Can communication across networks of people be optimized to share information, while at the same time lessening the likelihood of information bubbles and echo chambers? In Episode 54, we’re joined by Ida Momennejad and Ajua Duker from Columbia University and Yale University, respectively, to discuss their open access article “Bridge ties bind collective memories” which was published with Alin Coman on April 5, 2019 in the journal Nature Communications.
Websites and other resources
- Ida‘s and Ajua‘s websites
- Ida‘s and Ajua‘s Twitter feeds
- Alin Coman’s lab website
- The story used in the experiment
Ida’s workshop on graphs and collective memory:
Patrons of Parsing Science gain exclusive access to bonus clips from all our episodes and can also download mp3s of every individual episode.
Patrons can access bonus content here.
Hosts / Producers
Doug Leigh & Ryan Watkins
How to Cite
Leigh, D., Watkins, R., Momennejad, I., & Duker, A.. (2019, July 23). Parsing Science – Collective Memories. figshare. https://doi.org/10.6084/m9.figshare.9247865
What’s The Angle? by Shane Ivers
Ida Momennejad: An important aspect was how much there is overlap in the information that is shared over time in the network, and how much there is the diversity of information.
Doug Leigh: This is Parsing Science. The unpublished stories behind the world’s most compelling science as told by the researchers themselves. I’m Doug Leigh …
Ryan Watkins: And I’m Ryan Watkins. Today in episode 54 of Parsing Science, we’ll be joined by Ida Momennejad and Ajua Duker from Columbia University and Yale University. They’ll discuss their research into how communication across networks of people can optimize information sharing while diminishing the likelihood of information bubbles and echo chambers. Here’s Ida Momennejad and Ajua Duker.
Momennejad: Hello! I’m Ida Momennejad. I’m currently a research scientist at Columbia University, and I used to be a postdoc at Princeton when I collaborated with Ajua and Alin on this project. I’m originally from Tehran, Iran, and I did my undergrad in computer science, software engineering, and then I went on to the Netherlands and did a graduate degree in philosophy of science and philosophy of mind, and from there I went to Berlin and did my PhD in computational neuroscience and psychology. The entire sort of background of mine is very much centered around building computational models of how people think or designing experiments to understand how their brains represent the future in particular, and what is the relationship between memories of the past and representations of future, how these sort of representations inside the brain get updated, and how they lead to decisions in people.
Ajua Duker: And my name is Ajua Duker. I’m currently a second year PhD student at Yale, I’m in the social psychology department. So after I graduated college, I ended up being a research assistant at Princeton, which is where I met Ida and Alin who I had the awesome opportunity to work with. And there, I did work with Alin on stuff having to do with memory, stuff having to do with retrieval-induced forgetting, and we did a lot of social network studies which is sort of the paper that I worked on with Ida as well. Just a little bit more about me as a person, so both of my parents were born in Ghana but I was born here in the US, but I have a lot of ties to home, I have a huge Canadian family that’s a big part of my life. And then also I dance still even as a grad student, but I used to train in classical ballet pretty much all through my childhood and high school.
Leigh: Ida and Ajua’s study is an application of a theory regarding the spread of information among social networks popularized by Stanford University sociologist Mark Granovetter, in 1973. He proposed that when two people are interpersonally connected, their groups of friends are more likely to share common memories of their experiences. The model explains how close-knit clusters of friends can be formed both by the strong ties among groups of friends as well as by weak ties that bridge between two or more clusters of friends. We began by asking Ida for an example of how such groups might share common memories of shared experiences in everyday life.
