Parallel systems for social and spatial reasoning within the brain's apex network
October 5, 2021
April 27, 2021
All Captioned Videos CBMM Research
What is the cognitive and neural architecture of core reasoning systems for understanding people and places? In this talk, we will outline a novel theoretical framework, arguing that internal models of people and places are implemented by two systems that are separate but parallel, both in cognitive structure and neural machinery. Both of these systems are anatomically positioned at the apex of the cortical hierarchy, and both interact closely with the medial temporal lobe declarative memory system, to update models of specific familiar people and places based on experience. Next, we test foundational predictions of this framework with a human fMRI experiment. Participants were scanned on tasks involving visual perception, semantic judgment, and episodic simulation of close familiar people and places. Across the three tasks, conditions involving familiar people and places elicited responses in distinct but parallel networks of association cortex, including zones within medial prefrontal cortex, medial parietal cortex, and the temporo-parietal junction. Lastly, we address the question of how these systems emerged in evolution. By assessing fMRI responses in nonhuman primates viewing images of familiar and unfamiliar animals and objects, we identify subregions of medial prefrontal cortex with a similar profile of functional response and anatomical organization to human social reasoning areas. These results indicate that the cognitive and neural architecture supporting human social understanding may have emerged by a modification of existing cortical systems for spatial cognition and long-term memory.
WINRICH FREIWALD: Hello, everyone. I'm Winrich Freiwald. I'm hosting today's seminar.
I'm actually super excited about this. So Ben is presenting. I don't think he needs much of an introduction here.
He did his PhD work with Rebecca and with Nancy, was wildly successful. I was lucky to be in his thesis committee and snatch him away for a postdoc in my lab. Ben has been doing a lot of thinking and experimenting on high level social cognition circuits and some other questions.
And today's subject I'm particularly excited about is an idea that Ben came up with a couple of years ago. Being his usual very efficient and competent self, he's already done the experiments human fMRI experiments. And I think it's an exciting way for us to think about the neural circuits of social cognition, so tying a couple of different directions together that are normally viewed separately,
As Chris was already alluding to, this is already quite far progressed even though Ben has not been working on it for such a long time. It would be really, really helpful to get your opinions and feedback on this and questions throughout because it's really going to help us to turn this into a final product to send out soon. So without further ado, Ben.
BEN DEEN: All right, thanks. So it's nice to be "back," quote unquote, giving a talk at MIT. This is my first time giving a talk here since my thesis defense.
So I'm excited to tell you guys about what I've been working on the past few years. And since this is a talk about understanding familiar people and there are a good number of familiar people in this audience, I thought I'd start with a story. So this is the story of my first hiking trip with the Saxelab.
This is back in the summer of 2010, the summer before I entered grad school at MIT. Rebecca invited me to join the lab on a hiking trip in Franconia Notch in the White Mountains. And the plan for the first day of the hike was for the lab to split up into multiple groups and take separate paths all converging on the Circles Lake, Lonesome Lake for lunch.
I was in a group with then-postdoc Marina Bedny, grad student Jorie Koster-Hale, and two undergrads, Jackie and James. And when we got to the top of the ridge, we had sort of the decision point as to whether to take a short trip up to the nearby peak, which would have a nice view or, whether to just go straight to the lake. And at this point, Jackie was having some eagle pain. She didn't really want to go up to the peak.
But Marina, Jorie, and I were quite excited about going to the peak. So we decided we would just leave her at this intersection and make a quick trip up and then come back. There's us at the peak.
And this seemed like a totally reasonable decision at the time. In retrospect, we've come to understand that this is apparently not considered good hiking etiquette to leave an injured, inexperienced hiker alone on a trail. That said, we eventually all made it to the lake, and all this well in the end.
On the other side of the ridge, there were some other issues going on. So then grad student [INAUDIBLE] was leading a team consisting of undergrads Hannah and Swetha, and they ran into some more serious issues. They got lost a number of times. And I guess apparently when [INAUDIBLE] finally found the path and was confident that she had the right direction, she sort of went full steam ahead and arrived at the lake without the undergrads-- much, I assume, to Rebecca's terror.
But luckily, the undergrads came by a few minutes later, and ultimately, we all made it to the lake and had a nice lunch there. The next day of this trip was a smaller group pike, and I spent most of this hike chatting with Rebecca, sort of a long conversation on ideas about what I would research in grad school. There was a moment that sticks out in my mind when we were talking about the idea of writing a popular science book-- so a book about our research for the audience.
