Seminar Series 2026: 1. Dr. Anna Dornhaus

Ajouté : 2026-05-06

[00:07:22]Hey, hey, hey, hey.

[00:07:34]Hey, hey, hey.

[00:08:31]Nat.

[00:10:52]Heat. Heat.

[00:12:52]Hey. Hey. Hey.

[00:13:26]Can you hear me?

[00:14:27]Ah, I need a Hello. Morning.

[00:28:46]Dr. Dornhouse completed her PhD in zoologology at the University of Worsburg. She was then an associate researcher in the department of artificial intelligence at the University of Wsburg and a visiting fellow in the school of biological sciences at the University of Bristol.

[00:29:03]Since joining the University of Arizona in 2005, she has served as an assistant professor, associate professor, and is currently professor in the department of ecology and evolutionary biology. Her research focuses on how collective behaviors emerge from the actions and interactions of individual using social insect colonies, including ants, honeybees, and bumblebees as model systems.

[00:29:30]through laboratory and field studies as well as mathematical and individual based modeling. Her work continues to investigate coordination in foraging, collective decision making, task allocation, division of labor, communication, and the evolution of cognition in complex social systems.

[00:29:49]Today she will give a fascinating talk about what we learned about the world from studying collective behavior, algorithms, replicators and strong inference.

[00:30:00]Please welcome uh Dr. Anna Dornhouse and thank you for giving the talk for us.

[00:30:08]>> Thank you for that introduction Pro Bayan. Um yes I think um I don't know if my talk is going to be a little different than the normal research talk but I really um was struck by the fact that I think it's important to think about what are the general principles we extract from our research on you know specific aspects of biology or animal behavior. And I hope the title will make sense as I um start the talk. So um just a brief introduction. I work on social insects um ants and bees like these little teenorax ants here and the bumblebees. We often mark individuals to recognize um each individual um and then follow them around in the colony. And that's a picture of um my current lab up there.

[00:30:53]Now what I want to focus on is these social insects and the fact that they are um a collective. They live in colonies. So there are lots of individuals that are somehow working together um to do the work of the colony.

[00:31:07]And in this uh social insects resemble a lot of other systems that we may think about both in biology and outside of it.

[00:31:15]So for example, this is a cartoon drawing of a cell. Um this is a simple bacterial cell, but nonetheless you could see how the cell really is a um is a container of lots of individual um little machines if you will, enzymes, um proteins, other molecules and all these molecules somehow uh once you put them together um create a larger organism which is the cell and uh enable the cell to perform various functions of life through their interactions.

[00:31:47]And similarly when we go to multisellular organisms uh this is a volvox algae which is uh essentially a colony of lots of cells um again we have the same pattern of lots of individuals coming together to perform some collective function. So this is true in the evolution of multisellularity.

[00:32:06]Here's a picture of an embryo in which you very much can see how all these cells through their interactions are differentiating into different functions um and then producing the function of the organism uh in embriional development. Of course, we see um oh wanted I just wanted to get rid of this down um that so in the brain for example we again have lots of um different individual cells that are interacting with each other.

[00:32:40]Um and so it is in many other systems even engineered ones like collective robotics or uh cluster computing or here this is a a graphic um depicting the power grid in the US. So all of these engineered systems also are showing collective behavior and that there are lots of individual units that are interacting to produce some larger function. And of course in some sense this is also true in uh modern what we call AI or um large language models uh which this is sort of an abstract depiction but they are neural nets um which in some analogy to nervous systems consist of lots of individual units interacting to produce an interesting collective behavior.

[00:33:28]And not least uh we also see collective behavior in humans both in small teams in commercial operations and in large commercial operations. This is uh an iPhone factory where lots of people are working on an assembly line. And in all these organizations we also have multiple people contributing their own parts to a larger hole.

[00:33:52]Okay. So these are all systems that display collective behavior.

[00:33:57]And what I would argue is that all of these systems benefit from an understanding of these three fairly basic concepts. So let me talk about what I'm um I'm going to first talk about algorithms.

[00:34:09]And here I want to switch to um talking about ants. And I want to give an example that's um actually not from my work but from uh Nigel Frank's work. So um these are army ants. Um many of you have probably heard of them. They live in the tropics. They have very large colonies. They're essentially nomadic and um the feature of them that I want to talk about right now is that they have these large raids which are basically uh ant roads if you want to call them that with a lot of traffic in both directions. Some individuals are going out to um hunt uh other small insects or big ones uh and others are bringing back the loot that they've collected.

[00:34:52]And in these um you know traffic uh roads almost you could call them um ants are walking in both directions not in a chaotic mixup way but they actually form distinct traffic lanes. And these distinct traffic lanes enable the traffic to go more smoothly because there are less head-on collisions and ants can walk faster.

[00:35:15]But how do they achieve these traffic lanes? We don't know that the ants know that they're making traffic lanes. What we know is that all they do is if they're carrying something, if they are carrying prey back to the nest, they have a a stronger tendency to just keep going even if another ant is coming is is facing them coming their way. Um but when two ants are facing each other um and uh trying to go the same way, one of them has to move out of the way and the ones that are not loaded with prey are more likely to move out of the way. That is the only part of the algorithm.