Momennejad: Rich ties is the term that we use because we find it a little bit more useful, especially in our condition a bridge tie has only one connection. So it’s like the connection that if you cut it, two clusters are gonna get separated from each other — just like a bridge that connects two sides of a sort of let’s say a river. So imagine you and your friends go and see a movie, and another group of people go and see a movie, and let’s say that Ryan’s groups of friends are not connected to Doug’s groups of friends, and Doug and Ryan are connected to each other. If immediately after the movie Doug and Ryan text to each other what they think of the movie and then they talk to their friends, as opposed to first they talked to their group and then talk to each other, the two groups of friends that are only connected to each other through Doug and Ryan are going to be more similar to each other. So with the weak tie is Doug and Ryan, because they’re connected to each other but their friends are not friends with each other. So each of them has a cluster of friends, but they are the only connection between those two clusters. So if they talk first to each other and then they talk to their own groups, their groups are gonna have more similar memories of the movie, as opposed to if they first talk to their groups and then talk to each other. But weak tie has other notions, especially social psychology it could be that the connection they don’t talk to each other that much, so there are other notions of weak ties. That’s why we wanted to move a little bit from this Granovetter’s initial notion that has two meanings in principle. One meaning of it could be that it’s a rare tie between clusters that are not otherwise connected, and another meaning could be it’s a tie between people who don’t talk to each other as much, or a combination of the two.
Set-up of the experiments
Watkins: Next, Ajua explains how she and Ida set up their experiments to mimic these kinds of within group and between group relationships that happen all around us.
Duker: For this study, we wanted networks of 16 people, and so the way that they are all structured is that there are four clusters of four people within the larger 16 network. And so what Ida was saying before about weak ties, what that sort of looks like in our study paradigm is that you’re mostly having conversations within your smaller cluster. So everyone who did our lab study they had four conversations. Basically, everyone read a short article, and there are about 30 different independent items of information within the article. We assess people’s memory of the article at the very beginning. So they initially read the article, and then they were just asked to remember it pretty much immediately. So they just wrote down as many items as they could remember as possible, and then they were basically told to have conversations about the article themselves. So the way that this happened was that we brought all these people into the lab, and we had them sit down at computers, and we had them engage in sort of chats. Basically, it’s a live chat platform, the conversations that people were having were kind of analogous to how in the early 2000s people did Instant Messenger, just like really quick instant chats, and it kind of mirrors sort of Facebook Messenger or like DMS on Twitter, but basically you just see this chat happening. And so after they engaged in their memory task, they then had these chats, and we’ve instructed to again remember with someone else. So they have these conversations about what they just read about essentially. Now three of the four conversations happened between other people in your smaller cluster, but one of the four conversations happened between clusters. Basically, you have the smaller network of people that were sort of talking to one another that are like tighter knit, and then you have one set of conversations across the whole set happening between clusters. So we manipulated whether people had that sort of between-cluster conversation first or last.
Origin of the story used as an intervention
Leigh: The article that Ajua mentioned, which participants read in their experiment, could be said to be something of a rather unusual short story. Weighing in at just 384 words, it’s a tale of two boys who skip school and visit one of the boy’s houses. If you’re interested in reading it, you can do so at: www.parsingscience.org/e54. We asked Ida to tell us about the story’s origin and uses.
Momennejad: Okay, so there is this paper by Anderson and Fisher in 1978, and it has these different stories. And typically what happened was that participants would be split into two groups and each group was given a different schema, and depending on whether you heard, for instance, this is a rubbery or people are visiting a place, he would remember it differently. So that was the original idea. The whole aim was to see if schema processing would influence encoding and retrieval of the same memory of the same stories that people heard. The conclusion was that has some effect at retrieval and at encoding, so it wasn’t just the encoding. So we intentionally chose a story that we knew in the literature had led to divergent memories between different groups of people, with the difference that we didn’t give different clusters different schema, which we had thought about in the days of the design, but we allowed the cluster itself, and the information overlap, and information diversity to create different kinds of schema themselves. So it wasn’t as binary schema, as you know, robbery versus house visit or something like that, but it was more of the more emergent scheme of just the kinds of things that these four people happen to remember as opposed to the 16 people in total. So an important aspect for us was that if different clusters are focusing on different aspects of the story; for instance, it could be that different approaches to the same story might be evoked in different participants. The less there is information flow and diversity of information in the entire network, the less similar the memory of the clusters would be.