And I expressed that I might be interested in doing this, and Rebecca seemed kind of taken aback by this. She was pretty surprised that a serious scientist would want to do this. And she asked me why I would want to, and I said, well, I sort of like the idea that after I die, people might actually still think about my ideas and still care about them.
And Rebecca responded without missing a beat. Why do you care what happens after you die? And that once stumped me, and I think I still don't have a great answer to that question.
OK, so what I just told you was a combination of episodic memory, sort of pulling events from my own experience, and narrative, incorporating perspectives of others on this trip. And the story was thoroughly structured in terms of familiar people and familiar faces. And in this talk, I'm going to argue that not only do the brain and minds have specialized systems for understanding people in places, but they're actually parallel in architecture, both at the cognitive and neuro levels.
That's a brief outline for the talk. I'll first go into some background and theoretical ideas and then present two studies, both fMRI, of one in humans, one in macaques, that I used to test these ideas. So the CBMM core reasoning module posits that there are at least three core cognitive systems for understanding different domains in the world-- one for understanding other people, one for understanding places or spatial layouts, and one for understanding physics or interactions between objects. I'm going to ask two big questions about these core systems in this talk.
The first is, what is the architecture of these systems? Is it just sort of three separate systems as I've shown here with no meaningful relationships? Or is there some meaningful structure?
And second, what's the relationship between these core systems and long term memory? In the domain of spatial cognition, there's a rich history of studying spatial reasoning as integrated with long term memory. Spatial reasoning is usually studied with respect to places that participants have some familiarity with.
In contrast, in the domain of social cognition, there's been decades of research on how we understand others' minds and how children and adults understand others' minds. It's really been-- used tasks that place very minimal demands on long term memory and that involve reasoning about people that the participant doesn't know and was just introduced to. In this talk, I'm going to suggest that maybe we should be studying social cognition in a way that integrates long term memory more closely.
The research program that I just described has led to a model in which we understand other people using a theory like internal model of others' mental states and how they give rise to behavior. And this has been studied not only within cognitive science but also within cognitive neuroscience. This is a research program that Rebecca has pioneered among others, and it's led to a relatively standard way of defining regions of the brain that may be involved in theory of mind by using story comprehension tasks, tasks in which participants read stories and answer true or false questions, and then comparing responses to stories for which the answer requires understanding a false belief of an agent, the stories when the answer requires understanding a false photograph for physical representation, and their mental content.
And when we compare responses to these two conditions, we see a highly reliable, robust pattern of response with activity in the temporary parietal junction, superior temporal sulcus, superior frontal gyrus, medial prefrontal cortex, and medial parietal cortex. These regions fall within a broader zone of cortex that's been referred to as the apex or default mode network. This is a set of anatomical regions that also exist in the macaque and are strongly anatomically connected in the macaque.
In this talk, I'll use the term apex network, which I prefer because it refers to sort of concrete anatomical hypothesis, that these regions are positioned at the top of the cortical hierarchy. So you can understand the cortex in terms of anatomical connectivity as a rough hierarchy where regions near the bottom are sort closer to the sensory motor periphery, regions near the top are further from the sensory motor periphery closer to the limbic system and hippocampus, and the regions that I'm showing here in dark colors seem to be in terms of anatomical connectivity in the macaque and functional connectivity humans positioned at the top of this hierarchy. So since the original arguments for parts of the brain that were selective or preferentially involved in theory of minds, various folks in the literature staged counterarguments.
So I'll mention a few of these. One was that these regions seem to be perhaps more broadly involved in semantic memory or semantic comprehension. Another was that these regions might be broadly involved in episodic recall and future thinking.
So on the left, I'm showing you responses to recalling specific memories, on the right to imagining future events. And from the start, there was some sort of basic methodological and logical issues of these arguments. One is that all of these results used group level analysis of fMRI data in which data from multiple specific individuals is combined in a stereotactic space and averaged.
And this is a technique with relatively poor spatial resolution and a technique that's not really great for making arguments about functional specialization because potentially not overlapping regions can be blurred into overlapping. Second issue is that these contrasts very often involved some social content. So a semantic memory contrast might involve reading a sentence or a paragraph that very often would involve a human doing something. And episodic contrasts have the issue that almost all of our episodic memories had some social content. So It's somewhat difficult to dissociate episodic and social responses in these studies.
