 Okay, okay great thank you and of course thanks for being here to listen to me rabid on about fish so today I'm going to be discussing sort of an inquiry into the mechanisms by which animals influence one another when they're in collective social groups or just groups generally and taking this from a sort of animal behaviorist and evolutionary perspective and so to begin with I'll just show you a video here this is a video that I took while snorkeling out the front of the Adriatico guest house in fact right under the sign that says you're not allowed to snorkel there which I saw coming out but it not only is a beautiful location but it also encapsulates here in this collective this group of fish one of the fundamental questions that we're all really interested in when studying collective systems which is the question of how one of the entities in this system in this group is influencing the probability of a behavioral output in any of the other entities and so a fundamental aspect of that question is is both are there universal rules by which all animals are interacting in a bang or are there deviations and differences amongst the rules that are being used by members of these groups and collectives and so that's the question that I'm going to be dealing with today asking the question if there's a behavioral output by entity a what is the effect it has on entity b and how is that effect realized and so looking at this question of therefore collective or social influence there are many ways this can be manifested on I'm going to deal with three of those today the first is that if there are different individuals within these groups or collectives they may be doing fundamentally different things the behaviors themselves are different and so their effect on a collective is therefore different alternatively different individuals may be doing the same things in the sense the kinematic signature of behavior for instance is conserved but these things have a different functional effect in the network in the group that they exist or indeed salience so there may be no observable difference in the behavioral output but the way the information is processed may be different and alternatively the different individuals are doing exactly the same things and these things have equivalent functional effects but they operate on different spatial scales and so therefore may manifest in different emergent or collective outcomes I'm going to look at that in three different contexts the first is what I will call a derived context which is an experimental lab based context in which we know very much about what we're putting into the system what we're trying to measure coming out of the system the second is a natural context in what in the wild in my case in Lake Tanganyika in Africa where we are not modifying the system very much but we're observing how the system responds to its its natural contexts and finally an evolutionary context which is a comparison across species when maybe the rules by which they have interact or by which they interact have evolved and changed over evolutionary time so that we can then interpret or interrogate these different modes of interaction so we'll start with these derived contexts and so this set of experiments I'm just making sure you can see you can focuses around this group of sick lids called a Stato tilapia Bertoni an emerging model in in social neuroscience a very useful system because it has a very strong social hierarchy social hierarchy that's manifested not only in behavior but in phenotypic markers of dominance and so dominant males in these systems are brightly colored they're aggressive they hold territories and they're reproductive subordinate males are not colored so they show none of these sort of morphological markers of dominance and they're non-aggressive and they don't hold territories but importantly males switch frequently between these two states and so you have the ability to interrogate the behavior of a male both in the dominant state and in the subordinate state and if you create a social network an interaction network in this case of these systems you see that the dominant males are very central and important nodes they're highly connected nodes in these networks that's because they're attacking other males they're fighting their displaying their performing courtship to females and then you have the subordinate males out here that are peripheral nodes in these networks in some senses you may consider these unimportant nodes and so we can then ask the question if we place information or some kind of seed into one of these networks we can then ask the question is there a difference in the way this information propagates through these networks based on the the dominance position of the seed of that information and this gets us to the question of how are the the different types of males in this case the dominant and subordinate male influencing the behavior of others in their network and and altering emergent outcomes at the level of the group very similar to what we saw in Lucy Applin's talk yesterday and so the basic premise here is because you want to keep it very controlled is we don't want these animals doing a range of ecological tasks we want them to do one thing and in that case it's a simple association learning task where they learn to discriminate between a yellow and the blue lights which indicates food and so there's no initial response importantly from these animals they do have a sensory bias towards red seemingly not yellow and so they show no initial response but over time they come to learn that one of these colors is positively rewarded which is indicated by a color change here that you can't see but then we take these individuals at different social ranks so again here's the