The neural representational space of social memory

Social functioning involves learning about the social networks in which we live and interact; knowing not just our friends, but also who is friends with our friends. Here we utilized a novel incidental learning paradigm and representational similarity analysis (RSA), a functional MRI multivariate pattern analysis technique, to examine the relationship between learning social networks and the brain's response to the faces within the networks. We found that accuracy of learning face pair relationships through observation is correlated with neural similarity patterns to those pairs in the left temporoparietal junction (TPJ), the left fusiform gyrus, and the subcallosal ventromedial prefrontal cortex (vmPFC), all areas previously implicated in social cognition. This model was also significant in portions of the cerebellum and thalamus. These results show that the similarity of neural patterns represent how accurately we understand the closeness of any two faces within a network, regardless of their true relationship. Our findings indicate that these areas of the brain not only process knowledge and understanding of others, but also support learning relations between individuals in groups. Significance Statement Knowledge of the relationships between people is an important skill that helps us interact in a highly social world. While much is known about how the human brain represents the identity, goals, and intentions of others, less is known about how we represent knowledge about social relationships between others. In this study, we used functional neuroimaging to demonstrate that patterns in human brain activity represent memory for recently learned social connections.

Abstract: Social functioning involves learning about the social networks in which we 41 live and interact; knowing not just our friends, but also who is friends with our friends. 42 Here we utilized a novel incidental learning paradigm and representational similarity 43 analysis (RSA), a functional MRI multivariate pattern analysis technique, to examine the 44 relationship between learning social networks and the brain's response to the faces 45 within the networks. We found that accuracy of learning face pair relationships through 46 observation is correlated with neural similarity patterns to those pairs in the left 47 temporoparietal junction (TPJ), the left fusiform gyrus, and the subcallosal ventromedial 48 prefrontal cortex (vmPFC), all areas previously implicated in social cognition. This 49 model was also significant in portions of the cerebellum and thalamus. These results 50 show that the similarity of neural patterns represent how accurately we understand the 51 closeness of any two faces within a network, regardless of their true relationship. Our 52 findings indicate that these areas of the brain not only process knowledge and 53 understanding of others, but also support learning relations between individuals in 54 groups. 55 those two individuals within the network: that is, does the similarity of the pattern of 111 response to two network members increase as a function of the closeness of those 112 members? We also examined if the memory for tie strength between network members 113 was related to the similarity of the fMRI voxel pattern response to the faces of members. 114 To understand the contribution of the frequency of face pairing during network learning 115 to memory and neural representations, we compared a network in which centrality 116 differed between members (i.e. some members had more connections than others) to a 117 network with no individual centrality. 118 119

Materials and Methods 120
Participants 121 22 healthy individuals (10 females; age range = 18-34; mean age = 23; ethnicity 122 = 64% White, 18% Hispanic/Latino, 18% Asian) participated in a 1.5 hour learning 123 session immediately followed by a 1.5 hour fMRI scanning session. Behavioral data 124 from a total of 31 individuals was collected, but seven subjects did not meet the learning 125 criteria from the behavioral task, one subject was unable to be scanned, and one 126 subject's fMRI data was incomplete. All participants were right handed (self-reported) 127 with normal or corrected-to-normal vision. Participants provided written informed 128 consent in accordance with the Declaration of Helsinki and the Human Subjects Review 129 Board at George Mason University and were compensated for their time. network. This meant that in network 1 the centrality of members was varied (variable-152 centrality network), while in network 2 centrality was equated across members (fixed-153 centrality network). This also meant that the frequency of presentation of each face 154 differed in network 1, but was equivalent in network 2. Each trial consisted of a face pair 155 presented for 4 seconds accompanied by a question, and participants were asked to 156 make a comparison between the faces and decide which person better fit the question. 157 Questions consisted of behavioral and personality characteristics taken from various 158 personality surveys included in the International Personality Item Pool 159 (http://ipip.ori.org/). Half of the questions asked which person was more likely to exhibit 160 a characteristic, and half asked which person was less likely (example: "Who is more 161 likely to be easily intimidated?"). Network learning took place in alternating blocks, 162 where the subjects viewed 36 randomly presented trials of one network followed by 36 163 trials of the second network. Participants completed 720 trials in total (360 per network), 164 with the weakest network connections being presented a total of 20 times and the 165 strongest a total of 80 times. 166 After completing the paired face viewing portion, participants were explicitly 167 tested on their knowledge of the network connections. They were told that the faces 168 represented college students living in a dorm together, the faces that they saw paired 169 together previously represented friend connections, and the more often they were 170 presented together, the closer in friendship the pair was. They were asked to group all 171 of the faces into two separate halls, as a check to make sure that they could distinguish 172 between the faces in different networks. They were then presented with all possible 173 within-network face pairs twice and asked to rate their relationship on a scale of 0 (do 174 not know each other) to 4 (know each other very well). They were not asked about 175 cross-network face relationships. This explicit testing period was included to ensure that 176 participants learned the structure of the networks to an appropriate level before being 177 scanned. Participants who were within 2 standard deviations of pilot data (hit rate = 178 0.85, SD = 0.14; false alarm rate = 0.35, SD = 0.15) were included in further analysis.
Both parts of the behavioral task (learning and recall) were presented to the participant 180 using PsychoPy version 1.842 software (http://www.psychopy.org/). 181 The fMRI task stimuli included the same 12 faces from the behavioral task as 182 well as 12 novel faces as a control. Faces were presented one at a time for one second 183 on a black background with a 4.5 second inter-stimulus interval (black screen with a 184 white fixation dot), and participants completed a 1-back task to ensure they were 185 attentive. The task consisted of four runs of 9.6 minutes each, resulting in each face 186 being presented a total of 16 times (not counting repeats, which were included in 187 analysis as a separate regressor). Following the face task, participants underwent an 188