More recently, the field of memory research has sort of progressed a bit in this sense and has started using individual subject analysis. I'm going to highlight a few studies from Randy Buckner's lab that have pushed this line of work. This study found that the apex network seems to be comprised of two parallel interdigitated that some networks with zones across all of the sort of anatomical regions that are part of the network. These regions are strongly functionally coupled within network but have basically no functional coupling across the two networks.
And a subsequent study led by Lawrence Nicola found that this dissociation in terms of functional connectivity maps quite nicely onto a functional dissociation. So the regions that I'm showing in pink here responds to a theory of mind task, the one that I described earlier, and don't respond to an episodic projection task that involves answering questions about yourself in the past or future relative to yourself in the presence. OK, so I should note that the Denicola and Buckner study don't really make a strong argument that that theory of mind and episodic projection are the true functional specialization of these networks per se.
That being said, I'm going to consider the hypothesis that this is the sort of correct functional specialization and the right way to think about the function of these networks, and I'm going to ask, for a system involved in theory of mind, should we really expect that it's dissociated from episodic processing or from long term memory more generally? I'm going to make three arguments. The first two are a prior, the third is an empirical argument.
First one is that human episodic memories are thoroughly structured by social content. So I would argue that you essentially can't come up with an autobiographical memory that doesn't have social content, by which I mean specific familiar people, their internal states, and their actions and interactions. If you consider the memory that I recounted before, if you take away the social content-- [INAUDIBLE] excitement at finding the path, Rebecca's resulting terror-- there's not much of any content to the memory. It's sort of just a group of people wandering around in different parts of space for no apparent reason.
Argument number two is that mental state reasoning, particularly when it's applied to familiar people that we have knowledge about, relies critically on information from long term memory. So if I were to see this book in the bookstore, what would my response be? What would my emotional state in response to this be? I'm going to ask one of you guys to respond about this. Please respond so that I don't just--
AUDIENCE: I think you'd be annoyed at my hypocrisy.
BEN DEEN: Yes, I'd be annoyed. At the very least, I'd be surprised. And note that making that inference relies critically on long term memory, right. So it relies on the knowledge that Rebecca is not interested in writing a book.
AUDIENCE: I dispute that memory, by the way.
BEN DEEN: What was that?
AUDIENCE: I dispute the memory.
BEN DEEN: You dispute the memory? OK, that's interesting. I'd be curious about that.
I was wondering if you would have remembered that. In any case, the inference that Rebecca just made relies critically on long term memory, right? It relies on the knowledge that Rebecca does not want to write a book and also the knowledge that I know that Rebecca doesn't want to write a book.
And the last argument, which is empirical, is that representations of familiar people, person identity, have been argued to exist in a network that is at least anatomically quite similar to the social cognition theory of mind network that I just described. So this is from a series of studies from Diana Tamir and Mark Thornton showing that using multivoxel pattern analysis and fMRI, spatial patterns of activity in these regions can be used to discriminate which specific person a participants is thinking about. All right, so this is a highly speculative slide. Just take it with a grain of salt. I'm not going to directly test any of the ideas in this slide in the data that I'll show you.
But for the sake of concreteness, I want to provide sort of an idea as to what the model of familiar people that builds off of our familiar theory of minds model might look like. So this is sort of motivated by just a computational level consideration of what information about other people do we need to have in order to make these sorts of inferences. And I think what that would look like is something like theories of mind. So if we would have an abstract theory of minds that applies to anyone in the world, we would Additionally have internal models of specific familiar individuals that would incorporate sort of a set of beliefs, emotions, and desires as well as actions.
So my internal model of Rebecca might include beliefs like her beliefs about functional specialization of the RTBG, about theory of minds development, the fact that she's not interested in writing a book, and a collection of actions, like she went to MIT, published some landmark papers in the neural basis of theory of mind. All right, so I've argued that we should perhaps expect theory of mind to be closely integrated with long term memory in general and maybe even episodic memory in particular.
So if you buy that argument, then what should we make of the functional specialization of this other network, this parallel interdigited network? I think you can sort of find the answer by just looking at the stimuli from this study. So this is one example. Where were you the last time you saw a movie?
This is a question that's about the person's past, but it's also sort of a spatial question. It involves some spatial contents, and this is essentially true of nearly all of the past and future questions in the study, whereas the present questions don't have any spatial content. So another hypothesis is that perhaps this network in blue is specialized for spatial reasoning. And there's actually a bunch of reasons from prior work to expect this.