[00:35:53]That's the only behavioral rule you need in order to make these traffic lanes appear. So there's a group level pattern that seems efficient. that seems sort of organized, almost centrally organized and yet it is produced by this incredibly simple rule at the individual level without necessarily the individuals sort of consciously or or at any level understanding that traffic lane formation would be a good idea. So this is the phenomenon that I really want to talk about. It's that there's information about the collective organization in this individual rule um in a way that is not apparent from us seeing what the individual rule is. If I hadn't told you about this phenomenon, just saying, well, a loaded ant is less likely to turn out of the way and avoid an oncoming ant, you probably wouldn't have thought that it would lead to this lane formation. Okay, so we have an algorithm meaning a recipe or behavioral rule of what the ants the individual ants follow and we have an outcome which is this traffic lane formation and linking the two is often not straightforward.

[00:37:03]Um okay so I already said that the algorithms are the rules at the individual level about how to act or react and they imply this outcome but they don't state it explicitly. um either us or the ants may or may not read from the algorithm the information about the collective level app and there are many classic studies that are like this. So for example um in ants specifically the very classic ant algorithm uh that's now widely used in computer science to solve a variety of problems is inspired by pheromone trail foraging. So if you imagine an ant nest here at the bottom designated by N and a food source designated by F at the top and different ants um in this case we have an experiment with only two available paths but um even in a natural situation there are lots of available paths usually to get to the food and the first ants may uh accidentally discover the food by walking an arbitrarily complicated path. But when the ant has dis when an ant has discovered food, it goes back on the path that it knows and lays a pheromone trail.

[00:38:14]Now if many ants lay these pheromone trails and then other ants follow the pheromone trails to the food and they themselves lay a pheromone trail on the way back. What happens is eventually the colony finds the shortest route and it finds the shortest route simply as an outcome of the fact that when any ant randomly happens upon the shortest route and lays a pheromone trail on it, that pheromone trail because it's shorter will evaporate less during the time during any period of time. So in other words, the f the the ant that's walking on the shortest trail can maybe walk back and forth three times in 10 minutes. And so the trail will be reinforced three times in 10 minutes.

[00:38:58]Whereas an ant that's walking the long trail maybe can only walk once in 10 minutes. And so its trail can only be reinforced once in 10 minutes. And this combination of just the physical constraint of the movement and the fermone evaporation leads to the colony level solution to the problem of finding the shortest path. No individual ant actually has to know that it's on the shortest path um or has to measure any path length or compare any paths. No ant has to even know that there are multiple paths as long as they are likely to be recruited to the strongest pheromone trail and they lay a pheromone trail um repeatedly as they're going back and forth from the food to the nest.

[00:39:45]Now I want to show you a couple of examples of this kind of phenomenon from my work. So the phenomenon that I'm talking about is that information about the collective solution to a problem is in in the individual behaviors without necessarily being obvious from the individual behaviors. So the individual behaviors are local decisions that individuals make about how to act when confronted with an environment like foronging a pherommon trail or not. um they may or may not have any knowledge of the global problem that the colony is solving. And so the problem that I'm going to talk about next is uh in bumblebees. This is a bumblebee nest in the lab. It's sitting on uh this wooden cat litter, but the the darker brown structure is the actual nest. It's made from wax. And you can see these um bulges here are the larve that are being raised by the bumblebees into adults.

[00:40:39]And the bumblebees are all um of course like most social insects, all the the workers here are sisters. They're all offspring of the queen that you see up here.

[00:40:50]Okay, so the bumblebees um are trying to raise as many new workers as they can, but the workers once they emerge are surprisingly different sizes. So these are two sisters from the same colony. Uh this is bombus impatience is the species of bumblebee. And in bumblebees, it's widespread that there's quite high worker size variation.

[00:41:12]Now, we know that this the size variation is not genetic. They're full sisters. In fact, they're super sisters um through a genetic quirk of himopta.

[00:41:21]They're related by uh have a relatedness of of 75. But um that the size differences are caused by differences in nutrition during the laral stage. So they're are hollow metabolous insects means meaning they don't grow once they're adults. All the growth is during the laral phase and then they have a final m single final mold uh into adulthood that makes them turn into bees instead of maggots and um they don't grow anymore after that. So these differences are fixed differences that are generated by different nutrition during the laral state.

[00:41:58]And um the question I'm going to ask is well how do the bees organize this? The larve signal when they're hungry? So um who decides which larve get more food and which larve get less food? All the larve presumably want to be big. So how is that decision made and how is it implemented? And it turns out that in bumblebees, again, this is implemented probably by an incredibly simple rule, which is that there are some bees that are specialized nurses, meaning they do most of the feeding of the larve. And those specialized nurses are um faithful to very little micro territories in the middle of the nest. So, in other words, the the nurses just hang out in the middle of the nest most of the time in these central zones here. And so larve that happen to be in these central zones get fed a lot more than larve that are in the peripery where nurses only um walk around uh very occasionally. And uh we know this because we've shown uh I'm not showing the data here on the nurses actually uh sticking to the center. The data I'm showing is just that um these letters here correspond to the letters in the in this uh in these concentric circles. So what I'm showing is just that indeed the larve that are closer to the periphery end up being um a smaller pupil size meaning a smaller adult size too and uh this is because they are fe fed less often than brood in the center.

[00:43:33]So again, we have here a very simple algorithm that is just if you're a nurse bee, you know, mostly hang out in the middle of the nest. And from that single rule, automatically a collective pattern follows, which is that the colony raises workers that are quite um variable in body size.