dominant and the subordinate males we also included females in this design but they turns out they have quite a strong hierarchy within the females that we cannot interrogate easily so I won't talk about too much on the female side today then we take these individuals that now know this association and we put them into a new group of naive fish and ask the question how quickly does this group of naive fish come to learn this association task given that they now have a social demonstrator of different ranks otherwise these groups are equivalent except for the social rank of the demonstrator and this was done when I was a poor postdoc before I came to the Max Planck and Ian Cousins department and so it looks a bit ghetto but you'll have to excuse me here you're seeing two lights come on and here is the informed individual so this fish knows the association and you see that the other fish are afraid of the of the condition and don't show a response the group then comes to generalize about the light meaning food but there's no consensus within the group so they don't know which color and then finally we see a very strong outcome at the level of the group where they all know this association very very well and so the basic question is how quickly do each of these groups come to this informed state where they're all moving towards the queue prior to the reward being dispensed now if we were being naive or I was to present you with a strawman argument which I never would really stand behind you might argue that information should travel along these network edges as a function of their edge weight right so the the dominant males would be the best informants because they're the best connected within their groups certainly in in the structure of this behavioral network and so we can then go and ask that question and ask how quickly that each of these groups learn and so this baseline rate is a naive group so they learn in about 20 exposures to the stimulus and what you're seeing is that there's evidence of social facilitation here so a demonstrator in the group does increase the rate at which the group learns this association interestingly though here is the dominant male the line of the groups that had a dominant male with the information and although there is a significant increase in learning the important aspect is that it's the subordinate males or the groups with subordinate male demonstrators or or at least if it's not an active demonstrator role at least with the information seated at the demonstrator that the subordinate that learned far quicker and so the question then is how could this be true when these subordinate males are so peripheral in their networks well the answer of course is probably that we've got the wrong network right we have a behavioral interaction network which is dominated by aggressive displays and aggressive interactions and it's unlikely to be the same kind of network over which information transfers effectively and so this can be relatively easily seen here so I'm not going to tell you which animal here is the dominant animal but I think it may become quite apparent quickly which animal is the one that's chasing and and holding the territory and if it's not clear it's this fish number eight and so if we just look at this video here it's quite obvious in a sense as to why this might be a very poor effect or of change at least in this association learning task in this group two things to notice of course one is that this animal is chasing all of the other fish the other is that generally speaking it seems to be further away from them and and less connected in other kinds of networks and so then we can interrogate those other kind of networks and so here is a spatial connectivity network which just describes how close these dominant individuals are to the rest of the group and what you're seeing here is the the dominant male and so it's the color of the lines indicates a further distance away but that's better represented here and so you're seeing that the average into individual distance between the dominant male and all other individuals in the group is far greater okay so although it's behaviorally better connected spatially it's very poorly connected if we then also look at a visual connectivity network how much each fish can see of each other fish we see also that the dominant male this is now a different group where the dominant is number 10 is very poorly connected as well and so not only are they further away from this individual but they're seeing it less and so there's far less opportunity for the visual information of this rapid movement towards the the stimulus to be recognized and and perhaps encoded behavioral change finally what's also important is if you look here at the response to the to the learning task you see it's characterized behaviorally by this rapid swim towards the light yeah before the food is dispensed now compare that behavior to the behavior of a dominant male sorry never work with children or animals okay we see that the dominant male is frequently performing behaviorally similar similar cues right so it's dashing around the tank all the time and so there's a potential that the signal to noise ratio is also different for dominant males and subordinate males subordinate males are never swimming around at this rapid pace and if they do it's probably a very unusual behavior for them and therefore stands out and we can look at that numerically and so here's the average speed at which they approach the light stimulus and here in red are all the times the dominant male in this group made these acceleration bursts that were faster or around as fast as those dashes towards the light cue and so you see very many instances of of this behavior and then if we compare that to a subordinate