A darker side to collective memories?
Watkins: While clustering can enhance the prevalence of shared memories among a group of friends, Doug and I wondered if there might also be a darker side to such collective memory. Here’s what Ida had to say about the question.
Momennejad: So one of the important aspects of the other work that I do in neuroscience is that we make computational models of what happens inside a brain when activating particular memories is going to strengthen some other memories and weaken memories that are related but are not mentioned. And the neural mechanisms of that are interestingly also a network-based phenomenon. So let’s say you remember your sister, you don’t just remember your sister, that memory also brings up or makes it more available to think about maybe her children, or your mother, or your aunt. The more you’re thinking about your sister, that memories that are very close to her will get strengthened too, but things that are related to her but a little further will get a little more forgotten. This is just a phenomenon that happens because it could be useful for your daily events, that things that gets more retrieved because they’re more in your environment, or they’re more salient, or they’re more rewarding, or they’re more primary to your survival, will get to hire more prioritized representation. So I think that the most important purpose of memory broadly is to be able to predict, in a useful and in a sort of practical way, predict the future. So the predictive nature of memory would require that it’s useful for me to remember things that are important to my social interactions, and it might be useful for me to synchronize with them at least to some extent to know where my social groups’ memories are. But if a big community or big network of people is too clustered, what happens is that they will disagree on the memories of the same event. And I think that polarization, especially in politics, one of the things that it does is that there are two different versions of reality after a while, and these two versions do not hear each other because there’s not many bridge ties, and maybe the bridge ties didn’t talk early enough, so the information has been integrated into some models of the world, and it’s being replayed in ways that are meaningful to each of these two clusters. And therefore, it’s being reinforced within the clusters, but it’s a little perhaps too late, or it will take a lot more energy and effort in order for bridge ties to be able to even transfer what each side’s model of the world is compared to the other.
Leigh: Ryan and I wondered to what extent Ida and Ajua might feel that the social position of bridge ties might have an influence on the saliency of their ability to connect diverse clusters of people.
Momennejad: So in a way there are some people that we allow what they’re saying to influence our memory systems and restructure our memory systems a little more than others, and that depends on our social ties with them, especially given that a lot of human networks are very hierarchical, I think that it would be very interesting to also look at the effects of somebody who’s in a position of authority. For instance, a lot of sort of orators and like famous people, like think about the MLK speech. There are some speeches that’s truly synchronized nations and even people in the world in terms of particular memories or significance of events, in spite of the fact that they didn’t think so before. So some particular network positions if somebody occupies them, so it could be a position of power, trust, or knowledge, or for some people who are religious that are the religious authority, that person’s sort of reminding of them might be more effective. And it could also have a negative outcomes, it could be that the memory or the stories of a person who is in a kind of a dominant position would shape, reshape the other person’s memories. So in general, in many situations we would not take another person’s memory on the same way, but I think there are situations that in fact we take it even more into account, like situations of authority …
Ajua: I definitely agree with you. I think the context is the key factor here, that sort of ties into when your representation is not particularly clear about something. So in the case of speeches a lot, this authority figure or this figure of knowledge is often talking about something with great like painstaking detail about an event or some sort of past occurrence that it’s not particularly likely that most listeners have thought about or engaged in any sort of rehearsal in great detail. So in that case your memory is a lot more susceptible to taking on this new information compared to sort of co-remembering a past event that both parties have experienced themselves.
Policy implications and applications
Watkins: Ida and Ajua’s study illustrates the importance of when people share information across different groups with which they interact. We followed up with Ida to learn about the potential policy implications and applications of their research.