[00:43:52]And in general, uh in this case, there was this pattern with space in the nest.

[00:43:57]In general, space in the nest often seems to organize work uh and seems to be a central um organizing principle in a lot of different social insects. So this picture on the left here is um from this very nice paper by Michael Smith on the symmetry of the honeybee comb. But I'm not showing you anything here about the symmetry. I'm simply showing if you imagine this is a honeybee comb in the in a hollow tree or something like that.

[00:44:24]This is essentially a two-dimensional structure, but it's used by the bees on both sides. And you can see that the different areas of the comb are used for different functions. So this is my point here that the space is not space is not irrelevant. In fact, different functions of the nest are located in different places on the comb and vice versa. Over here we see um a cast of an of a harvester ant nest. So you have to imagine the surface of the soil is up here and these are tunnels going down.

[00:44:55]They look staircases almost and uh lots of individual chambers in which the ants may have um food stored or brood. These are not actually the ants from which this cast was made. These are just images depicting those functions. So here is a seed storage chamber by some harvester ants. And here is a brood raising chamber. Um of course opened up so we can see it. This is what some of these chambers down here will look like.

[00:45:22]And so across all these different wildly different social insect species, we often see that spatial structure is what drives uh organization and that different areas in space serve different functions.

[00:45:38]And um here again a picture from my work. So these are teenorax ants. They are very small colonies, usually at maximum up to a few hundred, but this colony is probably less than 100. Um, and they live normally in rock crevices, so fairly small uh not self- constructed cavities. And in this study, we actually made them uh artificial nests of different shapes. In the left case, um, just a circle shape. In the right case, a sort of long lengthy tunnel shape, but the two shapes are actually the same area. And um as earlier work by Anna Sedova Franks had shown individual teenthorax ants also have these what's called spatial fidelity zones. So individual teenthorax workers also like the bumblebees don't move randomly across the nest but instead stick to a specific area. So each worker has their own specific area that they stick to.

[00:46:33]And tasks are also not evenly distributed across the nest. As you can see here, the queen and the young brood is mostly in this fairly concentrated area and then older brood is usually in a different area. And then of course there are workers that are going outside to forage.

[00:46:49]And we find that despite this radical reconfiguration of what their space looks like, individual ants do maintain those spatial fidelity zones. So individual ants try to reestablish this uh social organization regardless of what the outer conditions are.

[00:47:09]All right. So, I've shown you all these different um examples of collective organization, and we've talked about how algorithms um may be organizing this collective behavior. So, um here's just um a reminder of how this works in general.

[00:47:25]So, on the right um I have a picture of an Andrew's this case cookies uh from my old German cookbook. And down here is the recipe. And the recipe is like the algorithm. It's a step-by-step instruction for how to create this end result. But from reading the recipe, you may or may not know. In fact, only with a lot of experience will you be able to read a recipe and anticipate anything about the texture and taste and consistency and so on of these cookies.

[00:47:57]Usually the only way to know what this algorithm will generate is to actually run the algorithm and see what the end result is. And this is very similar to how a lot of these systems operate.

[00:48:08]Individual units, whatever the individual units are, have specific behavioral rules, ways that they um interact with each other or react to the environment. But um it's hard to know from just reading the rules what the collective outcome will be. Nonetheless, individual rules, simple or not, can generate complex collective outcomes.

[00:48:32]And um as I already um alluded to, in a lot of cases, these algorithms rely on the use of a common substrate or a common space um and sometimes common materials. And it's often that the information between individuals is transferred through the use of space rather than through explicit signals.

[00:48:55]Now I want to say a little bit more about this difference between describing a system as a model or describing the outcome and describing the algorithm. So in um I feel like this is a central theme in our modern world in a lot of ways. Um AI is one example but not the only one. So to remind you, a model um is a description of a system. So if you have a blueprint of a building or a blueprint of what the cookie should looks like, that's a description of the end result. The recipe or the instructions for building it are the algorithm. They both may describe the same system, but they do so in very different ways. And the recipe actually allows you to construct something. The blueprint doesn't necessarily tell you how you get there. On the other hand, reading the recipe doesn't necessarily tell you what the end result will look like. And um as I said before, these algorithms nevertheless do contain about information about outcome which is realized in the interaction with the environment, but the outcome may be hard to see without actually running the out.

[00:50:05]And um I would argue that in biology actually in general is characterized by um algorithmic uh algorithmic behavior more so than blueprint like behavior. So when we think of gene sequences for example that form a phenotype, the genes are typically not a blueprint for what the phenotype looks like.

[00:50:33]they're typically an algorithm for how to make the phenotype.

[00:50:39]So, um more generally, I think we should rethink um the world, rethink how we think about the world and think about it more algorithmically. But it is hard to do because it's so counterintuitive to go from an algorithm to a description of what the system of interest is actually like. in um in the context of AI there's a lot of discussion about whether uh AI models have world representations do they have a model of the world do they understand anything I mean apart from the question about consciousness just the actual practical question of do they somehow represent what the world is like of course in some way they do otherwise they couldn't answer questions but they do not have explicit models they don't actually make a database of information that they can look up. They just have an incredibly complex algorithm for how to answer your question. So, the information is implicit in the algorithm, but they aren't looking it up somewhere. They they just have an algorithm for how to derive the next word from the previous word. And that word implicitly has information just like these social insect algorithms implicitly have information about how to organize a social insect colony. And the same may be true for our brains. Um, newer studies of memory, for example, argue that there's something called process memory, which is effectively the idea that our memory is also encoded algorithmically. And that may not be true for all of our brains organization, but it is something that um would fit with what we know of how the brain and how neural nets actually work.