Regions of Interest (ROI) and Mask Creation 216
Localizer data preprocessing steps were identical except the functional data was 217 registered only to each subject's specific structural image. Face-selective regions of 218 interest (ROIs) were created from subtracting the combined object, scrambled object, 219 and scene conditions from the face condition. These regions included bilateral posterior 220 STS and fusiform face area (FFA). Activity was thresholded at Z > 3.7 (p < 0.0001) for 221 most ROIs, although this threshold was relaxed to Z > 3 (p < 0.001) in one subject, Z > identified relationship ties significantly greater than chance across both networks (t(21) 284 = 8.08, p = 7.004e-08). Table 1 shows the average hit rate, false alarm rate, sensitivity 285 (d), and the correlation between true and reported perceived strength for ties and 286 relationship strength across subjects. Paired sample two-tailed t-tests revealed no 287 significant differences between recall measures for the two networks. There were also 288 no significant age or gender effects for any of the measures. When averaged together 289 across subjects, group perceived relationship strength was highly correlated with the 290 true network structure (r = 0.896, p < 0.00001). In order to assess whether our 291 behavioral task was comparable to previous forms of social network learning and recall, 292 we calculated performance measures used by Brashears (2013). Accuracy refers to the 293 number of ties correctly recalled divided by the number of total ties reported, coverage 294 refers to the number of ties correctly recalled divided by the total tie number in the 295 network, and performance refers to the product of accuracy and coverage. T-tests 296 revealed no significant differences between accuracy or performance measures in our 297 task and those of Brashears (accuracy: t(21) = 0.98, p = 0.34; performance: t(21) = 298 0.58, p = 0.56), and we actually saw an increase in coverage (t(21) = 3.58, p = 0.002), 299 although our networks were smaller, so participants did not need to remember as many 300