I'll just mention a few examples. Parahippocampal cortex is well-understood to be involved in spatial processing. Damage to this region can lead to topographic disorientation or difficulty with spatial navigation. And this area overlaps the scene-responsive parahippocampal place area, although part of the parahippocampal gyrus that's associated with this functional network seems to be positioned sort of the anterior-most tip of the parahippocampal scene response in an area that's somewhat less well functionally characterized. Same thing for retrosplenial cortex, or I'm saying retrosplenial plus because it includes some adjacent regions as well. Damage to this region can also cause topographic disorientation, and there's a strong overlap with scene-responsive area of the retrosplenial complex.
Medial prefrontal cortex is not classically associated with spatial navigation. However, some recent work showing that damage to this area can impair scene construction or sort of imagining novel scenes. And the specific part of the temporal parietal junction appears to be just adjacent, just dorsal, to the scene responsive occipital place area, and recent work has shown that this region, or at least a region that is anatomically quite similar to this area, is engaged during spatial navigation and during scene memory-- so imagining familiar scenes.
OK, so some would argue that perhaps the relevant association between these networks is not social cognition versus episodic projection, but perhaps both networks are involved in reasoning and long term memory. But the relevant dissociation-- like, for instance, the dissociation that's often cited between FFA and PPA-- is one of concept domains-- so for processing people versus places. We'll flesh out or model of how this might work in a bit more detail.
So at the level of neuroanatomy, we have two parallel interdigitated systems. The system in pink here we think implements generative models of specific familiar people using something like a theories of mind model. The system in blue implements an internal model of specific familiar faces using something like a cognitive map based on surface geometry.
I'm going to argue that these systems both interact closely with the hippocampal formation to update our models of specific familiar people and places based on experience, and such internal models can be used for a number of purposes. They can be used for long term memory, storing information about people and places, and structuring our memories of past events. They can be used for reasoning, understanding and explaining other people's behavior in the present, and navigating space and prediction-- predicting what people are going to do in the future, predicting where you're going to end up as you navigate through space.
All right, I just want to be a bit more concrete about in this model, which is how might these areas be connected anatomically to the hippocampal formation. So for the spatial network, this is pretty straightforward. There are a number of regions that are sort of well-known to have anatomical connections with the hippocampal formation.
Our hippocampal cortex is one of the main inputs to the hippocampal formation via entorhinal cortex. Retrosplenial neocortex has a direct bilateral projection with the subicular complex. And [INAUDIBLE] prefrontal cortex, or at least the agranular parts of mPFC in the cingulate gyrus, receive a direct projection from the hippocampus.
So OK, so there are a number of potential routes for the spatial network. What about the putative social network? Well, there's the same potential projection into mPFC. But what else? Are there any sort of cortical bi-directional routes that might link these two areas?
I would argue that a potential route is through the temporal pole. So the temporal pole has been identified as part of this network. In terms of anatomical connectivity in the macaque, temporal pole has strong connections with both remedial prefrontal cortex and with entorhinal cortex. It actually has as strong production to entorhinal cortex as do perirhinal and perihippocampal cortices. But for whatever reason, this seems to be sort of forgotten in the literature. And a recent study from our lab found a region of the temporal pole with a response to faces over objects and a particularly strong response to familiar faces. I'm also note that in terms of lesion studies, there have been associations with temporal pole damage in the context of Klüver-Bucy syndrome and frontotemporal dementia and social memory deficits.
OK, so we're going to claim that there is a link between the social network and hippocampal formation by way of the temporal pole and also just mentioned there are also sort of all manner of potential subcortical routes that could link these networks, including through the amygdala, thalamus, et cetera. All right, so this is our model. This is sort of a hard model to test, especially with coarse methods like human fMRI, generally speaking testing the idea that a set of regions in the brain implements an internal generative model is not that easy. So what I'm going to do instead is come up with a set of straightforwardly testable predictions from this model and then test those using fMRI experiment.
These predictions will be one that we should see interdigitated responses to familiar people and places across a range of tasks that involve reasoning and memory, stronger responses to familiar versus unfamiliar people and places, overlapping responses to the theory of mind localizer and tasks getting an abstract social cognition and familiar person tasks, a region of temporal pole with a similar response profile as these other regions, and basically functional coupling within person and place sensitive regions across these different domains of cortex but not between. All right, so we'll now get to some data testing these ideas.