[00:52:16]And so these two brains and AI uh specific examples lead to this question of how can you generate intelligence based on algorithms instead of having explicit world models.

[00:52:29]And of course um I already gave the genotype and phenotype example. There are no genes for most phenotypes.

[00:52:35]They're only genes for enzymes mostly and and other proteins. Mostly what the genes are are encoded algorithms that lead to certain phenotypes under the right environment.

[00:52:48]Which once you rethink this algorithmically, once you think about genes as being algorithms, the whole nature nurture debate goes away because you realize that an algorithm only works in the context of a particular environment or or rather it can work in different environments but only in a particular environment can you predict what the algorithm will actually lead to.

[00:53:09]And of course the same is true for our society. Rethinking the world algorithmically means understanding that uh for example who becomes a professor is not something that some central agency decides but it is encoded in the way we do things in the way we do education in the way we do hiring. All of these steps um are of course algorithms that whether we know it or not, small changes to the rules by how these happen can lead to changes in the collective outcome.

[00:53:44]Okay. So next I want to talk um a little bit more briefly about replicators. Um for biologists it's natural to think about replicators. Um but in a lot of other contexts um this is also a counterintuitive thinking that um in my view we need to spread more widely.

[00:54:04]So when we think about collective behavior we often would like to optimize it. We're thinking in this case these are sheep but you can think of your favorite uh collective organism including humans. How do we get them to perform well as groups?

[00:54:19]uh we often seem to think that individuals should want the group to perform well. Um but this is actually not an intuitive result. So how do we get good group behavior?

[00:54:31]Um in general I just want to very briefly review this. Uh this is an animal behavior seminar so I assume most of you have heard of evolution before.

[00:54:40]that um evolution by natural selection is actually an incredibly simple algorithm again um that can that can have a variety of intuitive and non-intuitive results. So the basic principle is just that organisms multiply in number. They reproduce. Uh here we have a little litter of puppies.

[00:55:00]So a dog two dogs have reproduced and made more of themselves. But the children tend to resemble their parents but are not identical. So we probably have the feeling that we can make pretty good predictions about what the parents of these puppies look like. So we know the children resemble their parents.

[00:55:19]There's inheritance and yet they're not the same. Variation is introduced um both by re combination and mutations.

[00:55:29]And just these three phenomena are lead to genetic change over time. Why?

[00:55:35]Because if children resemble their parents, but not everybody is the same, as they go out in the world and interact with the world, some of them will be more successful at reproducing again and others will be less successful. And just by the fact that their children will resemble the parents that were successful after each generation, whichever type was more successful at reproducing will become more frequent.

[00:56:04]So this is what leads to adaptation.

[00:56:08]Now in evolutionary biology we think a lot about what it is what is the unit that is actually doing this. Here I was taking the unit to be the organism the dog. But different things could be the these three rules multiplying, resembling and varying those could be true for different organ for different um organizational levels. So here is this little um picture of Matrioski showing um a few different of the levels that we can think of. Genes, cells, organisms, groups, they may all reproduce at their level. cells make more cells, groups make more groups. Um there is inheritance in all of these cases and there's also variability in all cases.

[00:56:58]So natural selection should occur at each of these levels and at each of these levels we should see adaptation.

[00:57:04]The only problem is that the speed with which this adaptation happens depends on generation time. So generally the smaller units reproduce faster um and so they become adapted more quickly than the larger units. And the consequence of this is that evolutionary biologists generally reject group selection meaning they generally reject the idea that groups become adapted just because the groups themselves are evolving. Um individual level selection is stronger and faster. And so whenever individual level selection leads to different outcome or drives individuals to adapt in a different direction than what would be the adaptation at the group level then that individual level selection is probably going to win out. Um the picture here um there's a lot of people in cars and each of them has this little thought bubble that says if these idiots would just take the bus I could be home by now. And of course this is um a classic image of what we call the tragedy of the commons. This idea that if everybody were to behave in a certain way, ultimately not only the group but each individual would uh win from that alternate behavior. And yet because each individual benefits from doing the selfish thing locally, um individuals end up doing the selfish thing even though that leads to a worse outcome for everybody. That's why it's a tragedy.

[00:58:32]and u mathematical models and empirical studies have shown over and over that this is the typical outcome um for groups evolving unless there are other mechanisms preventing that from happening and so I want to talk about these other mechanisms in a minute but just to remind us these collectives that I've been talking about some of them are in fact um evolved selected and adapted at the group level so in ant colonies as well In multisellular organisms including of course brains and also in this cell all the individual units are not reproducing indiv uh individually independently. In fact the fitness of each unit is directly tied to the fitness of the whole group which means group selection and individual selection individual level selection do not operate independently. they are not able to go in different directions because for the most part um individuals only reproduce as part of the group and only become successful as part of the group. Uh in these engineered systems of course we hope the same thing is true when we write software for cluster computing or we construct a power grid. We are not building in some kind of selfish individual reproduction on the part of the nodes. we are building that everybody is only behaving so as to optimize uh the behavior of the collective. But in many natural animal groups um especially in vertebrates each individual is independently reproducing and so group selection or the benefit of the group can inherently be in conflict with individual level selection. It's not always in a in many cases individual uh payoffs directly line up with payoffs with group payoffs but they do not necessarily and so in these kinds of groups what we should be looking for is individually optimized behavior and therefore the group as a whole may or may not perform.