Momennejad: We were interested in seeing, okay, let’s say that a network is extremely clustered already. Is there a particular order of conversation that is going to lead to more convergence than another? A very important thought sort of in the background of all of this was: how can this at some point turn into public policy? And therefore, can we find a way some equation that we can just figure out what’s the particular network based on marketing social media or other computational ways, and figure out who should talk to who first in order for the synchronization and the larger network to work better? So this practical applications and possibility of intervention was very much driving the question: what should be the order of conversations? And it turns out that if the bridge ties talked first, there is a much more potential for convergence both within the clusters, which was the kind of surprising part, and across the entire network. So an important aspect was how much there is overlap in the information that is shared over time in the network, and how much there is the diversity of information. So we came up with this overlap and diversity index, which was very simply the different conversations that were happening simultaneously in the network but between two people at any time. How much overlap was in this information? And how much was new information that was not mentioned in another conversation? It turns out, of course, that if bridge ties talked to each other first, the diversity of information early on is high. If the cluster ties talked to each other first, within the first two rounds of conversation, all of the members of the network have already shared everything they know with each other. But in the case where the bridge ties talk first, the first round of conversation is the same, but within the second sort of round of conversation they’re having a conversation with a larger group of people. So the diversity is high early on, and overlap increases over time. So diversity goes down over time and overlap increases. Whereas, in the case of cluster ties talking first very early on, they overlap a lot in their conversations, and people are very likely to repeat something that happened in the last conversation in the next one, and therefore, the diversity of information all of a sudden increases by a lot, and after everybody within the cluster has talked to each other, they start to talk to their bridge ties.
Application to typical speech
Leigh: Though their study was carried out via internet chat, speech is the primary vehicle by which people tend to interact in life. So given what we had just discussed, Ryan and I were interested in hearing their thoughts on how we might apply their findings to this more common mode of communication.
Momennejad: There is also work about the effect of abstraction versus detailed information about a particular event in how much you are open to otherness. There is some work that suggests that the more abstract things get, the more likely it is that I’d be open to seeing someone who is from another institution but treating them as if we are affiliates, as opposed to sort of being very distant. So it’s very interesting how a speech whose aim is to synchronize; for instance, recruit people for war, or recruit people for peace, or recruit people for civil rights, or recruit people to actually notice that, you know, the majority of the population is showing, for instance, racism. If that is the kind of speech that needs to happen, then this dynamic of using details versus abstractions is very interesting and it’s an art. The knowledge that we are collecting right now is a knowledge based on analytic skills and experiments, and making it explicit, making it sort of conscious and measurable, but people have been using this. So something that is interesting to me is the kind of knowledge that is the know-how of this, someone figures out when to use details, when to use abstractions, when to move from story to theory, to back, and when to excite, and when to empathize. So it’s very interesting that there are different modes of knowing these things. For some of us knowing that by means of measurements, and for some people is know-how. They have been somehow the entity that they are, the human that has been trained to use this algorithm, their brain is using this algorithm, but if probably you ask them about what the algorithm is, I’m not sure if they could tell you.
Duker: And I think there’s something to be said about how you can leverage emotional salience in your communication to affect the information that people remember. So basically, like, either rhetorically or directly referring to so emotionally salient events as a way to tie what’s being communicated — the actual message, so maybe some sort of political message, maybe it’s about anti violence, etc. — tying that to an emotionally salient event, it will also probably increase the degree to which people remember what’s being said, and how they’re thinking about the event itself.
What if there were more than four rounds of communication?
Watkins: Ida and Ajua’s study was designed to end after the fourth round of communication among participants, with some groups just getting information for the first time from bridge ties. This led Doug and I to wonder what they believe might have happened had the conversations continued beyond four rounds.
Momennejad: So round three is where they are similar to each other, and the reason for that is that in round three every person has had three conversations by this point. The three conversations that has happened, there’s a much more potential for overlap between what are the items that are available, especially given there are many items — my 30 items — like there is a, you know, by around three, everybody has had three conversations. However, the conversation in round four is extremely different for the diversity index, because in one group it’s not really giving new information, because now they’re talking within their cluster and that information already has gotten there. So round four is where they differ again— in the diversity and overlap both — but I think if it went beyond four rounds, I think that you would see that the diversity will gradually decrease for the bridge ties last, the Cluster Ties First, and the Bridge Ties First condition, their sort of diversity is already very low, probably might go a little lower, but probably their overlap will go higher. I can maybe explain a better at what I was just saying with an example.
