. Now what you can then do is you can basically work out, and all the orientations in a different process. Let's see is it's been written about. And then you draw on top of that orientation. And then you draw on top of that orientation. And you as you can see for this particular self is five and six. You have a very good correspondence as you have for some of the age cells. Now this is not true for all the different cells, and some which have higher sensitivity and other parts of the view, they may not match up as well as these groups. And for this, it's very clear. And it's very likely that the motion set to the elements and local elements feeding into the low level of the potential cells are all aligned in terms of input structure with the on to the eye lettuce. And you can think of the eye as some sort of original filter already. It's its general purpose, which is in this case, so it's a particular purpose, which is in this case, the analysis of the flow. Now it turned out something interesting about the sensitivity of these individual neurons to particular self motion components. This one here will be a 16 or literally responding actually during a role rotation. This one here to be as one would be responding best when the animal would be moving a pitch of notes of pitch movement, and that one isn't in intermediate between the role and pitch. So it's interesting, and they are actually 10 cells so you have 10 different axes that are represented. And it turns out if you look them as a function of a similar and elevation of these dots represent the best rotation axes the neurons are tuned to. Then there's a particular slide, well, and a very defined orientation of a plan that is set up. And this represents the represents actually different phases of what's called natural mode of motion. Some of you may have heard about that. So anything that flies or moves has has to obey to its own dynamics that produces particular modes of motion. So one of those modes of motion is a touch role as a phase shift movement where a role movement is coupled with a 90 degree phase shift with your rotation, and all together, it is some sort of swing oscillation problem with that is, it can get unstable. So it is a good idea to catch this mode, identify the activation of that mode, and then actually then this mode by motor action of combat by action of combat. And of course, there are other modes as well, which partially can come in favor. Now let me just illustrate that particular mode. This is actually a, that's wrong. Unfortunately, the camera perspective doesn't show you this really neat, repeated sort of oscillation through your and roll. Anyway, now that was the long introduction to something we are interested now. It still has something to do with the sort of modes of motion that we think the gender itself, in fact, related to or actually tuned to detect. And it took us and us is Graham Taylor Oxford and Sean Humbert at the University of Boulder now and it was a marionette as well to work out that in the end, the coordinate system they were set up by the local convention cells on the one hand side, but the coordinates is the new sort of control of flight are aligned. And actually, they are put together in a way to minimize the energy states to control the states of the offline. This is only shown for one animal and that is obviously kind of problem you have the most data in terms of electrophotology behavior system dynamics and basically the whole range that is required. But there's another animal that is flying in a very interesting way so my this is the bottom, we can actually see just the antennae and the legs to see the orientation. So we got interested in butterflies and I'm telling you I've only been working with flies. No, not true, but never butterflies. I think butterflies are amazing animals. Why? Well, a, they navigate. E, they have selection. Well, they have a very interesting television system. They have to select on what flowers they feed. And then of course, they need to stay stable in the air, which with this sort of wing design, the merge or fuse from the rear. This is not that easy. And you can see something here in this particular way already, which is quite amazing sound. You may wonder why it's so almost over shooting that picture. I think what's happening here is there's no visual structure around. The animal is trying to actually them from that pitch mode. It can't really be because the visual feedback is, is very low in that environment. In any case, so these neurons and these elements obviously have direction selective neurons. They have also the selling neurons very clearly, and that's exactly the target of our studies. But we took on this, this project and extending some work on polarization vision and in one insights. Now, let me see if I can show you. Okay, I think I'm, I leave it off. I'm not bothering too much with it. So, in the particular direction is asking very strongly responses, if you move in the opposite way, then you will have an individual of the signals. This was done in the tip of a butterfly. So what you do as a result. Oh, sorry. Yeah, so we can not only analyze directional tuning. We use these great ingredients for, but we can also look at the dependence of the response on the size of your stimulus of these options to bigger stimuli which makes sense in the process of processing, the bigger stimulus, the smaller response. We can also make the response dynamics in a sense by measuring the type of frequency during the new room. This one here pizza. Over 10 hertz, which is actually the fast level of data into the cells. So, in this experiment, we're looking at the activity in this one might be the same nuance. And stimulating the animals in front, there's actually some last three in local people directions, all these color coded orientations to local people directions. We have already quite interesting that may even be having information on that yet, again reflect the properties of the compound I have to check that. Now what you can then do from discriminating and one point against different positions along a similar. They are the results. You can do that up with motion, and you can do that have many different positions to obtain one of those vector maps and this one here shows you a very strong sensitivity to basically knows down pitch in this case upward sensitivity means the response during those down pitch. So you can do that in many different neuron types. These are samples and represent actually those colors here again for property sensitivity. We have some neurons that are sensitive to wide field, your rotation, more or less. We have some pitch sensitivity here. What's the most down. That's the most up pitch sensitivity in some intermediate as well. And with this data. I mentioned already that all these neurons, they will have three of the third rotation axis. And we can sort that out by running a particular algorithm to calculate what is the best rotational motion parameter that explains the receptive field. Let's go straight