[01:00:39]Um so to some degree multiple levels are always under whoops under selection simultaneously. So it's not that group selection doesn't in fact happen but usually it is so slow and weak that um any conflict with um selection pressures at the lower levels um prevents it from from uh dominating the outcome.

[01:01:03]Okay. And why is this so serious? I want to just mention one aspect which is communication which has been well studied. We know that of course communication happens between individually individual level selected individuals. So um between individuals that don't necessarily um that may have conflicts in in of interest. But whenever there's conflict of interest between individuals that are communicating um there's a famous article by um Krebs and Dawkins that says that all communication in these cases is either mind readading or manipulation. So mind readading because individuals want to extract the plans or the behavior the future behavior from the other individual. They're not literally mind reading but they're trying to anticipate what the other individual will do and manipulation because most of the communication is intended to get the other individual to do something.

[01:01:58]the the communication in these cases is not evolved to maximize information transmission because the sender of the signal has no interest in maximizing the transmission of information. What they want is just to get the other individual to do something that's in their interest. So, uh communication in these cases is sort of like advertising.

[01:02:19]Advertising may transmit information but only in so far as it achieves a certain goal for the sender. On the other hand, uh, communication in all cases where the interests are aligned, where both partners, the sender and the receiver, um, have the same interests. They they aren't conflicting, then um, information can be very rich and very subtle. Um, so this communication that's advertising is invariably incredibly loud um, loud acoustically or, uh, strong in some other modality. Whereas communication where the interests are aligned the the receivers are sensitive instead of evolving instead of evolving resistance and the sender may actually be interested in transmitting information.

[01:03:07]So the information transmitted is usually much more rich. Now this is the situation we find in social insects where the interests of individuals are completely aligned because they only for the most part reproduce with the colony as a whole. And so this is why uh you know good or optimal group behavior is more likely in these kind of scenarios where the interests of individuals are actually aligned.

[01:03:31]So how do we get good collective behavior? Well, if we're not a social insect essentially it's it's key to tie individual success to group success. And there's a large area of research that I'm not going to talk about in detail, but we know what the kind of things are that we need to do in order to make collective behavior be likely to be good. And in fact, in a variety of contests from multiplayer video games where companies spend uh a lot of money to try to avoid online hostility to things like Airbnb and Uber where companies also have multiple individuals interacting and are somewhat relying on individuals not falling for a tragedy of the commons for individuals to maintain high levels of cooperation. All of these companies are in fact interested in um applying principles and and um implementing principles that help the group as a whole uh perform well and and display good collective behavior. Um and some of those principles are of course uh reputation formation. So some way to rate other people. This is uh this is a key mechanism by which we can maintain cooperation is by partner selection and directing cooperation selectively at other cooperators. This is essentially the basic principle of how cooperation is stable through evolution is when cooperators can selectively direct their cooperation at other cooperators.

[01:04:59]whether it's because they're related to them and they have a good idea about who's a cooperator or whether they um can measure the reputation of other individuals or you know um get information about what they've done in the past.

[01:05:16]Okay. So all these processes that are well studied in biology are the key to um evolving and maintaining in an evolutionary system cooperation among even individually selected uh members of the group.

[01:05:33]Okay. So this is what I wanted to say about replicators and maybe as a footnote I want to say that these are basic bi biological principles that I think biologists have not been um as successful as they might be in bringing out to the world what we can contribute to these fields both in terms of thinking algorithmic thinking thinking about algorithms versus models and how complex behavior can be realized or complex information can be stored in algorithms and how um cooperation can be maintained in groups even when they're under individual level selection. These are inherently things that biologists have studied for decades uh whether with those words or not. And I think we uh we must realize that broadly outside of biology these are principles that are very necessary for improving everyday life and for understanding how the world.

[01:06:26]Okay. And now my last point is about strong inference and intuition.

[01:06:33]So why is it that the outcome of an algorithm is so hard to predict especially in collective behavior. This phenomenon that that's the case has been called emergence in the past. So we call it emergence because it seems like the collective behavior emerges from the individuals in a way that we didn't anticipate or was hard to predict. Um but the information about the collective behavior was there all along in the algorithm. we just somehow we're having trouble making that connection without actually seeing it play out. And there are two main ingredients that um create emergence um that have also been well studied. And those two ingredients are recursivity generally and positive feedback specific.

[01:07:19]Um so I'm going to explain what I mean by those terms. So recursivity thing missing. Oh, there it is. Um, recursivity is just the phenomenon that some process is using its own output as an input. And this could be as straightforward as something directly feeding back on itself. But often um there are multiple components and so some process is in this case inhibiting another process and then this process is uh upregulating this process and then this process is the one that feeds back.

[01:07:53]So there could be multiple steps but somehow an outcome from one thing influences back or goes back as input to the same process. And we know many examples even the very simplest ones in mathematics where this simply implementing recursivity leads to very surprising and varied outcomes. So the image here is of the Mandelroad set. Um it's what's called a fractal. It's a class of mathematical objects that are generated by these recursive equations.