male here in blue we see only four instances of a similar behavior and also that all of those instances were coincident with being chased in fact by the dominant male they weren't spontaneous they were a response to the dominant males behavior and so here we're seeing of course that the the the the effect that these dominant males have on the behavioral change in their group members is very different to subordinate males as a function of the kinds of behaviors and the scale over which they perform these behaviors we can actually go in and validate also the salience of these cues by looking at the neural encoding of the social stimuli asking the question for instance in the in the areas of the brain that process fear in the fish in this case the basal lateral amygdala homologue we can ask is there elevated activity when they're seeing these kind of behavioral cues coming from subordinate or dominant males and the answer is we don't know but we're trying to get to that point we do have data showing that of course when they learn the cue they have lower activity in in these fear processing regions and other parts of the brain as well but it's not clear if this is a cause or an effect of this difference in learning rates okay and so then we can now move into natural contexts because of course in their in their natural environment in the in the conditions in which they've evolved fish are not learning or are not tuned to do these kind of discrimination tasks they have a very different set of rules the four F's of biology as we talk about it in undergraduate course fighting fleeing from predators feeding and reproduction and so we're going to now look at this system here which is Lake Tanganyac and cichlid called a lamprologous calypterus and a very interesting system and so this is wild footage and this fish employs a collective foraging strategy by which it overwhelms the defenses of territory holders so here you're seeing one territory holder attacking these groups as they come in but by sheer force of numbers they overwhelm that territory defense and are able to feed inside these territories yeah and so they they transition between three more or less behavioral states these are juvenile animals so they're not doing any reproduction and they're not defending territory so they're not really fighting very much so their whole mode of life at this life history stage boils down to to avoiding predators and finding food yeah and so if you watch this video you'll see these behavioral states relatively clearly here of course we're feeding and getting attacked by predators then they will enter into a movement phase which is driven by some proportion of individuals starting to to to move then again they'll enter this feeding phase which is characterized by this disordered unpolarized state where they're making small movements and then it again transitions into this this directed movement phase here where they're off now looking for something else to do and so using this system we can go in and ask and and look at the both the collective level behaviors and also the the behaviors of the individuals and ask what the link is between these these states and the transitions amongst these states as a function of what the individuals within these groups are doing also of course as a function of what's happening externally one thing I do want to point out here is that in in a movement like this if you watch this this group for about 20 or 30 minutes you're gonna get about one or two of these predator or territory holder attacks on the system every minute so it's a very high density of of of interactions and and occurrences of very interesting behavior in these natural systems now of course tracking in these natural systems is very difficult well let's just summarize what I just said which is that we have these three broad states swimming which is a very poor sort of behavioral classifier a recent paper has just shown there are with a machine learning approach 13 modes of swimming in zebrafish larvae we can also have fleeing or being attacked as I as I demonstrated in them in the video oops and then these feeding swarms now of course this is not an exhaustive list of the behaviors and that's kind of the point we want to characterize what behaviors there are and how the transitions amongst these behaviors are manifested okay but but on to tracking of course in in natural systems it can be very challenging to track these individuals but that's been that the challenge and the goal we've set ourselves and so here we have one of these groups in the wild what we do is we catch an entire group of about 200 individuals we have fixed these barcodes to them which allow us to track every single individual and also maintain their identity or at least that was the idea and then with camera arrays above them as we swim with scuba with an array of 10 or 12 go pros above them filming these very large schools of up to 200 individuals we can track everything that's happening amongst them now interestingly initially we did design this this experiment with these barcodes in mind but you know this is the middle of Africa and things go wrong and as it turns out we couldn't reliably detect these barcodes in the same way that we intended which was with some software developed by Jake Graving one of Ian Cousins PhD students so instead we trained machine learning models to recognize these tags and even to pull out the IDs from the information encoded on those tags but what we're ultimately moving towards is entirely untagged systems so here's some experiments of mine with spiders showing our ability to use these neural net approaches to detect individuals in very very complex visual scenes so this is what we're moving towards as well in the fish entirely unmanipulated approaches