forward analysis. This is the result of that. And of course, you can say, we have basically two cells one in either side of the brain. And so we have, for each particular research to feel me, we obtain, we have two different rotation axes that you can do the same analysis for others as well. Now, for other cells as well. This is what we get basically for control rotation axes either side of the brain. These are the ones that are sensitive to vertical rotation of the axis. That looks a little bit more all the place of mind that this presentation. This line here is actually presenting the point in a sphere. So they are all very plastered. And then, with a version of orientation is basically a scandal and Sketch about what blow flies as we look at the distribution of axes we found in better fly designing you understand stuff. Maybe. No, We just edit the second set of very contralateral preferred protection access to that. So you can have the full picture of what all the protection accesses are represented in our data set so far. OK, now, behavior. We think we have a very good chance to ultimately, and at some point, we even are recording in fly handlers. The first step very clearly is observe the behavior of the animals and then analyze the trajectories which we can reconstruct in three dimensions. If we have that information, then of course, we can just say, OK, we have a certain rotation component in that, in that those trajectories will compute in a straightforward way the optic flow as it develops over time. Because we can see the individual trajectories in a result in the temporal domain. And we can see that that is part of one of those trajectories. And the corresponding input to the array is shown down there. That's actually quite useful already. Simply because we have this schematic data, we can actually do all sorts of analyses that will maybe or hopefully help us in which way in butterfly the neurons actually tune to particular modes of motion. Right. Now, one of the possible analyses is basically taking the cinematic data, running it through a principle component analysis. And no surprise, what you get is the strongest variance in the trajectories, this pitch component. And that's actually just simulated here. These are the axes that were coming out of the analysis. Basically, what you can then do is, and that's actually something I didn't do in the past, but just to give you an idea of how the response in these neurons would look like over time. So here's the optic flow which we can reconstruct from looking at the trajectory. Now, we take the optic flow here and project it into the receptive field of these individual neurons. And then we can look here on the y-axis, what is the result of the projection? So you would imagine that during this sort of pitching oscillation, the neuron will be alternated and basically stimulated and inhibited. And then there's a certain time of course that's actually representing the transition between moving upstream and moving downstream. Right, I think we have one more interesting that was almost toppling over, like example of the Dr. Price, but it again emphasizes that I suppose these two neurons from being information about pitch are pretty important for the animal in terms of flight control. Now, for using these normal pitch invention cells and then passing the information onto descending neurons and then riding the motor system, the input to the system or the orientation of your head should be stabilized. If it is rotating, then you would have to correct the input for that rotation. So in that animals, we find that in case of evasion or comprehensively high movements, a lot that at the front end, the system is trying to get rid of those disturbances. And to different levels or different degrees of course within the physical limits. Now, the other thing that's associated with at or gaze movements is we use emotion blur, what scanning we learned, that's exactly the opposite in the way if you want. And then of course, it has for insects that also have other sensor systems on the head. It makes it that easier to stay in line with all these sensor systems and the gravity vector if they have space or not anymore. So there are a couple of very good reasons for gaze and evasion. And here is a simple experiment you can do. You oscillate the thighs, you monitor what the head is doing and then you can plot the results of their compensatory head movement how well is the system performing as a function of frequency. You do a typical frequency response and end up with, you know, jump if this just shows that the information from the lower-gray to evasion cells into the more convenience is basically not changing in terms of motion sensitivity. So there is no transformation of axes to evade. I skip it just in the interest of time. This is a very nice thing, just showing you how basically the shortcut between the sensory and motor transformation is being established by the integration of local information or local motion information with these neurons and then the output of these local axes and evasion cells are aligned already or can actually run axes of the head which was studied by Stephen Houston in another couple of years ago. So from local coordinates to motor commands, it's taking place in the same, well, in a related or aligned coordinate system. Well, for the stabilization reflexes in flies, we have to consider that there's not just visual interaction but also mechanism information and important. So we apologize that the information to the system and they respond really fast. Well, if you remove the heart yes, the whole system becomes very small and actually may even become the feedback signals that are provided by the visual system may face in that wrong time and rather than stabilizing the gaze the gaze deviation from the hologram orientation maybe, etc. I really wanted to show you this. We did studies of many different species like five different species and we found that for most blood flies or for the blood flies and for the horse flies we had a typical blood class characteristic of the whole system. So you have comparatively good performance of low frequencies and when you get faster and faster that somebody rolls off. This one here is a hover fly. Well, one of the not exactly the species might not work and what if it's a hover fly and ultimately it's rotated at 25 plus that's more than 5,000 degrees per second and the faster you drive the animal or rotate the flies the better the performance. This is one of the most amazing things we have ever seen. When I saw it in the lab I thought, okay, well how tears must go how teny must go we switched on one after another system everything I knew could happen anything and it just stayed. Now when we look at that yeah just for you to understand what this is this is just the whole experiment with that sweeping stimulus so low frequency minus frequency and the minimum low frequencies again if the system would be performing perfectly well we would have a flat line because the