So because we think that responses in this network are going to be particularly strong to familiar people and familiar places, we take a somewhat different approach to running the human fMRI study, which is that we sort of design our study sort of in a way that's very much tailored to each individual subject. So we have each individual subject lists their 10 most familiar people and most familiar places. We choose six from that list to use for the experiments.
It's a set of tasks-- I'll describe these in more detail, but visual perception test, semantic judgment task, and episodic simulation or future thinking. We scanned nine participants so far. We're scanning the next one starting next week.
And we collect sort of a relatively large amount of data for an fMRI study. So we scan three times each for 2.5 hours each. We get 10 to 15 minutes of data per condition per task as well as 60 minutes of resting state data and 40 minutes of relatively high resolution anatomical images.
So I'll go through the tasks in a bit more detail. This is a visual task. In this one, the participants are just seeing images and performing a one-back task.
So anytime an image repeats, they press a button. And we have 20 images, distinct images, per category and identity-- so a total of 600 images per subject. The unfamiliar or familiar people and places are yoked to one another.
So the people are controlled for age, race, and gender. Places are controlled on semantic category. There's some important constraints. So the place images don't have any people in them.
I took all these images. So it was quite fun to wander around New York City and try really hard to get images of places that didn't have any people in them. The person object images don't have any discernible spatial structure-- things like corners or edges.
And none of the images have any legible text. The people in object images also don't have any clear context clues. So I tried to avoid pictures from a wedding or that sort of thing. All right, and all the categories are also controlled for mean invariance across images separately within each participant on luminance, contrast, and saturation.
Next is the semantic judgment task. So in this task, we ask for basically ratings of adjectives. For the familiar person condition, we ask for ratings of personality traits. For objects, we ask for ratings of physical properties. And for places, we ask for ratings of properties that relate to spatial or navigational information. And so we basically have the person in the scanner just moving a dot on the screen left and right to choose a rating from 0 to 4.
And lastly, for the episodic task, for this task, we went with a future thinking or imagination task rather than an episodic recall condition because as I've argued, it's rather difficult to dissociate spatial and social processing in the context of episodic recall. Most memories have some spatial and social content. That being said, it's easier to dissociate them in the realm of imagination or future thinking.
So we have people imagine familiar people-- have participants imagined familiar people talking about specific conversation topics, imagine specific objects engaging in physical interactions, and imagine navigating or moving through subregions of familiar places. We have the participants list sort of a set of conversation topics and subregions that they're familiar with before this game. We additionally include two localizers-- the theory of mind localizer that I described and Ed Federico's language localizer, just reading sentences versus listed nonwords.
And to mitigate signal dropout in the anterior temporal lobes, use a multi-echo pulse sequence. We've developed methods for analyzing this data that are sort of tailored to multi-echo data. They involve a simple but robust pre-processing pipeline that gives us high temporal signal to noise ratio across the whole brain, including in the anterior temporal lobes, without any need for a spatial smoothing apart from a single interpolation step that combines motion distortion correction and registration to a functional template. And if anyone's interesting, these scripts are available on my GitHub.
All right, so I'm going to show you what the responses to these tests look like. I'll start with just one example subjects and the comparison of familiar people versus places in the semantic task. So this is what we see. I'll walk you through this.
So within medial prefrontal cortex, you see sort of an alternating pattern of regions preferring familiar people, places, people, places, and people. Medial parietal cortex, sort of similar pattern of alternation between person and place preferring areas and the expected place preferring area in the parahippocampal cortex. Within the temporoparietal junction, one region that prefers familiar people, another that prefers familiar places.
Within the superior frontal gyrus, again, this sort of pattern of alternation between person and place preferring areas. The superior temporal sulcus, we don't see much in the way of areas that prefer places but at least two subregions that prefer familiar people. All right, that was the semantic task. This is the episodic task.
Just flip back and forth between those two. You can see the responses, at least within the apex network, are rather similar. And we again see this sort of pattern of interdigitated regions that prefer people and/or places. This is the visual task. Again, you see a similar pattern. Across all these domains, we see subregions that refer people and/or places.
OK, we're next going to ask how do these regions respond across all the conditions in this experiment. I will do this for the region of interest or ROI analysis. So I'll start by defining search spaces using the Human Connectome multimodal parcellation, which we registered to individual subjects' brains using surface-based registration.
We define search spaces just based on regions that anatomically we expect to be part of the apex network. We then define ROIs as the top 5% of person or place preferring voxels using the semantic task. And we extract responses in independent data. So for other tasks, we just use the full data set from the semantic task and extract it in the other tasks. For the semantic task itself, we'll use leaf one run out cross-validation approach.