to tracking wild animals from these tracks within use behavioral decomposition approaches and I won't go too far into these of course because I think we're all relatively familiar with them to to create maps of the entire behavioral space of all the individuals within these groups and and of course the group state itself but just to briefly discuss that we we look at these tracks they take through the world we take points within these tracks which are parameterized in highly dimensional space depending on what kinds of behaviors were feeding into the model and also the time windows we're using to populate those and then we ask the question how similar are some areas of some points within this behavioral track to others and the idea being that you should get these clusters of different behavioral states in an unsupervised and agnostic manner now of course we also use supervised approaches for this machine learning but what I'll talk about here is unsupervised stuff and the kinds of parameters we feed in our kinematic signatures based on the motion and and well the behavioral motion of each animal and so here you're seeing quite a clear signature in the green fish here of that characteristic driving display that it did and so the point is here that we can really pull out with with some of these parameters very very powerful signatures of the kinds of behaviors that they're performing okay and so then we can feed these parameters into these models here we have five parameters average speed average positive acceleration average negative acceleration absolute turning angle and trajectory directedness just a measure of the path tortuosity and here there we're seeing how well they contribute to each one of these these classes and so here we're using dimensionality reduction with one dimensional t-sne space many of you are probably familiar with the two dimensional t-sne space this is an argument that I had nicely the other day with Gonzalo about which is the best approach here and I'm open to all of these discussions this being a very new and emerging way of understanding behavior so then we can take these these behavioral categories that the model spits out and we can map them on to the track of the entire motion of these groups as they move through the world and then start to ask are these a reasonable behavioral categories do they map to something we can observe and understand and quantify and be asked about relationships between the behaviors of one individual and the group and so as a ground truth we can identify already some quite apparent behavioral states which are easy to interpret so here we have one of these feeding swarms which is characterized again by low velocity by low degree of polarization and short stochastic sort of movements and then we have these these times when the group is attacked and so they're going from one of these relatively normal or relaxed swimming modes into an increase in acceleration and directedness as they're attacked by one of these predators or territory holders from the rear of the group so they just burst away and he's sort of what it looks like one of these attacks this is here we have you'll see bang an attack on the group and so here is another kind of behavior that also pops out which one of great interest is it can we quantify where these attacks are occurring how frequently just based on these automated approaches and for instance from where the network was attacked which node was the initial responder and how did that transfer through the rest of the group and so if we look at that well sorry no I'll go back then we can overlay these these these color encoded behavioral states on the group as it moves through the world with the goal ultimately to overlay those on on the videos and attempt to marry this sort of unsupervised approach this this statistical breakdown of behavior with behaviors that we can observe and understand and I think this is the great tension of using this kind of approach to quantify behavior because ultimately it needs a ground truth but if your intention is to remove subjectivity from your assessment of behavior it becomes sort of chicken and egg problem which way around you perform this process but going back to this this predator attack we see here quite an interesting phenomenon where there is this deviation by one individual or some subset of the group from the otherwise overriding group state and I think these are very very interesting phenomena and we have of course many many instances of these and so this for instance could be an attack on a subset of individuals maybe they entered a territory here and the territory holder attacked these individuals which had to take a deviation in the path or of course it could be the avoidance of an obstacle and this is another aspect to this whole process this is occurring in the real world where there are things there's terrain there's things to to navigate and and avoid and of course we can also quantify those things because as we move these camera arrays through the world we're getting a lot of information on the structure of the terrain just as you saw in in Ari's talk and using this approach structure through motion we can then reconstruct the entire terrain through which these animals moved and start to ask questions not only about how they're interacting with one another and how they're interacting with other hetra specific species but also how they're interacting with the world that they have to navigate and we can reconstruct the camera path sorry about the low frame rate it's the cost of rendering these things and for each one of these behavioral traces that we get they're typically about 45 minutes long we can reconstruct the entire path typically we're getting between 250 and 300 meter sort of path lengths and