[01:08:27]This particular one is generated by this particular recursive equation and um effectively this is a graph. This is a coordinate system that you can imagine underneath here and each dot each p each uh coordinate point is colored according to the behavior of this recursive equation whether it converges on a fixed value or not. And um and the color is is also how how fast converges. And so um if you've never heard of fractals before, I encourage you to Google it. Um they are incredibly complex objects despite the fact that they're mathematical and they're basically just colored points in a coordinate system, but you can zoom in on these areas uh around the edge of the shape and find an infinite amount of uh structure there.

[01:09:18]And so um people you know nobody had an inkling that you would get that from this simple recursive equation but it's the fact that you um compute something and then the result of that goes back into the equation is what generates this incredibly complex structure.

[01:09:37]But we know that there are many other cases. For example, in the logistic map here we have bunnies or any other biological organism when you look at population size over time. This can fluctuate widely in some cases depending on the parameters of the um of the system and how fast they're reproducing compared to what um what the carrying capacity of the environment is. But we know that we can get chaotic behavior here very uh fairly easily. And this chaotic behavior again comes from the fact that the the output in a sense here is new bunnies and the new bunnies are the input for their reproduction in the next generation. So every population growth is inherently a recursive process.

[01:10:27]And of course here, this is just a cartoon depiction of the classic swarm model of fish or birds um having a zone of attraction and a zone of repulsion and a zone of alignment. So um when they're too far away from other fish, uh they swim closer to them. When they're too close, they swim farther away. And when they're in between, they try to align their direction with the other fish. And this creates um a fairly complex uh swarm or or school I guess when they're fish um movement that also seems hard to anticipate from just knowing something about the rules. So in all of these cases there's in this case the recursivity is just that the behavior of the other fish feeds back onto this fish and then this fish this fish's behavior feeds back onto the other fish. So the fish uh influence the behavior mutually and of course this is the rule for all social systems. Um in all social systems we get this recursivity and recursivity inherently is the thing that creates unpredictability and a particular version of recursivity is positive feedback and by positive I mean not saying that you did something well but um um meaning amplifying. So in control theory when we talk about positive feedback we mean that when the input when the output of this process whatever it is becomes larger um the feedback creates it makes it uh enhances the process even more to make the outcome even more large and the key u behavior of positive feedback is that it's amplifying meaning that small initial differences can lead to big differences in outcome and this leads to extremely nonlinear effects. The maybe classic version of that, the classic um I don't know if I should say example or metaphor is the butterfly effect. Um this idea that a butterfly flapping its wings, very small initial difference in air movement could cause a hurricane somewhere far away. That would be the big difference in outcome. Whether this is really true for butterflies is not clear. But uh the fact is that weather also has very nonlinear effects in that small differences can lead to big outcomes. And there are many many examples in society of this. Um I don't want to argue uh with anyone about the Mona Lisa if you feel very strongly about it. But the painting by Leonardo da Vinci which is in the Louver in Paris uh is estimated to be the most valuable painting in the world. It's uh when I wrote this slide, which is now probably a couple years ago, was valued at about 700 million. Um but it's a very small painting if you've seen it. It's probably no bigger than my computer screen. And um it hadn't been and it had existed for well over hundred years before it was stolen from the museum in the 20th century. And before it was stolen, it was just considered one of the many paintings that Leonardo da Vinci did. and it wasn't particularly recognized as special. But after it was stolen, there was a lot of press about it and so on and then it was recovered and then suddenly it rose to popularity.

[01:13:40]And why is this relevant? Well, if the Mona Lisa was the best painting in the world by some inherent quality measure, one would think that people would have recognized it in the hundred years before that. um but they didn't which um makes it likely that what this is is one of those positive feedback processes where a small difference in this case being recogniz being in the newspapers from being stolen and recovered uh you know gave it that uh little push to become trending as we would say now. And once it's trending, there's a positive feedback process where people share information about it just because other people were sharing information about it and it becomes extremely popular.

[01:14:28]Okay, back to ants. So, so positive feedback of course often is inherent in social information sharing. So, just like when we're talking about the Mona Lisa, if um ants are talking to each other about food and my cartoon version here, the ants have four identical chocolates to choose from. And um by happen stance, one of them gets discovered a little bit earlier. And whichever one gets discovered earlier gets more recruitment and the recruit more recruitment. And so, it's possible that most ants end up at this one chocolate even though all the chocolates are objectively the same. And we've actually empirically studied this phenomenon which is called symmetry breaking. The chocolates are initially symmetrical symmetrical but now they're not because the ants are preferring just one of them. So the ants are breaking the symmetry between the baits or between the the rewards. In this case in the field version we actually used little hot dog pieces. So there's a little maze here. The ants can climb up in the middle just like in the cartoon.

[01:15:28]And then they have four paths of which they can choose one and at each path is an identical reward.

[01:15:34]And what happens in this case is interestingly that some ant species um have strong symmetry breaking. So the y- axis here is the degree to which they break the symmetry the intensity of how many ants go to the most popular hot dog. Um and other ant species have um have less symmetry breaking and in fact if they're at 0.25 25 it means a quarter of the ants go to the most popular hot dog which basically means they're all um well actually no this is this is adjusted to zero. So all of this is asym asymmetric but nonetheless the most popular hot dog is not much more popular than the least popular hot dog. So this species seems to rely more on flexibility and discovery rather than focusing all the ant foragers on a single resource. And we think this directly relates to dominance, ecological dominance in the habitat because these are fire ants, solenopsis and they dominate food sources precisely because they recruit a lot of individuals to them.