ties. 301
When exploring network recall, it is important to not only look at the correctly 302 identified ties, but also at the pattern of mistakes made. Specifically, we wanted to see 303 whether there are systematic biases that could be predicted by the level of relationship 304 strength of the friend pairs. We assessed recall by relationship strength by looking at 305 the relative direction of the errors made (i.e. how much subjects overestimated or 306 underestimated the strength of the connection). A linear mixed effects regression model 307 (fixed effect = strength; random effects = subject, residual) revealed that relationship 308 strength affected recall error compared to a null model (χ 2 (1) = 226.9, p < 2.2e-16). This 309 pattern shows that overall, weak ties were reported to be stronger than they actually 310 were whereas strong ties were reported to be less strong (Figure 2a). This reflects a 311 general tendency to assume a mid-level relationship between observed people when 312 the relationship is not explicitly known or is unable to be recalled. This central tendency 313 effect seems to be robust, as it was also observed in a separate subject sample (N = 314 23, 17 females, mean age = 19.6 (sd = 2.4)) learning a larger social network (N= 9) and 315 a larger possible range of relationship tie strengths to choose from (0-6) (χ 2 (1) = 362.84, 316 p < 2.2e-16) (Figure 2b). In order to be able to compare network memory performance 317 to the neural patterns in response to each face in the network, we converted the relative 318 error for each subject to absolute error, which gives a measure of distance from the true 319 network structure, regardless of the direction of that error. The absolute error measure 320 for each subject for each network was then used as a dissimilarity model for RSA to 321 elucidate what neural patterns underlie these errors. 322 323 fMRI Results 324 During fMRI scanning, participants viewed the original faces from the social 325 network behavioral session, as well as 12 novel faces and were asked to press a button 326 when they saw a face repeated to guarantee attention. We first conducted a GLM 327 comparing the 12 familiar faces from the two networks to unfamiliar control faces. 328 To examine whether information related to social network recall is represented in 345 the brain, we carried out RSA searchlight analysis on several DMs representing 346 different types of information about the networks. The first compared neural pattern 347 similarity to social tie strength, with more similar neural responses to any pair of faces 348 representing a closer relationship between those faces. Neural pattern similarity that 349 reflects this network structure would indicate that the brain carries information about the 350 true relationship between individuals, regardless of whether people recall those 351 relationships accurately. We did not find a significant correlation between these 352 measures in our analyses. As the network properties differed between network 1 and 2 353 (see Methods section for details), we compared the two networks and found no 354 significant differences. 355 While the pattern similarity to viewing faces was not significantly associated with 356 social tie strength, it was significantly associated with the subjects' memory for that tie 357 strength. We assessed this by measuring each subject's absolute distance from each 358 true network structure and the 1-correlation distance between the neural response to 359 each face viewed in the scanner. An association between these two measures would 360 indicate that the more accurately a subject perceives the true relationship tie strength 361 between a pair of faces, the more similar their neural pattern response is to those two 362 faces. In other words, this association does not rely on the actual connection strength of the relationships themselves, but the subject's memory of that connection, reflecting a 364 second-order knowledge or understanding of a social relationship. Neural pattern 365 similarity in the left TPJ, the left fusiform gyrus, the subcallosal cingulate cortex, the 366 cerebellum, the left thalamus, and a small portion of the left lateral occipital lobe was 367 significantly correlated with the recall accuracy model, suggesting that neural 368 populations within these areas are important for accurate perception of social 369 relationship strength (Figure 4). Table 3 reports MNI coordinates, cluster size, and 370 peak voxel activity of results. As with tie strength similarity, we compared the two 371 networks to each other separately and found no significant differences. This indicates 372 that the significant findings are not due simply to frequency of the face pairs being 373 presented, as this differed between the two networks. 374 We also conducted RSA searchlights using two other dissimilarity matrix models: 375 Our results indicated that the relationship between memory for tie strength and 405 neural pattern similarity was not due to factors such as the frequency at which different 406 faces were paired with others during the learning of the network, as we found no 407 differences in memory performance or RSA results between a network in which some 408 faces were paired more often with others (variable-centrality network) and one in which 409 all faces had the same number of connections to other network members (fixed-410 centrality network). In addition, participants saw each individual face the same number 411 of times as they learned one of the two networks and there were no significant 412 behavioral or neural differences between the two networks, and therefore our results While most of our subjects were able to accurately report relationship ties, there 471 were individual differences between ability to recall relationship strength (measured by 472 the correlation between the true structure and the reported structure of the networks). 473 Previous literature does indicate that there are individual differences in social recall. 474 Individuals tend to report group and relationship averages or norms more accurately 475 than individual interactions, but more experienced observers show more accurate recall, 476 especially when group structure is transitive (Freeman & Romney, 1987;Freeman, 477 1992; Kumbasar, Romney, & Batchelder, 1994). It has been suggested that humans 478 use cognitive heuristics such as triadic closure in order to remember social ties (De 479 Soto, 1960;Freeman, 1992;Brashears, 2013;Brashears & Quintaine, 2015). 480 Overestimation of symmetric ties for less central network members, and 481 underestimation of more central network members, has also been reported previously 482 (Krackhardt, 1987). There are also differences in the ability to perceive and remember 483 non-social patterns, but evidence suggests that learning, remembering, and storing 484 social information might be distinct from traditional learning and memory systems 485      Table 3. Coordinates, cluster size, and peak activity for the group-level significant clusters from 750 the recall error model.