We'll define ROIs using all but one run and then extract it in the left out run and iterate across runs so as to extract data and define our ways using independent data sets but still maximize our statistical power. And statistics we'll run using a linear mixed model across runs with subject nucleated as a random effect. All right, I'm first going to just show sort of a cartoon image of what we predict based on the model that I outlined earlier.
So this is the visual experiment. The person conditions are in red. Places are in blue. Objects are in green.
So the visual experiments, we expect sort of a weak response to any face over the object and place conditions and a strong effect of familiarity-- so a much stronger response to familiar faces or people than unfamiliar. In the semantic task, we expect a strong response to the person condition, not the object or place conditions. Same for the episodic.
We expect this region will also respond strongly during tasks getting at abstract social cognition. We don't expect a response to the language localizer. Right, and that was an ROI defined with the person versus place contrast, the reverse contrast place versus person. They expect essentially sort of the reverse pattern of responses, strong responses to the familiar places across all the tasks, and in this case, no response to the theory of minds applies here.
All right, so this is what we see. We'll start off with medial parietal cortex. These are responses of the person preferring voxels with the medial parietal cortex in the visual task.
So we see an effect of people over objects and places generally as well as a fairly strong familiarity effect, a stronger response to familiar faces than unfamiliar faces. The semantic task, again, a strong response to people over objects and places, and the episodic task, again, a strong response to people over objects and places. This region also responds more strongly to the belief than photo condition in the theory of minds localizer. Doesn't respond to much to the conditions in the language localizer.
If we look at place preferring voxels for the medial parietal cortex, we see a similar pattern of response-- so strong response to all of the familiar place conditions and an effect of familiarity in the visual experiment. And I want to emphasize this lack of response to the social episodic condition. To my mind, this provides a strong argument against the notion that this network is specialized for episodic processing per se in favor of the arguments that the relevant dimension driving functional specialization is one of content domain.
Right, we'll now go through sort of the rest of zones of cortex and medial prefrontal cortex, see a similar pattern of response. Replace preferring voxels immediate prefrontal cortex-- again, a similar pattern to what we saw earlier. And this is, to my knowledge, the sort of first demonstration of a place selected for sensitive set of subregions of medial prefrontal cortex.
The superior frontal gyrus-- again, we find regions that respond to familiar people across all of these tasks. And interdigitated areas that respond to familiar faces across these tasks-- again, this is, to my knowledge, the first identification of a subregion of superior frontal gyrus that's specifically sensitive to places. Right, in the STS, social preferring voxels, we see a similar pattern. For place preferring voxels, we see something that's kind of intriguing, a little bit weird.
This area seems to have a weak place preference, although it is consistent across the tasks but also has a strong language response. This is an intriguing response profile, although we don't take this as evidence for our hypothesis because, as I mentioned before, we did not predict a response in this area to a language task. And then TPG for regions preferring familiar people over places.
Again, we see responses to all of the familiar person conditions as well as the theory of mind localizer. And in sort of a nearby region, a place preferring TPJ responds to all of the familiar place conditions. And lastly, sort of as expected, place preferring parts of the parahippocampal gyrus responds across all of these tasks to familiar places.
OK, so getting back to our predictions, we've shown that there are interdigitated responses to familiar people and places across a range of tasks. All of these regions had stronger responses to familiar over unfamiliar faces or scenes. And we see quite strong overlap between responses to the theory of minds localizer and familiar person tasks.
All right, so next, I'm going to ask if there is a region of the temporal pole with familiar person responses. And indeed, we do see a region of the temporal pole that responds to the familiar person condition across all of these tasks. And this is one subject, but we actually have identified this across all of our nine subjects so far. If you look at the responses in this region, they're quite consistent with what I showed you in the other regions.
This area responds across all of the familiar person conditions and to the theory of minds localizer. All right, so yes, there's a regional temporal pole with familiar person responses that could possibly link the social network with the hippocampal formation. And lastly, we want to ask about the functional coupling of these sets of regions in resting state data.
So when we look at resting state correlations, we find as predicted that they're much stronger between regions across this network that have a social preference and within regions that have a spatial preference. They're rather weak, zero or sometimes less than zero, between the two networks. So just showing this in bar graph form-- but within network correlations, they're positive. Between network correlations, they're generally zero or slightly negative.