[01:16:40]But what is the mechanism by which this happens? Well, in an in um model of ant pheromone trail foraging, we looked at different decision mechanisms. So, these are all um this is a model of ants making a decision at this intersection here in the middle when they're confronted with four different trails or four different paths that they can go on. In all cases, we assume that the ants prefer the strongest pheromone trail. But what does it mean to prefer the strongest pheromone trail? Is the preference for the trail just a linear function of the strength of the pheromone? If that is the case, then larger colony sizes actually have less asymmetry um because ultimately all the trails uh end up getting visited. Whereas if um whereas if the ants have a ranked preference, so in other words, 90% of them go to the strongest trail and going to one of the not strongest trails is just sort of a stochastic event that happens with smallish frequency. then we get quite high symmetry breaking and symmetry breaking stays high even at larger colony sizes. So in other words, the specific the details of the rule by which the ants select the strongest feromone trail even if they always prefer the strongest trail but um how exactly do they prefer the strongest ferommon trail? Do they have a ranked preference or do they have a preference that's just linear in the amount of pheromone on the trail? That very small difference or apparently small difference leads to a big difference in collective outcome like this where in one case we get potentially ecological dominance as a result large numbers of ants on the same source um whereas in the other case um really here we only have strong asymmetry at very very low forager numbers.

[01:18:36]All right. So again here there's information in the algorithm about the collective out outcome. The individual choice mechanism predicts what the collective out be.

[01:18:47]So um in summary I just want to say this is a sketch of the scientific method which hopefully most of you have seen before. The point is that emergence is common especially in collective behavior and we can't trust our intuition about what a change in the algorithm will do.

[01:19:05]any kind of decision-making role um individual behavior role that has changed whether it's in ants or in our society can lead to counter counterintuitive effects for collective behavior and we can't rely on our intuition to predict correctly what's going to happen. Um but we do have the tools as scientists to um to find out what will happen um by using empirical data by focusing on alternative hypotheses.

[01:19:35]Okay. And I think I'm going to skip what I was going to say about the waggle dance which is just an example of how we can use different hypotheses in this case about the benefits of communication in the waggle dance. uh is the communication in this case is the main benefit in uh reducing search time is the main benefit finding the biggest flower patches or is the main benefit in having to uh in not having to learn the rewards of lots of different flower types. We find that it's the latter.

[01:20:07]These are true alternative hypotheses about what in this case the benefits of a particular collective behavior are that all seem intuitively plausible beforehand but empirical data show us that only one of them is correct. So this is just uh an argument for really focusing on collecting empirical data and not lying on in not relying on intuition um to anticipate the results of changed individual levels.

[01:20:37]Okay. And with that, um, I'm going to close. I'm going to say, um, I guess specifically that we should play a lot of, um, attention to spatial structure and how the algorithms can only produce the results we anticipate in the same environment.

[01:20:53]And with that, I'm just going to just um thank everybody who's worked with me recently or a while ago. These are lab pictures uh from different heroes and thank you for your attention was really that was a really fascinating talk and I really liked how you connected all those different aspects.

[01:21:20]Uh while we wait for the questions to trickle down we already have a few questions. So we have a question from David. You said that the army ants will follow the strongest trail to a source.

[01:21:33]Do the ants uh create multiple paths to food? If so, how does the algorithm work to avoid them all just to follow the strongest trail?

[01:21:45]>> Yeah, that's a great question. I mean, um I guess the argument I'm making is that um ants like army ants and fire ants want to all follow the strongest trail. So in a sense it is part of their strategy to arrive anywhere they're going to arrive in forest. Um and so from an from an individual level point you might see you might say oh that's that's not ideal because then there'll just be crowding at the food source but essentially the benefit that they have from that is that they can beat up any other ants that that are there or any other competitors of any kind. Um uh you know in both cases we have ants that are also fastm moving and they obviously will quickly exploit the food source and then all these ants will become unemployed that were focusing on the one food source and then have to search for a new food source again. So um it is a very dynamic system but but you're right that one would think that if there is no competition um that it's that is really not an ideal strategy to all crowd around the same thing but um but it does enable them to competitively um exclude all their their um other ant species.

[01:23:02]Another question from David. Uh, are there any books or resources you would recommend to people wanting to learn more about emergent behavior, algorithms, and etc. >> Um, great question. Um, you know, I think there are probably newer books. Um, I always liked The Self-Made Tapestry by Philip Ball, but that's an older one. But if you want something um that's that's kind of uh you know not focused on social insects, a very general u popular science type book that also includes chemistry and and physics and a variety of other um scenarios. Um I think there must be better ones. If you email me, I'll I'll look up what what I would recommend with the most with the latest ones. Um yeah, >> we have a question from Natasha. Uh really inspiring talk. Thank you very much. Were the experiments repeated several times for each species to establish asymmetry?

[01:24:13]>> Oh, great question. Yeah, the the ant experiments with the hot dogs is I think what you're referring to. Um yes I didn't um this otherwise I would refer you to the paper. Sorry dot is actually a separate experiment in that graph. Um so here um each of these dots is actually one whole separate experiment with a different colony uh on one of these mazes. And so you can see there's a lot of variation, but the fire ants on the whole have higher asymmetry than um uh than these doicarine ants. Um and this is Furelius.

[01:24:56]So these ones are in the middle and they all have good amount of variation. Um like I said, this isn't published yet.

[01:25:03]We should probably have a box plot here to better show um how consistent the differences between the species are.

[01:25:14]Next, we have a question from Katherine.

[01:25:16]Uh, thank you for the talk. Uh, I have a question. In the Bumblebee experiments, if you block the central cells, uh, do nurses do the nurses start to disperse in the periphery, do they develop a new preferential zone?

[01:25:33]>> That's also a great question. I don't know the answer because we haven't done that. Um, we actually did do an experiment in which we cut apart the nest and rearranged the parts to see if they would be uh if there is somehow a signal in the wax. Um, but the problem with that kind of experiment is that the colony is basically your level of replication, right? So, so the sample size is essentially how many colonies you can do it to. And um ultimately we didn't have enough colonies to feel like we could confidently confidently say one way or another what happened. So um I think it's a great question. We don't really know what makes the nurses stay in the center and what they would do if they couldn't. um in I mean in teenorax ants what I can say in addition to what I did say is that Anna Sedova Franks back in the 90s did experiments where the whole colony immigrates to a new nest and she was able to show that they reestablish the same spatial order among the individuals when they immigrate to a new nest. So that's not exactly the way that you were asking it but just that there is some resilience to reestablishing the pattern.

[01:26:42]While we wait for more questions, uh I have one thing I would like you to talk more about is like uh I know you are really big on like uh scientific outreach. Um you have done a lot of scientific outreach. So I would like to talk more about how you would suggest someone to go about scientific outreach especially to like school children or people who don't have as much access and especially in the field of animal behavior.

[01:27:13]Yeah, that's a great question for Bayan.

[01:27:15]I feel strongly that uh outreach is something that people should be doing, scientists should be doing. Um I want to say two important things. One is I think direct the direct answer to your question is that the best way to engage in outreach is to think of what audience do you already have a connection to? um are you um you know do you have a second major and you relate to people in linguistics or or you know or do you have um do you work with mathematicians or computer scientists I mean if you work with other people these are examples from within academia then those are the most easy to reach targets um but maybe you're also maybe you um volunteer in an animal shelter or maybe you do something else in your home life.

[01:28:06]Maybe you go diving or whatever it is that you have going on in your life that connects you to a group of people or whether it's your neighborhood or your family. My personal view is that the most effective outreach is with the people that you already have a connection to. It's very hard to break into some other group of people that you don't yet have a connection to. So, I work a lot with schools, but I also have kids. And so I had, you know, I knew something about what it's like to be in a school. I knew something about what it's like to be a parent or a teacher.

[01:28:40]And then that made it easier to make more connections to schools. And once I learned more about what teachers and schools are like, I could make more connections that way. So it's hard to start these things unless you have an inn. And it's more likely that you'll convince people, frankly, um, if you already are known to them as something.

[01:29:00]I want to make a slight other footnote on this topic. I recently realized that actually a lot of PhD students spend a lot of time doing outreach. And I think that's lovely. But I also want to qualify that by saying I think the professor should be doing the outreach.

[01:29:17]the PhD students. I know this sounds perhaps selfish, but uh while I think the outreach is incredibly important for our society, for you as a PhD student, in order to secure your ability to do more of it in the in the future, you also need to focus on your publications.

[01:29:34]So, if if you're a student in the audience, I would actually say the best long-term strategy is to get your publications out first and and just slowly build up a life. Um and and once you have that you'll have audiences, you'll have connections to groups of people that you can then do outreach on.

[01:29:56]>> Yeah, I think uh I would say that is kind of my experience as well. A lot of outreach have came through other professors. Uh also we have another question in the chat from Andrew.

[01:30:09]Excellent presentation. How do you think the body size of individuals or the colary size affects uh the dominance or discovery? For example, are D inas morphologically similar to as solenopsin sylon?

[01:30:27]Ah great question. Um so um the body size in this case uh is is very roughly these ones are bigger and these ones are the smallest. Um they are not radically different in body size, but in general um we often find that these strongly recruiting ants are somewhat small body size and large colony size. And I say this with a lot of hesitation only because um ant people, ant researchers for a long time have had this gut feeling that larger colony size means smaller individuals and means more trail foraging and more competitive dominance.

[01:31:11]And I think that's true in many cases, but whether it really is a causal pattern, I'm not really totally sure.

[01:31:19]Um, of course these are different species, so a lot of things differ between them and and my graph here is really just to introduce this idea of a small individual behavior being able to lead to a large collective outcome, but to be really sure that that there's this gradient between species, I would like to actually add more species to this graph um and also better represent the variation within and between species.

[01:31:47]But thank you also for the compliment.

[01:31:52]>> I think those were all really great questions. Uh unfortunately we are out of time. Um and I'm sure Dr. Dhouse would be happy to take any questions if you just send them by through email. H thank you again everyone for attending and participating in the chat. And thank you Dr. Dhos for this great talk. For our viewers, please make sure to keep an eye on our social media account for more information about upcoming seminars and other events potentially our an annual conference and we'll see everyone hopefully in our next seminar or our next event. Thank you all again.

[01:32:33]>> Thank you.