the really nice thing is that the the paths often overlap so that we have a sort of ground truth so that we know where we are in space the idea is to combine this with GPS markers which is not currently achievable under water so we can reconstruct entire sections of this of this terrain in which these animals are living this gives us really really nice ability to then interrogate how they're dealing with environmental obstacles so here you see a particularly tall rock and you see a decision point at the level of the group here where some of them went around this way and the rest of them went around this way and of course this is happening frequently but I'd also like to point out especially in the context of the the talk we saw on the bird flapping yesterday is here we get a similar ability to measure individual tail beats by the by the tortuosity of the path here and so you're getting evidence of how many tail beats they're doing the frequency and can compare that to the speed for instance when they're escaping a predator or making a different behavioral decision okay and so now to take it into the evolutionary context and to bring it back to this point so we may have alternative mechanisms for differences in social influence and here if we're asking questions of a system in which we know that there are differences in the mode of living we know that there are differences in the degree of sociality we know that there are differences in the collective structure of these different species we can then go in and ask the question over the time of evolution that has changed these species which one of these aspects has changed yeah so we can go in and interrogate the basic behavioral repertoire that these animals have to ask has that changed and that leads to different emergent collective outcomes again to remind you we could we could then embedded in a network to ask the question if the behaviors are the same is it that they differently transmit through the network and have a different effect on the rest of the group or is everything the same but it's just occurring at different spatial scales and so to answer this question again we look to the cichlids and this is a particularly special group of of cichlids in Lake Tanganyika the lamprologian cichlids obviously if you know anything about cichlids which you may not of course they're famous for this this explosive adaptive radiation so we have over a thousand species in the African rift lakes or from a known founder stock of of a handful of species that have since diverged into a wide variety of modes of life and so this particular group comprises about 40 species they are all very very small so very amenable to to experimentation and and field manipulation they all eat the same things the micro benthavors the same things eat them they all live in snail shells they're all about the same size they all look roughly the same but they vary dramatically in their social structure so some of them live entirely solitary except for the breeding condition and some of them live in permanent social groups with our parental care helpers at nest and all of the sort of classic hallmarks of sociality and so the way these animals interact is fundamentally different yet everything else about their ecology and life history is relatively conserved and so that we can then go in and say given that there are these differences amongst how they interact what is actually different about them and here in in in sort of distinction to the stuff I showed you earlier about only the kinematic signatures of of these swimming schools in the wild here we're very much interested in the interactions that are occurring between these animals and so here we have three traces that describe this behavioral interaction now from a traditional ethological perspective we would look at this interaction and we would characterize this as an aggressive act on the part of the green fish towards the orange fish and that might be the end of it and here we could we could quantify that here we have that if you can see it I apologize it's it's small and faint here we have the speed of the two animals so we have a rapid increase in the speed of the green animal here we have the relative orientation of the two animals so when it's zero one animal is facing the other and here we have the inter individual distance between them and so using supervised methods as I said and unsupervised methods we can break down all of these different types of behavior and ask the question are these behaviors different amongst these different species importantly here we have one of the social species and what you'll notice is of course in the in the attack this orange fish is not leaving it's staying in the territory it's performing what we probably would interpret as an affiliative display here a sort of peacekeeping display and it's allowed to stay which we very different we would hypothesize to the non-social species in which this interaction would lead with lead into this fish vacating the area entirely now leaning on the the the long history of study of this particular group of animals we have a broad range of approaches to then interrogate this this question about how the behaviors that they display have evolved we can take traditional ethogram methods where people have categorized behavior into into these categories these these distinct behavioral states which is of course potentially hazardous because this is probably much more of a continuum of states nevertheless we have a great deal of evidence for and literature supporting this kind of approach we can take this and combine it with of these some of these objective approaches like t-sneed to create behavioral categories and let the data tell the story about what kind of behavioral categories there are and