OK, so again as predicted, these sets of person and place preferring areas are functionally coupled. OK, so to conclude from this study, it seems that the brain's apex network contains parallel subsystems specialized for understanding people and places. We identify a novel component of the social subnetwork in the temporal pole that can potentially link the system with the hippocampal formation.
And we think that these networks are candidate neural substrates for core reasoning systems for people and places. I mean, we think that this may indicate that these reasoning systems have a parallel architecture with cognitively and neurally. And I want to end this part just with some sort of computational level discussion of why should we even expect that there might be parallel systems for understanding people and places.
On the surface, people and places are very different beasts, right? Just intuitively, not very similar. And I think if we take a broader view of the computational challenge that we face in the domains of spatial and social cognition, it's actually not so surprising that we might see parallel systems here.
So for spatial navigation, we're trying to choose movements through an uncertain and constantly changing environment in order to maximize reward in the context of social cognition or decision making or choosing what to do and say among other people that have uncertain, constantly changing internal states. So there are sort of parallel computational problems here of difficult optimization problems with a bunch of uncertainty involved. And I've argued that in both of these cases, we solve these problems by integrating abstract knowledge-- so an abstract understanding of other minds, an abstract understanding of space-- with declarative knowledge about specific people and places.
That sums things up for the human fMRI study. We next want to ask, how did this social network in particular evolve and to what extent is it uniquely human? And to do that, we turn to our close evolutionary cousin, the macaque.
So in terms of functional and anatomical connectivity, something like an apex network has been argued to exist fairly broadly across the mammalian kingdom-- so not just in humans, but also in macaques, which diverged from us around 30 million years ago, and even in rodents, both rats and mice, which diverged from us something like 80 million years ago. That being said, it's much less clear whether there's any social specialization within the apex network in any non-human species. So we're going to study macaques because they're a social species. They recognize one another based on faces and voices, and they arguably have some at least basic ability to understand others' behavior in terms of internal states.
All right, we're going to consider a few potential hypotheses about specialization within the macaque apex network. One is that there is no social specialization. So perhaps they have this network, but it just doesn't respond to social stimuli, isn't involved in social cognition.
The other extreme-- we can imagine that maybe macaques have a fully human-like social network, of the default-- of the apex network. I mistakenly called it default here. And then we can imagine a range of possibilities in between these two extremes. Or macaques have some social specialization within the apex network, but the specialization perhaps hasn't propagated across the full network.
So it's not easy to get macaques to do a social cognition experiments. It's quite hard. You can't just ask them what they think about familiar animals.
But what you can do relatively easily, at least, is show them images. And as I showed in the human fMRI experiments, you can elicit this network dissociation just by showing images of familiar people and places. So we're going to do the same thing in macaques. And in fact, when I say we're going to do, what I actually mean is others in the lab have already done this.
So I was lucky when I arrived in the lab that then-grad student Sophia [INAUDIBLE] was just finishing her thesis in which she showed macaques images of familiar objects and faces. These fell into three categories-- personally familiar faces and objects where the macaque has direct personal experience with the items in question, visually familiar faces and objects, which are images of macaque has seen before but are animals and objects that they've never personally interacted with, and unfamiliar faces and objects. This is a case where the macaque has never seen these images and not interacted with the monkeys or objects.
All right, so when we compare responses to faces versus objects, this is what we see on the lateral surface. So you see sort of the familiar set of face patches along the extent of temporal cortex in the macaque starting from posterior and leading up to anterior regions AM as well as the region in the temporal pole which Sofia discovered in this data set. We also see previously described regions of lateral frontal cortex.
But when we flip the brain around, on the medial surface and medial prefrontal cortex, we see a set of responses which has sort of an intriguing pattern of what looks like an alternation between socially preferring and non-preferring regions across the dorsal to ventral axis within mPFC. And this is a result from one animal. Sophia scanned four animals. So we replicate this results in a second animal, third animal, and fourth animal.
OK, so we next want to ask how these regions in the macaque compare to potentially homologous regions in the human in terms of anatomical organization, response profile, and functional connectivity. So in terms of anatomical organization, there's already sort of a striking parallel, which is that in humans, we see interdigitated responses to people and places along the dorsal to ventral axis, medial prefrontal cortex. In the macaque data set, we don't have responses to places, but we see something like interdigitated regions that do and don't prefer faces over objects.
