 Hi, thank you first. I wanted to thank the organizers for having me giving me the opportunity to speak to today About the work that I've been doing in Ian Cousins lab at the Max Planck Institute for ornithology and constants Like many of you here, I'm sure We're all quite excited about collective behavior and I'm particularly interested in those those sort of collective faculties that come from Interactions between individuals that are not Ex that are not exhibited by individuals themselves Perhaps the most interesting of these are the idea of collective perception. For example, we know that Collectives of fish can perceive a large gradient. That's well beyond the scope of their own environment, but through individual detection and Interaction among each other. They're able to perceive a gradient for example So the idea is so if I want to think about collective perception Decision-making and cognition Well as a neuroscientist, this makes me think about a very large field within neuroscience and that is Individual collective individual perception decision-making and cognition. And so what I want to tell you about today is how my My efforts to to adapt so a very popular Experimental technique from neuroscience to collective behavior That task is this one It's called the dot motion task for perceptual binary decisions It's been in use for about the last 30 years and it's been a very very lucrative task for studying decision-making The task works like this if I set some motion on these dots with some degree of coherence I could I could ask you now What direction is the majority of the motion so maybe show of hands who says the dots are moving to the left? Who says that to the right? Okay Good so far. So even though there is some component of random motion in here You can you may be able to see that there's not all the dots are moving to the right This was a pretty easy task. However, if we change the ratio of Coherent to incoherent motion the task becomes a bit more difficult And so if I ask you the same question again who says the motion is to the left, okay? Who says it's to the right? Who's not sure? Yeah, so clearly much more difficult with with a low coherence low information Congratulations to those of you who said it's to the right you've your Perceptual decision-making is about on par with the most highly trained monkeys This task has been adapted to lots of different organisms over the last 30 years. Whoop No, my video is gonna play So here on the left is a rat performing the task here on the right is a human I think and Generally the goal within neuroscience is to record neural activity and try to understand what neural mechanisms underlie perceptual decision-making Why are all of these coming up at once? Okay So here for example is an fMRI study done in humans during a perceptual decision-making task where they identify a region of the brain that seems to Represent the accumulation of evidence and they find that with higher coherence the evidence accumulates faster Similarly in rats using olfactory mixtures, you can identify neurons whose activity correlates more with the confidence in a decision also by mixing acoustic signals you can identify neurons Whose activity records more with the the readiness to make a decision than then the evidence itself? So my goal is to adapt this very lucrative system for studying decision-making to animal collectives primarily with the goal of using a very well-designed experimental procedure to study to learn something about collective decisions in collective decisions and my secondary goal is to maybe be able to learn something in general about a Distributed information processing that might inform something about decision-making in neuroscience So how do I want to do this? Well, I'll take advantage of the fact that fish schools respond to moving contrast which I hope you can see oh god So here's about a thousand sun bleaks swimming in a tank And I hope you can see that there are black dots projected on the floor of the tank And that the dots are moving in the same direction now looking from above you can see that the the swarm of school of fish is spinning in the same direction of the dots and even as I change the direction of the stimulus The animals respond quite quickly and changing their direction as well. This is extremely robust It works nearly 100% of the time with large groups small groups large tanks small tanks It's a it's an extremely robust response And this forms the basis for how I will adapt the decision-making task to fish The responses may be easier to see if I here now where we record videos in infrared infrared and we block the visual Stimulus from the the video so that we just see the activity of the fish And here in a second you'll see the stimulus come on indicated by an arrow And what you'll see that when the stimulus is congruent with the the milling direction Of course, they continue milling in that direction and then as it switches you'll see there'll be a moment of chaos sort of it seems That like the system becomes a bit unstable until it switches and the motion is then congruent with the stimulus So the way I'm thinking about it now is that this milling direction Represents a sort of by stable state of the system that it can go in one direction or the other and the stimulus is adding Influence to go in in one direction or the other And also in a second, I think you'll see the stimulus will stop Yeah, and you'll see the milling direction persists So at least to some extent the the information that was presented previously is still represented in the current state of The by by the milling direction. This may be analogous to something like hysteresis or working memory So we can track the fish using a computer Vision software developed by a graduate student in the lab Tristan Walter And Tristan has really optimized the system to work really quickly so that we can collect lots of data with very large group sizes And not be bogged down with with a long processing time that a lot of systems require So we get data that looks something like this we can quantify the rotation direction by just Calculating the rotational order parameter, but keeping the direction So here when presented with clockwise and counterclockwise rotation, you see that the order parameter Goes from zero to one to negative one in response to the stimulus So this lays the framework for the adaptation of the random dot motion task to fish schools We simply project patterns under the floor of the tank record them with camera and we can Present a school with a very easy decision to make or relatively difficult one because there's So many parameters in the system we developed a high throughput system to parallelize the data collection We have seven identical tanks set up with projectors and cameras and so on and we've developed a data analysis pipeline that that gives us a trajectory data Synchronized with stimuli and stimulus information so that we can associate the response of the of the collectives to the presentation of the stimulus So I'm still just in the early stages of analyzing the data that I've collected so far And for now, I'm just going to do a very Sort of crude representation of the behavior by thresholding their behavior in two different directions One is a threshold on the polarization