Leigh: We’ll hear Ida’s example after this short break.
Ad: Science Pods
Example of how bridge ties might operate in a conversation
Leigh: Before the break, Ida was about to provide an example of how bridge ties might operate in a conversation among a group of people.
Momennejad: Imagine that the four of us, we have different orders of talking to each other. Let’s say that I talk to Ajua, Ajua talks to Ryan, and Ryan talks to Doug, then my information reaches Doug with a distance of three, but Doug’s information reaches me with a distance of infinite because it never reached me, no one ever talked to me after talking to Doug. Now let’s imagine that the first round of conversation is me – Ajua, then it’s Ajua – Ryan, then is Ryan – Doug, but then there is another round of conversation which is Ajua directly to Doug. So this means that my information reached Doug in two different ways: once with a distance of three, and another time with a distance of two. It was introduced later so maybe it was a little more discounted, and there is some overlap between these two conversations. However, it might be that Ajua just conveyed something that I had said that Doug hadn’t heard already from Ryan, and then at another point, let’s say that I talked to Doug personally, and what happens is that now there are three different ways in which my information from my personal memory has reached Doug by the end of these conversations. While at that point is the first time I receive any information from him, he has received so much influence from me that many of the things that he would tell me would already have come from me.
Network analysis techniques
Watkins: Ida and Ajua’s study tracked the sharing of information both within groups and across groups in order to track how that sharing influenced the collective memory of the groups. This required sophisticated network analysis techniques, so Doug and I asked Ida to tell us more about their analytical approach.
Momennejad: So there is a way that I made equations that take very simply account of this, and I used an algorithm that I used for the brain in my other studies here, and that algorithm is called the successor representation, and the simple thing that it does is that it takes into account multi-step dependencies. You consider all of the multi-step connections, and you have a mean of them when you want to see what is a particular effect. You sum the different ways in which particular multi-step connections happen, and you discount them by their distance, and I think that is a very interesting sort of an important aspect which was using an analysis that I’ve done, and modeling that I’ve done in other studies for within neurons, but now applying it to how conversations propagate, because it’s simple things about propagation in networks or in graphs. And somewhat, not entirely, so it has some other measures like recency and primacy of memory, so it’s not entirely like that, but to some extent, you can think of it in terms of communicability distance on graphs which is: if you start a random walk from some points on a graph, how often will you end up on another point? And you will get there from different paths, of course, sometimes directly, sometimes a longer path, etc. On average, how many times will you get from point A to point B on a graph that let’s say has like 30 different nodes if you just infinitely have walks there. And so that’s the kind of similar mentality. By round 3, we are at a point where the echoing and the propagation of information is more similar, but by around 4, because all of a sudden from every cluster people talk to people from other clusters, all of a sudden they get diverse information, all of a sudden they get information that is not reaching the other members of their clusters, so they go further from each other. So even within cluster their memories become more distant from each other, for the Cluster Ties First. But in the Bridge Ties First, they’re just talking to the next person, and they already have reached information about what I know has already reached them in previous conversation. So, well, not a lot of new things are happening.
Leigh: The participants in their study tended to communicate differently when they did so when interacting within their in-groups, as opposed to when doing so across social groups. We were interested in hearing Ida and Ajua’s thoughts on why this might be the case.
Momennejad: There’s different ways of thinking about it. One particular idea is that if someone is more similar to you, or if you consider them more similar to you in one way or other, you might be taking up the information that is coming from them similar to information that you would be receiving yourself. And if you perceive them to be more different from you for whatever reason — artificial or not — you might be perceiving the information that comes from them as information that needs to be checked more, information that requires more vetting. It could go back to this kind of primal idea that someone from my tribe is going to give me information that is useful to my survival, and someone from another tribe or someone that I perceive as a rival, might give me information that are detrimental to my own acquisition of value and my own survival. So it could be that that’s the primal link that makes a distinction, but on the other hand, there are benefits to vetting information that comes from people that might look like they’re not my allies, and the benefit of that is that there would be a lot of people who’ll try to manipulate you. and a very important aspect of human interactions is basically using deception, and it would be, of course, always beneficial to detect deception. So there is an aspect of it as that is rational, which is I should make sure that somebody who seems like they might not be my ally is not deceiving me, but there is another aspect of it that might be, all right, who do I consider an ally might be now on the basis of sort of an unjust social algorithm. So I think there’s two aspects of it that are both important.