and ask how they match with these kind of approaches and then if we've constructed these these maps of the behavior we can go in and ask are these behavioral maps different amongst the different species using exactly the same parameters using exactly the same thresholding and and and division amongst states we can then ask just as Gordon Berman has done with different species of Drosophila are the behavioral repertoires that these animals have different and is that's what's causing the difference in their collective function at the level of the group we can then attempt if we want to ground truth some of these behavioral maps in the very very exhaustive ethograms that are produced by many of the groups around the world and I think this really does encapsulate the the challenge to this kind of question in the modern era because here we have an ethogram from one of the groups that studies one of these species and it contains about 35 behaviors if we look at another group that studies this species they'll publish an ethogram with about 20 behaviors and then you have someone like me coming along and showing a behavioral space or a behavioral map here which has eight categories and the real challenge and the question is how do we reconcile these different approaches to asking the question how different are behaviors this becomes very very important when we're doing comparative work to ask how the behavioral interactions amongst species have evolved because if we're not dealing with the same template we have a very very difficult task in comparative work now of course how I choose to create these different categories is subject to debate and argument but at least I think in this case we're dealing with a common set of data and that other people can interrogate and interpret and re-evaluate and re-analyze versus this kind of approach where we're going under water with a clipboard and a pencil and we're putting check marks near the different behavioral categories with no way of really comparing them post-hoc ah that's what I wanted to say about that one if I'm able to oh dear sorry interestingly when we do train models in a supervised way if we train you can't see it here but this is the the aggressive chase category if we train one of these models on one of the species in the chase category and then we apply that predictor to one of the other species which is identical in size and and everything I said it fails so there's something about the kinematics it seems that means that the predictor generated in one species does not work in another species perhaps suggesting that there are fundamental differences in the kinematic structure already um or the effect that these things have amongst these different species okay um now of course one of the things you can do is create behavioral repertoires as we saw in the previous example based on the kinematics of the animals alone but these are social animals living in collectives so this doesn't really make sense uh conceptually because we're interested in the interactions and so the the subsequent step and one thing we're we're now incorporating into these analyses is to not only include individual level characteristics but also these sort of Markov chain approaches um this is a paper um by Jens Krauser who's with us here um looking at the probability of a behavioral shift and then perhaps embedding those in natural social networks to ask not just what the animal is doing but how does what the animal is doing affect what the other animals around it are doing getting to this fundamental question of these rules of interaction and these rules of influence amongst these different species okay and the final aspect that I mentioned was this sort of structural or this spatial question because of course when we're comparing these different species that live naturally at different scales and asking what's different about them it may be that the answer is already there it's simply that they live at different scales and if you forced them to live together you would see that they immediately update their behavioral rules such that they no longer show these differences in behavior because everything about them converges and so this is what we tried to ask with this experiment so we forced these different species to live in this standard sort of spatial world where all of the different shells are equally spaced and so their interaction rate and their spatial relationships are all now identical and asked do the behaviors remain different or do the behaviors converge on the same behavioral set now of course this is a lesson in the fact that your study system can be smarter than you are because instead of accepting this condition what they did was rearrange sorry about this the architecture of the tank so here we have the social species and instead of leaving the shells where they were they dug out the shell so that they fell together and then lived in these little couplets the asocial species did the opposite in which it chose one shell and then buried all of the other shells around it in the tank so no fish could come anywhere near it I think this is a really beautiful result demonstrating this relationship between the spatial structure that the animal produces and is given and the behavior that it produces and of course the fact that this is a reciprocal relationship we asked the same question in the field and so here we're mapping out the sort of interaction space amongst all these individuals by by mapping where they are in space in three-dimensional space and then characterizing their their their home ranges also looking at how they deal with individual shells and then of course how the groups and the networks that they produce are arranged in space okay so I'll finish there and take your questions thank you