parameter. So high polarization. It gives us a polarized state And then when polarization is low we give two thresholds one representing when the Rotation order is high we get congruent milling when it's low at negative one we get incongruent milling This is direct directed in the opposite direction of the stimulus and then in the middle is the swarm state So now the task is quite quite simple. We just present different stimuli with different coherences to to fish schools And for now, I'll just just show you Three levels of coherent coherence low moderate and high these represent three difficulties high moderate and low Here's what the data typically look like. Oh There is a gray bar here representing what the stimulus is it starts at about here and it ends about here So you'll see initially there's a swarm state in no matter what coherence the level the stimulus is This is likely just a starter response then the polarization drops because no matter what the coherence level there is generally some Milling response and then what's interesting then is the two different milling responses either in the congruent direction or incongruent direction So first looking at the green. This is the low coherence level. You'll see that it's about the same in in both congruent and incongruent but then if you look at the at the high coherent Experiments and there's an increase in congruent milling and a decrease in incongruent milling so I think this is Promising if we're if we're trying to adapt the random dot motion task to to fish This tells us that a group of fish can indeed Perform a task and they perform it better when it's easy and they perform it poorly when it's difficult But is this collective? Well, if we vary the group size we see that the performance drastically increases With larger groups. So here. I've varied the group size from eight all the way up to 128 And you'll see that as a group size increases the amplitude of the response increases particularly for the high and moderate Coherence levels that's maybe easier to see if I show you just this is the the mean success rate during the last half of the stimulus the last 90 seconds and So here in green is when the coherence is low There's very little information presented and no matter what the group size they generally Have a success rate of about 50 percent. This means that they're milling it at random When the coherence is high for group sizes above With 16 and above it seems that they perform maximally Suggesting that even 16 fish is sufficient to detect a fairly strong signal in the in the system And then for the more moderate levels It seems that the smaller group sizes perform very poorly But it seems that as the group size increases the they start to get better This is still ongoing work, and I'm in the next couple of months I'll be expanding this plot in that direction all the way out to 4,096 Going in powers of two And also I'm working on more detailed data analysis looking at local interactions trying to find for example Points of divergence or or measuring the stability of the of the school Before and after the correct or or error responses So for the last part of my talk, I want to introduce something that is Haven't started yet. It's a Collaboration that I'm excited about and that is getting at the question of what social information is necessary for collective cognition so In the experiments. I've shown you so far. It's very difficult to tease apart What information is social and what is a social in the collective response to to the Moving dots paradigm, but if we move to a non-living system or a completely a social system We can start to tease these things apart. So I'm quite excited about a collaboration with a Cecilia Lozano. She's a postdoc in the Bellinger group at the University of Constance And she's working with these synthetic micro swimmers These are asymmetrically coated particles that respond to light and will move Down a light gradient With a quite an interesting relationship between the velocity and the intensity of light so here as the intensity increases Velocity increases up until a point where actually then At very high intensities you actually get a positive phototaxis in the opposite direction So what you can do with these micro swimmers is if you give them a light gradient You expect them all to go in one direction from high to low as long as you stay within sort of this region So you can give successive Gradients in like a sawtooth pattern You can't see those dots at all. Can you? Well, they're all moving to the right. This is probably the most difficult task this time And so you think we can drive these particles continuously by giving repeated sawtooth patterns similar to what we did with the The fish task if you orient the gradients in a circle you can At least theoretically get continuous responses from the particles Then of course we can in theory flip some of the gradients to change the coherence of the stimulus and And and ask whether the the group of particles as a whole will will come to the correct solution of the gradient direction So now we have I've shown you two methods for steering collective decisions one and fish where we just present them visual stimuli one using these Microswimming particles where we can moderate the coherence by changing the the orientation of grids of gradients on a grid And we've come up with a strategy Sorry Here in fish this is of course Taking advantage of social information And this is a completely a social system. So we would expect that the fish will perform this task In a far superior way to these non-living a social particles However, if we can mimic some aspects of social interaction in this non-living system We can determine what what components of social interaction are essential for improving the Performance in this task. So we've come up with a way of how to do that and that is well, if I just quickly remind you of something probably no one in the room needs reminding of the a Common model for social interaction in animals are these zones of attraction orientation and repulsion if we If we project in closed loop a gradient around each particle in the system That gives some information about its motion. So perhaps it's elongated in the direction of its motion Or maybe its gradient represents the amplitude of its velocity We can also reproduce these zones of attraction alignment and repulsion as Particles come close to one another. So as as a one particle approaches another they would attract and then At this point of alignment to the switch to negative phototaxis and then you get alignment and then collisions would represent the repulsion So using this we can test hypotheses about what's the minimum social information that's necessary to To improve the performance of a detector in the of a collective detector in a noisy system Not only that if we do repeated experiments and Reward successful detectors and allow them allow them these parameters to vary on their own in a Sort of evolutionary loop we can ask how collective systems might evolve Their behavior to to improve their their collective perception and cognition, so I'll stop there and First thank everyone in the Department of Collective Behavior And especially those listed here who contributed directly to work and I'll take your questions. Thank you