path would always stay horizontal if it's getting worse then obviously the oscillation of the response becomes bigger and that is hardly any complication from this you can see already quite a big difference between the hover fly and the hover fly and a caliphora and it's even more remarkable if you look at the way the data are this is the dynamic of the stimulus so the oscillation frequency and this is actually the remaining slip speed on the eyes so how much of a mismatch is there now if there was no compensation whatsoever then that would be a linear relationship but when the sensor systems click in then this curve will go down and it is true actually with lower dynamic range on to certain learns and something like that all the three fly speed sheets they are doing better because they have the sensors that helps them now then you can actually remove the the halters and then all the curves will shift up and it gets worse here you can actually see that the responses are making things worse for the also flies and the slip speed is even though there is still visual feedback not everything but this that's the power flies and it's unbelievable with and without halters it is performing amazingly well so what we think is happening is that this is actually caused by inertial torsional spring if you want simulating the negative crystal so it's a inertial mechanism that is heavy enough so if you have a very fast rotation of the torsional spring it's a horizontal orientation the faster you're still you'll drive the system and in the lower dynamic range it is still using all the different mechanisms driven by vision and other mechanical systems right okay so I'm running out of time I'm very sorry about that I would have loved to at least talk about this which is absolutely amazingly built by Mike where we look when we drive at home well that's linked somehow to the really tightly relationship between motor action and sensor because one of the things that's in the gaze angle monitor the device that Tom has been explaining the gaze angle and the steering wheel angle are plotted on top of each other and you can see the high correlation of that so theoretically very simply knowing what signals you send to your eyes to your eye muscles to move the shift of gaze could be used as well to command the arms steering the wheel now that is that brings us to something I didn't mention earlier I was just staying here in the sort of inner room reflex control business but of course we have all these other things further, polarization, position distance, all things that are very important in the different behavioral contexts and the question is how can you control these actions or goal direct behaviors without being trapped by the inner loop control because he mentioned if you want to do something voluntarily and you stimulate by doing this your sensors, your sensors will generate a signal that's sent to the inner loop control and muscles and counteract it this is one of the things where we have now done a lot of theoretical work I have to jump as you will kill me but it gives us at least an idea of where we have evidence in the literature that there is an implementation of its forward model based control with other words you use basically an inference copy and you generate your own voluntary mode of command an inference copy that anticipates the sensor signal if you do a voluntary movement and that anticipated signal is subtracted from the sensor canceling the sensor response and avoiding the reflex and the civilization we participate that strategy and we have some evidence for that now it's of course very good for two reasons the first reason is you avoid any sensor systems driving into a separation subtract always whatever you think is going to happen if you perform a movement and that means at the same time you remain in sensitivity for anything that comes as an external disturbance now the other thing is due to that structure where we have a forward branch here we send motor commands into the system and here we subtract even if this subtraction is not perfect the nature of the medical feedback will still make sure you don't crash and all those are even maybe no features we don't want to end up with this anymore and I'm very sorry to skip in my summary but I have made a big mistake so we made a big mistake to the overwhelming and that's what it's like next up we made some work here related to the the butter slides and that was the comparative study on the gaze derivation and well many other people to say and this is my extended influence well obviously we thank him for what he has done for the field but I think the other and I think Simon has mentioned that already really good proof that there's no contradiction in being a nice person and an excellent scientist who will be an entirely community by and large thanks very much I hope I have to add that I didn't get to the end of that video but I've been made or they are more to it in the end one question the the motion processing and yeah exactly yes so you have basically in flies you have what's called the log plate which is part of the third of the from Europe and and in many other insects similar functionality like optic lobes that you would take place in your level but you also have a slightly earlier there's some motion processing that takes place not necessarily always but integration yeah thank you very much you have a beautiful clip on the fly show this unusual way you can wonder if there is possibly related to some of your late slides you have that integration scheme we do expect the visual system is kind of something at the time at the convenient time yes I mean this is yeah I mean I just dumped the results on you of that projection and the modulation of signals in those elements but what I think is happening there and that's the next step if you have these modulations and the modulations exist in more than one unit but they have different phase units if you have the disturbances that is actually pushing the activity the time-modulated activity in one of those into a different direction you could always imagine that you have a non-linear mechanism that is combining inputs and if you have a correlation of these signals at a high degree at a one and if it's actually decorrelated because an external disturbance you get a zero and then you take that information and modify your flight mode we don't know enough about the or I don't know enough about the control parameters in the flight mode in slides it's more straightforward we know that you have three parameters that's angle of attack on wind windbeat amplitude and pronation angle of the windbeat plane and we also know that there are many steering muscles which control the activity of the wing hinge and the way in which the steering muscles impact on the wing hinge and the inspection is dependent on the stage at which they generate an activity so that holds more or less some sort of clever integration mechanism that takes into account the permanent activity in the sensors okay