We sort of want to test if this is actually a reliable pattern in the data or if we might just be sort of reading our theory into noise. So to do so, we split our data in half. We put those points on the surface corresponding to three face over object preferring parts in medial prefrontal cortex.
We're able to find at least three subregions in each animal and each hemisphere and also non-preferring interposed regions. We draw splines across these points on the surface, and then we extract responses. And specifically, we'll show you the face versus object contrast as you move across these parts of cortex.
This is what we see. So we see a pattern in which-- see an alternation of parts of medial prefrontal cortex that do and don't have a social preference. This is in the left hemisphere. We see the same pattern in the right hemisphere. So this seems to be a reliable pattern in the data of alternation across the dorsal to ventral axis.
All right, next, we want to ask how this region responds across the conditions in the experiments. And just to remind you, this is sort of the prediction we have based on the human data that I just showed you. So in humans, we see an overall response to faces over objects as well as the familiarity preference, a stronger response to familiar than unfamiliar faces.
In the macaque, we're going to use basically the same sort of methodological approach that I described earlier. We define a search space anatomically and search for the top 5% of face over object referring voxels within that space and extract responses from those ROIs using a leave one run out cross validation procedure. And these are the responses that we see. We see the strongest response to familiar faces, an effective familiarity-- so the response to familiar faces is stronger than the response to unfamiliar faces and visually familiar faces-- and a strong effect of just faces versus objects overall. This response profile is rather reminiscent of what we see in humans.
All right, and lastly, I'm going to ask about functional connectivity of these areas. So just to remind you of the prediction from humans, in humans, if we were to assess the functional connectivity of socially preferring and non-performing regions of mPFC, we would expect to see functional connections with parallel but non-overlapping areas within the apex network. So do we see anything like that in macaques?
I'm going to show you first a [INAUDIBLE] map of functional connectivity using a socially preferring region of mPFC as a seed. This is what we see. So it's a rather apex network pattern of connectivity. It's connectivity with adjacent medial prefrontal cortex, medial parietal cortex, the parahippocampal gyrus, dorsal SDS, the lateral parietal region 7A, which is sort of intriguingly TPJ-like.
And this is what we see for the non-social seed-- so a seed region that has no social preference. Essentially, the pattern across space is virtually identical, and we actually see no significant differences between patterns of connectivity across socially preferring and non-preferring areas of mPFC. So we essentially think that these areas in macaque mPFC are indeed functionally connected with an apex-like network, but we don't find evidence for sort of full parallel subnetworks across the entire network.
All right, so I've argued that there are socially preferring subregions in macaque medial prefrontal cortex that are similar to human social reasoning areas in anatomical organization, response profile, and functional connectivity. So getting back to our hypothesis, we think that these data strongly rule out the hypothesis that there's no social specialization within the macaque apex network. They're most consistent with the hypothesis of partial specialization, that perhaps this specialization has emerged in medial prefrontal cortex but not elsewhere in the network.
But we can't fully rule out the hypothesis that maybe there is differentiation in the other areas. It's possible that maybe if we used something more like an engaging social behavior task or social reasoning task, we could come up with something like that for macaques. Maybe we would see differentiation elsewhere.
All right, so to conclude, we identified parallel networks within the brain's apex network we argued to be involved in understanding familiar people and familiar places. We identify evolutionary precursors or potential evolutionary precursors to the human social cognition areas in medial prefrontal cortex in the macaque. And generally speaking, we think that this full set of data in both the human and macaque experiments indicate that as humans, our remarkable ability to reason about other people may have evolved in a way that built on top of pre-existing systems for spatial cognition and long term memory.
Right, so that's it. We'll just do some acknowledgments briefly. I want to acknowledge Winrich for his support as an advisor and useful conversations about these ideas.
For the human fMRI study, I want to thank the Cornell CBIC for enabling us to collect all the human fMRI data that I just mentioned during a global pandemic in a safe way, all of our participants who stayed extremely still for 7 and 1/2 hours and also had to request familiar face images from a bunch of their friends and family members, and Gazi Husain, an undergrad at Hunter who I'm mentoring who played a role in sort of collecting the images and processing images. For the Macaque fMRI experiments, the main person to acknowledge is Sofia Landi. She plans these experiments and collected all the data and a few others in the Freiwald lab who were involved in animal training and data collection for that project. And last, I just want to thank institutions that have supported my postdoc-- the Helen Hay Whitney Foundation and Leon Levy foundation and CBMM for supporting the lab. So thanks.
Associated Research Module: