 Great. Thank you, Mark. And of course, thank you, Dean. It's very nice to... I'm going to put the mute on. Great. I don't hear the echo. I want to thank you all for hosting me. And of course, it's a little bit better to join you in person, but I'm very happy to be able to join you virtually. So I'm going to talk today about a project that I'm working on with a co-author. Talk about interdisciplinary. This co-author is actually in physics at the University of College London. Andrea Baron Kelly. And the main question that we're trying to answer with this project is how do social norms or social conventions emerge out of a state of nature in which there's an originally sort of disorganized population in which there's no form of social coordination to start with? And the literature on this problem dates back a very long time. And most recently and most formally, we've been trying to study this as a problem of diffusion. And my own work actually starts with that sort of basis for thinking about these collective dynamics. And so the typical way of thinking about this is that there's some sort of seed in a network that's sort of exogenously placed into it. This is really starting from kind of a disease diffusion model of exogenous kind of intervention. And then that's sort of spreading from neighbor to neighbor through a contact network. And my work on this problem on complex contagions really emphasized the differences between simple contagions like disease or information that spread very quickly through randomized networks. And complex contagions like behavior change, which tend to require social reinforcement and actually spread more effectively through plastered networks. Now my work building on really a tremendous literature in applied math and physics used computational models to study this. And then most recently I've been developing experiments to evaluate those models in which we've been able to sort of discriminate the effects of network topology and these diffusion dynamics. One of the beautiful things about being able to collect those kinds of experimental data is that they actually not only are able to test the formal models, but able to inform the formal models. And that's really where this project starts off, is that the more I thought about this problem of behavior change, the more I thought about that empirical test of how any kind of behavior change would spread from the spread of a collective action to the growth of a new social norm to the adoption of a technology, the more I realized that when we think about norms and social order more generally as a problem of social coordination, the problem is actually quite a bit more complex than just the diffusion of behavior. In fact, what winds up happening is people have to adjust their behaviors based on their expectations of what other people will do. And this really takes on the form of an evolutionary dynamic. So instead of an external behavior seeded into a population and spreading, really individuals interact with one another and based on how people respond to their behaviors, they're able to kind of creatively innovate and change their behavior set, come up with new options and see how people respond to those. And of course, as people respond to those options, they develop expectations about other people's behaviors and also expectations for how other people are interpreting their own behaviors. And this becomes kind of a thick space of beliefs about beliefs and emotional responses and so forth. And the more that we kind of unpack this process, the more we realize this space of expectation formation about behaviors really isn't the same thing. It's just a single behavior either spreading or not spreading. But it's this sort of creative process that requires a lot of spontaneous innovation by people. And not only do I have to interact with one other person and figure this out, but simultaneously I'm also interacting with other people and that person's interacting with other people and of course those people are interacting with other people and so forth. And so this is a very complex process of social evolution that allows an entire population to coordinate on some sort of collective norm or social convention that at the outset may not even exist in a population. It has to be sort of generated through these creative interactions. The sort of most amazing thing about these norms and this is something that sociologists have spent generations thinking about is that once they become established, it appears as if there could not have been any other norm. And we see this in people who grew up generationally in societies where norms or conventions are fairly concrete or locked in socially. They tend to carry a kind of moral weight where people believe that that social convention is simply the correct one. But as social theorists, we know that in fact any number of social conventions could have happened and when we look across cultures, although we see tremendous coherence and much nity within the culture, different cultures have very different standards for gender conventions, for race conventions, for norms about politeness, for standards of reading, standards of fairness. All of these things vary tremendously which tells us there's no one right answer. In fact, there's a social process that gives rise to some kind of collective convergence. And in fact, I think the default in several generations ago, I've been to assume that if there's a social norm that's locked in, it's somehow welfare producing. There's this fantastic quote by Canara where he actually says that's what norms are. They perform a social function. But several pieces of work, including my own work on the emperor's dilemma, studies how norms that are explicitly not functional. And in the case of the emperor's dilemma, we studied norms that were unpopular in the sense that they were undesirable both at an individual level and a collective level. So they were universally welfare reducing. And nevertheless, they could become socially reinforced throughout a population. People would not only comply with them, but actually enforce them on all of their neighbors. And so what this tells us is that the existence of a norm in society, this evolutionary process, doesn't tell us anything about why it's there. The norm itself is a kind of function of this social process, which really raises the question, where do these norms come from? How do we actually get a grip on them in the first place? And of course, this is a problem that dates back to antiquity. Aristotle first discussed this question. But to my mind, it was really the Enlightenment philosophers who first formalized this question very clearly. Hobbes, Hume, were so all kind of addressed this problem in their own way. And to my mind, the Humean approach or Hume's puzzle is the one that really frames the question that we're after. And this is actually the question I'm going to try to address today. He says two neighbors may agree to drain a meadow that they possess in common, because it's easy for them to know each other's minds, and each may perceive that the immediate consequence of failing in his or her own part is the abandoning of the whole project. But it's difficult and indeed impossible for a thousand persons to degree any such action. So Hume is really very clearly problematizing the micro-macro problem. He's saying two people through a process of trial and error can eventually figure out that they have a common goal and agree on how to achieve that common goal together. But as those people interact with other people and still other people, that process of local coordination takes on very complex dynamics, and it's hard to imagine how that would ever aggregate to an entire population. So that really frames sort of the theoretical puzzle of social order, of how interactions ever aggregate to produce kind of a social norm at the population level. This also faces us with a very clear empirical problem, which is the case that any of the social norms that we want to study from gender conventions to standards of fairness to linguistic conventions are already occurring in society, which means that there's tremendous history to all of these interactions. There's institutions that have helped to facilitate them. And getting an empirical grip on what the kind of origin story is of how this process happens from the ground up is very, very difficult. And so the approach that I'm going to present today is an experimental approach, and I want to make it really clear that I'm not talking about a psychology experiment, which is the way that we've typically thought about experiments in social science, which is that we give a person one kind of experience, and we give a different person a different experience, and then evaluate the differences in their individual responses. And then we'll repeat that process again and again to get some kind of statistical evaluation over many trials. The strategy here is a sociological strategy. And by sociological experiment, what I really mean is that we want to do is independently manipulate the entire structure of a society and see if it produces a collective outcome that's different. And I think that methodologically, the kind of origin point here is really Durkan's work on suicide, where he takes this idea of being able to independently compare entire society side by side and tries to identify suicide, not as a behavior that one person engaged in. He doesn't try to explain causally why one person or another person did or didn't commit suicide. He takes an entire society, describes its social structure, and then looks at the distribution of suicides over that population. And he takes an entirely different society and says, look, there's a different social structure here, but everything else is held constant. And now we get a different distribution of suicides. And so each entire society is a single observation. And that's exactly the strategy I want to use today. I want to try to describe how comparing different societies side by side will allow us to evaluate how simple changes to their underlying social structure will endogenously produce the capacity either for social conventions to form or the failure to achieve large-scale social conventions. And of course, our assumption at the outset is that the success or failure of social conventions doesn't depend on a specific answer being present in the population. You can imagine the same ecology of options circulating in a population, but based on the social structure, either an ecology will produce coherent convergence or it will fail. And there's many ways of thinking about social structure. I want to emphasize that the networks approach isn't the only approach. Of course, you can think about institutions for redistribution taxation. You can think about distributions of trades or demography. You can think about group sizes and heterogeneity across a population. All of these things are important forms of social structure. I specifically focus on network topology as the variable I want to look at because of the, again, because of the framing of the puzzle originally in Hume, I think is fairly close to the way that we're thinking about it today, which is essentially that we want to look at this sort of pairwise problem of people trying to coordinate with one another and see what this means in terms of the aggregation process of sequential coordination throughout an entire network. So my approach here is going to be going to present a little bit of theoretical background in the model that I'm using to study this problem. But the real focus is going to be here on the empirical works. We're going to do the experimental design. And then, of course, the empirical findings, which I think are actually fairly exciting. This paper isn't out yet. We're still kind of developing a presentation for it, but we're very, very excited about these results. So the theoretical background for this question really starts with the fundamental Hume puzzle, which is where do norms come from? How is it possible that societies can self-organize at a very large level? And the answer has been pretty much for centuries now that there's some form of institutional mechanism that governs this process. That institutional mechanism can either be an authority figure, and here we're thinking of, in the main, something like the government that mandates driving them left or driving them right, it mandates the sorts of languages we use, and punishes deviance and thereby kind of enforces global coordination across the population. But this kind of incentive can also be a positive incentive. Organizations frequently attempt to explicitly or implicitly manage the collective behavior of their populations by incentivizing people to engage in certain behaviors, not others. And so what this does is this is able to collectively drive an entire population to converge on a certain kind of collective behavior, which, of course, once it's achieved is self-reinforcing. Now, it's often the case that we see examples of large-scale collective behavior and large-scale coordination without any obvious authority figure in your government or organization driving this coordination process. And so it kind of problematizes the question of how could this ever happen if there isn't a centralized authority incentivizing it. And one of them I think most compelling answers has come from Harksonian-Selton's work on payoff dominance or risk dominance in the alternatives that people are choosing between. So they won the Nobel Prize for this work. And what they essentially argue is that there's some fundamental bias in the options that people are choosing between, where one of them is either just obviously worth more than the other, or one of them is safer. It's less risky than the other one. And what they notice is that through the dynamics of iterated interaction, the option that's either the higher payoff option or the option that's less risky is individually preferred. And that individual preference incrementally aggregates to produce a collective preference where individuals notice that other people have a slight bias towards one option and of course they have a bias towards that option as well. And then over the entire population that can then result in collective lock-in on that one particular option. So that provides a solution of how this can sort of happen without an authority figure. However, there's fairly thick institutional assumption still. We have to assume that everyone has a sense of what the collective payoffs are for the options, that there's only a finite set of options. And that everyone knows what everyone else's payoffs are as well. So there's this sort of, you know, this common knowledge that pervades the entire process of normative convergence. And this eliminates the question of how people coordinate on options when there's no differences intrinsically or a priori in the payoffs or choices that they're evaluating. And further, it doesn't explain how this happens when there's no real sense of what other people have in their heads in terms of their evaluation set. And so when we tend to think about this occurring in kind of the larger space of cultural media or cultural norms or linguistic norms where there's many, many options people can innovate and there's no real differences between them, we're left again without an explanation of how that can happen. So one of the most popular explanations in sociological literature on this is informational feedback or keynote advantage, which is that one option really gains traction just through sort of sarcastic processes. And then that option becomes sort of more available in the population and reported as such. And then that generates a kind of feedback effect. This solution, however, relies on some sort of, typically some sort of polling mechanism that can sample over the entire population simultaneously and then report back in a kind of collective form the instantaneous state of the population. So we think here of like best-seller links or polls which are very common in sort of voting situations. We say what's the state of the population, what's everyone doing, or what's the most popular book, or what's the most popular movie, what's everyone doing. And this helps to facilitate the feedback and then ultimately coordination. Again, it also requires really thick institutional mechanisms in the sense there needs to be a way of polling the entire population simultaneously, extracting the information about what everyone's doing and then reporting that back in a coherent way that everyone knows how to interpret and furthermore that everyone's actually already paying attention to. They know that there's this reporting system. So these three solutions have all, I think, made important progress in trying to understand this problem but they still leave open the question of how social conventions, the social norms, ever get started in the first place without any institutional mechanisms in place to already facilitate the process. And the answer that I think is as most compelling in the literature now is the social evolutionary response, which is that just through the process of individuals interacting in social networks, these micro-level interactions can aggregate, sort of, you know, sort of build up to produce a collective level coordination throughout a social network without anyone ever intending it, without any incentives for trying to do it, and without really any information about what the global state of the population is. So really, as an unintended consequence of individual interactions, the collective population winds up all agreeing on this behavior. So the way this has been approached from a kind of a formal point of view, it's a really study that's a coordination problem and this, you know, in the kind of formal game theoretic sense. And when we think about that from a network's perspective, this has really specific implications. So it means that, you know, two people who are trying to coordinate one another, on the left-hand side, you can see a spatial network, and in the middle you can see a random network, and on the far right you can see quite of the homogenous mixing case. And in most of this literature, the far left and far right have been a two that have been compared. It's either a spatial local network or a homogenous mixing network. And the argument has been, and series of models have shown this, but the spatial networks are far, far superior for generating rapid coordination. Now if you let any of these networks go, like a ball for an infinite time period, just the process of social interaction and stochasticity will ultimately lead to convergence on some sort of norm in all of these populations. But the real question is one of timescale. The real question is, what's the rate at which a population will be able to converge? And of course, for our purposes where we're trying to detect this empirically, we want to know, well, what's a timescale that we could actually measure? We could really see this process of social local interaction aggregating to produce a kind of collective norm across the society. And the answer has been consistently that it's the clustering, the clustered lattice network. And the reason for that is first of all that if two people are interacting in the lattice, they could keep interacting with the network partners, which establishes reinforcement for the answers that they're receiving from that same person. Furthermore, two people tend to share neighbors in common. So Bill and Tom are interacting and they agree on a norm. And Bill is a friend of Bob and Tom is a friend of Bob. Then Bill and Tom can both coordinate to force Bob to engage in the same behavior. And that sort of power of pollution makes these clustered networks extremely powerful for being able to enforce the social order locally, which then spreads from neighborhood to neighborhood and makes global coordination very effective. As ties are rewired, you can see in the middle random network, people still have the same fixed neighborhoods. They're interacting repeatedly with the same people. But those people aren't interacting with each other and that makes it more difficult for coordination to get off the ground. And then in the far right-hand side, people are just picking others at random. So each interaction is a random interaction without any really repeated contact. And as a result, it's very hard for people to figure out expectations about what the next person they're going to interact with is going to do. So this is the standard way of thinking about this problem of the emergence of conventions from a local self-organizing perspective through the lens of social networks. The approach that we're going to take complicates this a little bit by asking a further question, which is that almost all of this work has assumed that people are choosing between two options. For example, A and B, which means that at the outset, although people are coordinating locally, people in remote parts of the network have the same options that are available. Everyone starts with either A or B available. And the question is whether the particular equilibrium gets chosen. But in real social conventions, in social emergence in the natural world, people have this capacity to spontaneously innovate. And this complicates the story considerably because it means that new options can pop up in different regions of the network where other regions have no sense that those options are even a possibility. And so people in one area can be all working on trying to sort of agree on one set of options. And the set of options that sort of emerge spontaneously in a different region can be entirely different. This has dramatic implications for network structure. And in fact, all of the work on looking at these kinds of innovation models and also on the idea of information diffusion the spread of kind of new options across a population has emphasized the value of random networks. So the randomly mixing model on the far right hand side is the ideal case for this because it means a new option can spread almost immediately across a population. And so what we're interested in with this project is trying to combine the fundamental question of local social coordination with the opportunities that people have to innovate as part of this process and to see how those two interact to affect the ways that networks generate collective coordination in a population. So the way that we study this is called the name game is by looking at the process of the emergence of linguistic and linguistic conventions. And this approach to studying social convention states back to Wittgenstein who first really kind of formalized this idea of thinking about social conventions as kind of a two-person linguistic coordination problem. And the basic sort of setup is that two people independently look at an object and their goal is to come up with a common name force that can communicate with each other. Now they independently sort of propose names for this object and then each time they both propose a name they can see which name the other person proposed and they sort of accrue a memory of options and then as this memory increases and increases people are able to eventually through process of trial and error figure out what option the other person is going to say next and then they're able to come up with a consensus term. Now, so that's the sort of coordination part that maps fairly clearly onto the stated coordination game way of thinking about this process of convention. However in this process of trial and error people can spontaneously suggest new options and creatively respond to the options that they've been given. So they can combine different words or just suggest a new word and all of this then creates an innovation aspect to this problem. And of course our question is how does the interaction network affect the way that this local process then aggregates over an entire population and how does it really affect the possibility for conventions to emerge at all. So we've done a series of formal papers and Andrea has done works really since 2006 building kind of large theoretical literature around this question. Our goal is to be able to study this empirically. Now when we move to studying this empirically you can preach that there are several very difficult problems that make it a difficult problem to study. Now there's been a lot of work and a lot of discussion about the difficulty of studying diffusion in social networks the simple problem of identifying a contagion dynamic and identifying how social interactions affect that dynamic and really the causality of spreading. And people I think over the last five years have done tremendous amounts of work to advance that discussion. Matt had a great presentation yesterday on this kind of work thinking about very controlled observations of essentially what's natural experimental settings trying to figure out the exact network structure and what the exact path is of behaviors through that. And then my own work also thinking about it from an experimental point of view is used on my networks to be able to control the network structure and look at diffusion. And so with that sort of I think advance in our ways of thinking about this question this new question of the evolution of norms actually presents a set of new challenges that need to be handled in addition to the kind of well understood challenges of addressing the dynamics of contagion in networks. One of the most prominent of these challenges is group size. So there's been a tremendous amount of work starting in the 50s and 60s of MIT and Harvard on small group lab experiments. And this group research really developed through the 80s and 90s and in fact most recently in 2007 Richard Salton, the Nobel Laureate was doing very interesting work on language convention and experiments of just two to four people. The important thing about being able to formalize these kinds of dynamics in models before we're able to study them empirically is that it gives us a sense of what the relevant sort of parameter spaces for group size and what we're able to find is that in these particular models of evolution of social conventions for groups that are below a size of 20 the dynamics are exactly the opposite for groups that are above a size of 20. That's incredibly important because almost all of the work that's been done or in fact all of the work that's been done looking at the endogenous dynamics of cultural evolution in social interactions the sort of thing that we want to study producing social conventions for two, four, eight, or twelve people which means that those dynamics don't tell you really much at all about what's going on as you generalize to a hundred, a thousand, or a hundred thousand people so you need to study larger groups in order to really understand how these results would aggregate and really for the question of external validity how they would apply to other situations now the advantage of this information also tells us that we don't need to study 61 million people at the same time all we really need to do is to study group sizes that are well above that critical point of 20 and if we're able to do that then our results according to the model should generalize to larger and larger populations and that gives us insight into what happens when we aggregate up to a hundred million or nine billion because those dynamics are expected to be the same and so what we do then is to identify this critical point and then start above it and incrementally look at how these dynamics aggregate there's been attempts to solve using economic games where people have been doing lab group experiments with groups of size 30 or 36 Michael Kern just wanted people and that's incredibly important work that sort of pushed the boundaries of thinking about what you can do in a laboratory unfortunately because of the complexities of studying real-time behavioral dynamics those experiments have really focused on economic games where people are given one or two set options they're given fixed payoffs for those options and they're not allowed to innovate and furthermore they're given global information about the state of the world and so this kind of all those provides sort of important methodological innovations into thinking about how to study these problems it doesn't address the fundamental question of how social evolution occurs just through local interactions with spontaneous innovations now the natural response from a lot of researchers these days is to move to big data and it makes a lot of sense it's a very good idea because big data provides the behavioral fluidity of natural evolutionary interactions and it also of course allows people to get very very good longitudinal data very precise data and also data at scale where networks tend to be fairly well resolved the problem with studying these kinds of questions and with a big data format is that we lose all of our experimental control on causality is to say we don't have any information about what sorts of leadership is emerging in those populations what other exogenous influences may be available to some people and not others and then forms of social control and focal points as well and so the idea that this process can happen without institutional mechanisms playing a role is very difficult to test in a big data environment because we're not able to eliminate those institutional mechanisms so all those three problems are kind of the default problems that we need to be able to solve in order to answer this question which we haven't been able to answer yet with existing methodologies there's a final and fourth one which hasn't really been worried about that much and it's because it's such a hard problem to get at given that the other three are already difficult and this is the question of replication or reproducibility so imagine you were able to solve all those three first problems you're able to study in a society where you knew the full structure of the network you could study the real-time evolution of that change and on a measurable scale watch the formation from the state of nature of a new social convention and you're able to do that across let's say comparative societies that were otherwise in every way independent so that would be remarkable but that would really according to the Durkheimian methodology it would only give you a data point where the society of 100,000 people all connected in a network is a set of interdependent data points which means each of those 100,000 people it's not IID so if you want to study the collective process of aggregation and use the entire society its structure is the independent variable and it's a collective pattern of social coordination as the dependent variable but you need to be able to study the society replicated several times to get any real grip on whether there's a causal effect of the social structure on the collective process and that's been the real linchpin of I think moving these questions in social science into kind of a more of a hard science direction and so the solution to this at least the one that I'm proposing is that we can think about this as a problem that can be studied online using web-based experiments and the trope or metaphor I like to use is the social petri dish and the idea is that just like a petri dish we can kind of create two sides side by side see them in exactly the same way and make invisible changes to their social structure, changes that are unknown to anyone in the population and nevertheless just by virtue of people interacting with each other in a completely natural way one society will grow a collective behavior and the other society will fail to and this would give us real traction on the capacity of these sort of changes in network structure to be causally responsible for the emergence of collective behavior and of course the real punch line of this is that once we can do that once the fact that we're online and the fact that all of this is built in a virtual community allows us to replicate this process seamlessly again and again and again much in the way that we would do in wet biology or in bench science it's a sort of straightforward thing from a technical perspective to do this twice once you've done it once whereas that's not always the case in the laboratory environment and I've used this kind of replication strategy in several papers over the years and these papers have all been looking at typically at diffusion dynamics or changes in network topology and so the question here is whether we can use this to study the spontaneous production of collective behaviors and the emergence of social convention so the experimental design is based squarely on the name game we're using that kind of formal approach of thinking about dyadic interactions on coherent naming conventions as our empirical domain or our template about social conventions more generally so participants are recruited in the World Wide Web at large using online recruitment ads places like ad week have been very successful also through Reddit and other sources and then once they're recruited they join the study and are basically put into a pool of other people who are then sort of fed into an online social network where such topology is already created once they arrive in the study and the study starts they're given an object to look at and try to name in this case it's a face and of course they interact in pairwise fashion with other members in the community just as we would imagine the name game to try to coordinate using the same name and our goal is to see whether some coherent social norm emerges based on the interaction structure of the network so when people actually are recruited and arrive to the study the arrival site looks like the kind of scrolling list of instructions that people are able to read while they're sitting there waiting now you may ask why are they waiting they're waiting for other people to show up and so if you think about this in terms of the actual sort of mechanics of running an experiment if we imagine three different network structures a spatial lattice or a random network or a sort of fully mixing population we want to run all of those simultaneously so if each network lets say has 20 nodes we need to recruit 60 people at a time and then wait for when the 59th person has arrived everyone is still sitting in this waiting page and then the 60th person arrives and then we randomly allocate people network structures so 20 go to the spatial, 20 to the random and 20 to the mixing and then we initiate the game dynamics and when we initiate the game dynamics this is the screen that people see the the center of the screen just shows a face and then an entry box where people can enter whatever name they choose left hand side shows a list of players and the right hand side shows a list of rounds and so the game will keep track for you which rounds you're in up to round 20 and the game has a fixed limit of 20 rounds and tells you the state so you're currently in play on round 1 and then once you complete round 1 it'll tell you whether you matched or not now when people enter the name and hit send choice it tells you to wait for your partner because people are choosing simultaneously and then once your partner enters the name it reports back the name that your partner gave and of course reports the name that you gave to your partner so if you matched it tells you that you matched and pays you the winning amount which is nominal 25 cents for matching and if you fail then it actually tells you that you failed and tells you what the other choice was and so that way you can accrue a memory of what the other choices are in your population if you fail and you can also of course know that your other name is being used if you succeed one of the important things about this design is that that list of players on the left hand side doesn't change and that's important because in the spatial network you only had four neighbors that you'd be interacting with and that was also the same in the random network where of course in the modestly mixing case and so telling you that you only had four neighbors or telling you you had 20 neighbors would tell you something about the population structure so to avoid revealing anything about the population structure we just provide that complete list of names for everyone so that there's sort of complete opaqueness about the differences across experimental conditions so I have the mute button on I'm going to turn that off for a second I'm going to hear my own echo now I want to make sure that people are clear in order to understand the results for this experiment it's important to understand this particular design so if people have any questions about this user experience they can ask them now before I move on to the results Mark perhaps you can moderate here are there any questions coming up okay everyone gets it all right great so in that case then I'm going to move on and move on to presenting the results from each of the networks okay we're going to play for you now it's a movie in which the data are just sort of played back in real time one of the advantages of doing this kind of data collection is experimental online environments is that instead of being able to or instead of being forced rather to process the data to present fancy statistics that explain what happened you can actually just present the raw data this is something you're told never to do when you're in grad school you never just show raw data but what I'm going to do is to just play for you the exact raw data of what happens and it actually has a tremendous clarity which is always sort of surprising to see in data collection so what's going to happen is that as people choose names colors are going to show up on the screen so each node corresponds to a person and the color of the node corresponds to the particular name they're choosing I'm also going to show you the list of names that show up so you can see the correspondence so here you can see people are playing the game and as two people interact a link forms between them and if the link is white it means they failed to coordinate but when the link has a color then it indicates the color of the name that they're all coordinated on so you can see here the name Sarah is blue the name Elena is green Susan is sort of green and what you can see very quickly here is that this is interacting in a spatial network what you can see very quickly is that coordination starts to get off the ground in the spatial network on the top people are coordinating on the name Sarah and on the bottom they're coordinating on the name Elena and so because of this I mean you have a third of the population coordinating already it seems like this should take off very quickly and this is sort of consistent with the traditional kind of coordination model approach but what winds up happening next is pretty interesting so you can notice between right let's see here if the mouse shows up right here between the Elena and Sarah group you've got this name Julie that shows up and these people together according to the name Julie even though they're surrounded by two sort of dominant groups on the other side you have a similar thing with this name Samantha and the fate of this interaction process is indeterminate all of these options have equal viability in terms of eventually in the long run becoming an option that everyone winds up using and so there's no preferred option in the population or series of groups competing with one another and through this process of local competition they actually fail to achieve any kind of global coordination over the sort of span of the experimental study so because of this failure then the idea was to move to a random networking study this process again and now because of the sort of long distance ties perhaps this would facilitate the diffusion process that would allow coordination in the network so again in the same way you can see the process evolving people choosing various names and ties forming but of course these ties are now showing up across the network structure the interesting thing that happens is that the choices that people are making tend to look a lot like the dynamics that we saw in the first case which is that you see the name Lisa and Jane, different names interestingly Sarah still showed up but Lisa and Jane becoming sort of and Jennifer becoming popular names that are used in the population and sort of small groups starting to form among people who interact with each other but then outside of those groups again there's different names that are starting to achieve dominance and again you get the same sort of dynamic where people are competing with one another in these group structures these emergent group structures which have no intrinsic value but take on a life of their own yet there's no global coordination and if you look and see some of the options people have chosen there people chose no match which is kind of a clever way of trying to get out of the problem of choosing one name or another but to choose a kind of name in the previous people entered things like blue eyes people said beautiful right so people entered categories instead of names and this was very much along the lines that we think maybe a good solution will diffuse maybe there's an option that's a more attractive option that will generate coordination but in fact those options were completely ineffective which is to say I found that very interesting because it really goes to the kind of theoretical point that the dynamics have much more to do with the population structure in this case the network structure then with the particular options that are present so just suggesting a particularly clever solution doesn't affect anything in terms of whether the population is able to coordinate on that solution and so the end state of this dynamic looks very much like the end state of the spatial network and so that's also surprising because typically when we think about diffusion we think about the random network being the opposite extreme from the spatial network and also in the other direction we think about coordination so to see very similar dynamics in the random network and in the spatial network tell us that the dynamics aren't quite understood by the models that we've been using to study this coordination and diffusion so now we're going to study this from the classic case of homogenous mixing and in this case I used a male name we did all of these studies with male and female pictures to study to make sure there was gender parity in the dynamics and you can see as before one name has some early traction but something sort of different is happening now than happened before which is that one name you can see the name John is starting to gain traction and it looks like it's gaining a level of popularity that none of the names ever really achieved in the other network which is to say names got early starts and some names got popular quickly but none of the names ever gained a kind of majority share of the population they all just were competing with other names and in that sense there was sort of symmetry in the likelihood that any of the names would gain dominance but all of a sudden you have one name really taking over the population now you can see it's sort of just exploding and you get this kind of spontaneous creative emergence of a naming option throughout the population and what's exciting about that social process is that it generates coordination in a way that people who have never interacted with one another before total strangers towards the end of this process total strangers are meeting and then when they meet they're able to interact spontaneously and coordinate with one another and this is something that of course didn't happen in the other case in the other case it required a repeated sequence of local interactions to coordinate and then when they were interacting outside of their groups they failed to coordinate the other thing to notice also is the variety of names in this case you've got lots and lots and lots of names that I think are interesting also because of their cultural diversity you've got names like Tejash and Radu mixed with Aaron and Martin and Michael and David so you've got just this tremendous variety of options showing up and so the diversity of options present in the population actually isn't in any way an impediment to coordination coordination is being driven sort of solely out of the interaction network again with very little to do with the operation we're able to do with the operation or the available behaviors that are at play in the population so when we look at these dynamics sort of more carefully in terms of the overall result you can see that the spatial network and random network really had dynamics that were essentially equivalent the mixing population however converged to full convergence very very quickly there's a couple of really interesting things to look at in these results the first is that you saw 40% coordination in the spatial network as one of the maximum amount of success that any of the names got so you're looking here as the y-axis is the fraction of people who've chosen a given option and the x-axis is the time scale for the evolution so 5 rounds, 10 rounds 15 rounds up to 25 rounds and on the equivalent time scale all of these ran for 25 rounds we see that really no options gained more than sort of 40% share of the population in the spatial network we look over to the random network remarkably we see an equivalent thing we see that in fact no options ever gained more than 40% within this time window and so that really tells us that not only are the dynamics visually and qualitatively similar but quantitatively we're actually seeing the exact same dynamics in the spatial network and in the random network now we move over to the random mixing case what we see is that of course we get global coordination but not only that we get on a very fast time scale and interestingly if you look at about you know round 8 or so right here you see a little achievement of about 30% and at that point this option already has more popularity in the population than any other option and then that option takes off and gains dominance now 30% you would think maybe kind of a magic number but this is something that when people think about critical mass dynamics one of the very popular things to do is to say what's the critical fraction what's the fraction at which we get takeoff the really compelling thing about these results is that they show that the fraction at which you get takeoff is totally not the right question to ask but in fact it really has much more to do with the structure of the social network so in both the spatial and random networks you get 40% of the population coordinating and yet it simply doesn't take off and so the mixing network is able to get takeoff with a far lower coordination size as the initial group to motivate a dominant answer now what I want to do next is compare these results which are the empirical results to the theoretical results I'm going to play you the model that we used in this experimental design which is the model of the name game and what you can see here is when we play out the model and report the exact same thing which is the popularity of different names over the same time scale the most remarkable thing about these results is not only that they're qualitatively similar but they're quantitatively almost identical which is that we see that the dominant name gets about 40% of the population spatial and random networks and that you get really fast growth out of the mixing network on about the same time scale pretty interestingly the mixing network is actually faster in the empirical case than it is in the theoretical case and so what's going on here is really what's called symmetry breaking and the underlying dynamics that govern this process have to do with the change between every option being symmetrical in the sense of every option is a possibility every option could in the long run win versus one option breaking symmetry with the others and just becoming a clear dominant win and on the time scales we're looking at the spatial and random network are unable to have any options break symmetry the networks actually constrain the ability of any of the options of any popularity but in the fully connected network the symmetry breaking takes place fairly early on and then a sort of dominant winner emerges at that point it becomes strongly path dependent it doesn't actually reverse the process now if we look a little bit more closely at it there's a very clear dynamical clue that this is happening which is the distribution of names in the population so you can look here around five we've got a close up of the distribution of names and you can see that this on a sort of log log scale this exhibits a slope of about negative one this is like a zip law where these names where some are more popular than others but none are dramatically more popular which is to say there aren't names with many orders of magnitude more popularity than others and what this does is this means that every names have all of the names that are in have sort of an equal foothold into the population now what happens at around 15 when the split starts to take place is that distribution becomes more extreme and now you can see there's a couple of names that are starting to look like they are moving up in orders between them and the lowest names so they have significantly more membership in the population in the lower names and around 25 where the split has become final here you see that there's basically one name that has almost everyone in the population and it's orders of magnitude more than most of the other names in the population and that's when that's sort of the defining signature of symmetry breaking that can take place and it kind of a where take all state the next question to ask once we got these results was how do these results generalize the scale now of course these experiments for a population size of 24 and our model says that above a population size of 20 we can expect these dynamics to generalize so we could just stop there but empirically it's the most important question to ask is whether or not that model really gives us any empirical insight into what happens when we start looking at hundreds or thousands and so our goal was to then look at larger populations this is a very hard thing to do and a little bit risky in the sense that we were able to show the result we wanted to show and get some traction on this problem that one network versus another network makes a significant difference and you're always worried that if you move to a larger population you start to lose clarity on the results just because the difficulty empirically of managing these kinds of dynamics of larger scales nevertheless we decided to conduct the experiment again with networks that were twice the size so this time with 48 now we know that because coordination failed in the spatial on the random networks with 24 nodes that it's going to be even harder with 48 nodes so we did replicate those but they looked exactly the same did nothing happened and nothing would be expected to happen the real question is whether the fully connected networks, the homogenous mixing cases would be able to sustain their success as we move to larger scales and so here we replicated the study with an equal of 48 and there are a couple of stunning things about these results the first is that they're extremely clear that you know we got on the same time scale we got the result and the second is just how quickly the the dynamics take off that people were able to coordinate in a population of 50 people randomly interacting with neighbors you can see by about around 15 you've got full coordination right so this means that you've interacted with only a given person has interacted with only about 14 or 15 other people which means there's been no replications you've interacted with 15 strangers and by the time you interact with a 16th stranger you and that person are saying the exact same even though you have no history of interacting with one another and then that just gets repeated throughout the population so that's a tremendously to my interesting result because it gives us some real traction on the the generalization of these results to larger scales and it also tells us something interesting about the time scale which in this case is actually faster than theoretically and then just as a kind of test and we were a little bit nervous to do this we thought it would be kind of cool to then double the population size again you know what if we could study 96 people or essentially 100 people interacting online simultaneously in this space where they were totally blind as to how many people were interacting how many people they were playing with and with the overall population size was so we actually some trepidation ran this experiment and got this result which again surprised us because of its clarity not only did the norm converge again on an empirically tractable time scale but it converged very quickly by round 25 there was a clear dominant winner and it was almost the process of convergence and then by round 35 we essentially had convergence throughout the population and so this gives us again confidence that the theoretical model we've been using to describe that the naming game model is actually a fairly accurate way of thinking about these coordination dynamics when there's sort of innovation at play as a final test we replicated all of these studies several times just to make sure that we were getting the results and that's always again a sort of nerve-wracking thing to do as a scientist because you never expect to get results that clearly match with your model and when you do there's always a temptation to like not touch anything but we replicated everything several times and this gives us a very clean picture of the collective dynamic so the top row is the spatial networks and the random networks at different scales replicated and the bottom row is the modular mixing networks replicated at different scales and what you can see is that in every single case the dynamics are essentially identical both in terms of the aggregate success in the local networks and also the time scales at which success occurs and so this gives us a causal test where we can say let's compare the locally connected networks to the homogenous mixing cases and we can see with we have a significance of .01 that there's a significant causal effect of introducing network connectivity or increasing the number of people that I was interacting with on the likelihood that spontaneous social conventions will emerge despite the fact that people have no idea how many others are interacting with it and so from an application point of view this provides some really nice intuitions about the sort of increasing connectedness in our mind domain frequently it's the case that we know who we're connected to but we don't know who our friends are and so we have very little idea how large that space of potential communication is nor even who those people are and what these results say is that as that space becomes very large of friends and friends we can still have rapid large scale coordination on normative behaviors that are spontaneously introduced by people in the network which then sort of speaks to the question of what happens in terms of cultural homogeneity and these sort of homogenization processes more general on large highly interconnected online networks so in conclusion the question that adapter at the very beginning was this question of the orders of social order and social conventions really endogenously emerged without any of these institutional mechanisms in place to guide the social process and I wanted to contrast this sort of coordination perspective on this problem with the capacity for people to independently innovate and what we found was a very clear empirical result that increased connectedness is able to spontaneously generate social conventions in a replicable and reliable way and this gives rise to an exciting new way of thinking about the question of coordination and collective behavior in social networks and particularly online social networks that can be studied rigorously using these online experiments to think about evolution social