Duker: I would just like to add that the extent to which you may be perceiving information from someone who’s not a part of your group, so not an ally — like all sensibly not an ally — the extent to which you may perceive that information as potentially deceptive also varies as a function of sort of how many conversations you’ve had within your in-group about this information. So that ties back to our paper right now where people are sort of having conversations across their bridge ties, and they’re more likely to integrate that information that is not technically information from members of their smaller cluster. They’re more likely to do this if they have that bridge tie conversation earlier, then they do later further, for that sort of reason, right? Because they’ve rehearsed this information. So the similarity in terms of the representations that people have is quite high within the cluster, and then they’re having three conversations about that before they finally have a conversation outside of their cluster. At that point, their idea of what happened is quite solid, and so now they’re encountering this new information, and they’re way more likely to engage in these sort of fact checking processes, and also likely discard that information as invalid.
How to diminish our information bubbles and echo chambers
Watkins: Given the polarization of the American electorate leading into the 2020 presidential election, we were interested in learning Ida and Ajua’s thoughts on what we might do to diminish our information bubbles and echo chambers even when we might not realize that they constrain us.
Momennejad: The first sort of immediate implication for the current situation is if you think that there are some things about political affiliations or political decisions that are important to you and you think that there would be different ways that other people from another party think about it, you should start very early on and talk to as many people that are on the other side as you can, and in pleasant ways, so that they can sense a feeling of affiliation, and they can put up, have the opportunity, or give you the opportunity to synchronize memories together, or synchronize ways of looking at the actual problems together. So I just, like, that’s the fastest most immediate thing.
Duker: Right, if you have interactions with people who are less like you, who sort of have maybe a different vantage point that they’re coming at whatever your information is from, those could be the most valuable sorts of conversations because they could shape the way that you’re looking at problems, they can help you correct for blind spots that you may have. The earlier you have conversations with people who think differently from you, the better your product will be — whether that’s knowledge, whether that’s literally a physical product if you’re producing something — because at that point, you can shift your fundamental ways of thinking about something in a way that may make whatever you’re making, whatever you’re involved in, more accessible more beneficial to more people. So the more people that you talk to that are like you, the more siloed, more specific, and less generally accessible whatever you are trying to produce will be.
Links to article, bonus audio and other materials
Watkins: That was Ida Momennejad and Ajua Duker discussing their open access article “Bridge Ties Bind Collective Memories” which was published with Alin Coman on April 5th, 2019 in the journal Nature Communications. You’ll find a link to their paper at: www.parsingscience.org/e54, along with bonus audio and other content we discussed during the episode.
Leigh: Reviewing Parsing Science on Apple podcasts is a great way to help others discover the show. If you haven’t already done so, head over to: www.parsingscience.org/review to learn how to do it, or if you have a comment or suggestion for future guests or topics, visit us at: www.parsingscience.org/suggest, or leave us a voice message toll free at: 1866XPLORIT. That’s 18669756748.
Preview of next episode
Watkins: Next time, in episode 55 of Parsing Science, we’ll be joined by Zuzana Musilova from Charles University in Prague. She’ll discuss her research into the unique way that some fish in the deep oceans darkness may be able to see in color, even though they’ve previously been thought unable to do so.
Zuzana Musilova: They go deeper and they live only deep water zone. Then they don’t use the cones anymore because they are not, you know, sensitive enough. And then they start to express the rhodopsins also from the center of the light spectrum.
Watkins: We hope that you’ll join us again.