norms and also think about what this means for the connections between some really specific policy problems and basic research on network dynamics one of the policy problems that I'm thinking about most recently is for example the spread of beliefs about vaccination behaviors whether or not vaccination is safe or dangerous and you can think about this is obviously has a strong normative quality to it and doesn't just take on the dimensions of a single behavior or single belief or idea spreading but actually has a kind of creative responsive element to it where people are sort of learning from each other and then developing their own ideas and so forth and when we think about how these sort of beliefs that particularly ones that are externally seem sort of odd ever get traction or grow in a population these dynamics help to give us a way of thinking about how these dynamics evolve and then from a social engineering perspective what we might too think about in terms of being able to alter these dynamics collectively so that's the, I want to sort of stop there with talking about the collecting dynamics there's more to say in terms of lots of people have questions about individual behaviors and so forth but we can sort of we can get into that if you want to see more but I'm willing to take questions now and hear about people's reactions so thank you very much yes of course I think they can even hear me here yeah yeah but both will hear me it works so the example you gave it converged on the name John and I was wondering if that was a general pattern where people converged on popular names so maybe like conventions feedback on conventions kind of thing yeah so the names actually differed and we did some analyses on which names were selected there was no significant bias in one name versus another the name John showed up in one of the other networks with a male face the name I think Brad was chosen the name Steve was chosen another one so in terms of a popular name yes I think that there's some sense that there's a kind of a cultural reference here in the sense that the name Hussain didn't show up in the names or if it did show up it wasn't selected but I think that the interesting those two interesting points about that one is that just simply having John in the network in the set of choices available didn't mean that the network would converge so it's not sufficient to have a name like that or to generate convergence nor is it necessary because the other names could be chosen when we think of those popular names we can think of yes they may show up but if they're neither necessary nor sufficient they don't really tell us much about how these social norms emerge I can see Matt so Matt can I ask you that's a very interesting point can I ask you a follow up about that because I want to understand whether the dynamics you're talking about are driven by something like modularity in the actual topology or is the group structure that he's discussing emergent like the groups that we saw here that weren't a function of the topology in the sense that the groups weren't defined by more or less connectedness they were really just an emergent function of patterns of interaction right exactly yeah I agree entirely and in fact there are some very interesting results on more modular networks which as I mentioned before all of these three topologies eventually lead to you know an infinite time eventually lead to convergence the question we're interested here is the question of time scale but yes it's true that there's some very nice results by that author I know the ones by other authors thinking about modular networks more generally and so can show that with high levels of modularity you actually can lose the capacity for convention formation which is the other thing I think you're saying just that you actually get stable heterogeneity across the population yes absolutely oh that's exactly right that most of the theoretical work on the coordination problem has focused on this question of when you say bias we tend to think of that as either payoff dominance or risk dominance so in the vein of the harsh on result solution and most of the work here has been done on risk dominance you can think of like ellison here and a lot of the really interesting work by young others who have sort of run off of that that work has really emphasized the rate at which the risk dominant solution is chosen in the population and you're absolutely right that work has emphasized the value of spatial connectedness for that process this is where I tried to emphasize the value of obviously equal options but also innovation which really adds this sort of diffusion component into the process where it's just not there in the same way if everyone has the same set of options available to them at the outset okay sorry I think that's a fascinating point and I've done some work on this question when it was applied to the problem of collective action and the way that we think about that in collective action context is something like coalition formation if you separate a smaller group in the population allow them to interact you can solve coordination dilemmas that actually exists at a larger scale just through the reduction of cost of interacting I'm not sure that modularity solves that problem however because in the way that you're thinking about it modularity would need to somehow be dynamic you need to have a small group and then in order to go interact with the rest of the population then you need to sort of subsequently have a different network structure for interaction so you need to have a kind of sequential network of some sort where it was initially modular and then somehow more connected in order to get the initial critical mass exposed to the population and the question of in some ways critical mass which is the question that it seems like you're getting at is incredibly interesting in this case because for example all of the faces that were female faces solicited female names and all of the male faces solicited male names which we saw very conspicuously in those models I played or in the movies I played but we also ran experiments trying to see if we could generate using kind of a critical mass approach generate norms that were completely culturally unexpected a female name for a male face names that violated racial expectations for the color of the face and names that violated both gender and racial expectations for the face it turns out that you can actually do this fairly effectively even in the fully connected network there is actually a kind of a critical minority side if you get a minority above a certain amount which turned it's about 25% you can spread an unconventional norm that then will become the sort of locked in social convention I can just show you a quick example of that if I said what name would be selected for this particular face you can see here that you can see the name in red, green, blue and purple and the interesting thing is the red name wins but at around 5 it's actually not the dominant name so it's not the case that the committed minority was always the most popular name but through repeated interactions the name that actually won was Nia for that face compared with Eric, Sean and Hans and to my mind that gets to the kind of thing you're talking about and I think it's actually one of the more interesting implications of these kinds of dynamics when we think about what this means for social policy more generally the kinds of things that can happen online whether small groups could self-organize and actually generate normative convergence over a population one of the more interesting things to look at also is not just can you generate it from the initial state but can you flip a norm once a norm gets locked in in a population how much work does it take to actually change the equilibrium state from one to another and how flexible is that process once locked in is achieved what we're going to do now is we're going to pin people off and we'll see the rest of you back for the next session thanks speaker great thank you