Social structure learning in human anterior insula

Tatiana Lau, Samuel J. Gershman, Mina Cikara

TL;DR

This study reveals distinct brain regions involved in forming social groups based on either individual similarity or the overall social structure. It shows that the right anterior insula (rAI) is crucial for inferring group structure, and its activity better predicts ally choices than activity in regions associated with individual similarity.

What Problem Does This Paper Address?

Humans are inherently social creatures, constantly navigating complex group dynamics and forming coalitions. A fundamental aspect of this is the ability to distinguish "us" from "them." While previous research has often attributed this to simple similarity judgments (e.g., "they are like me"), this paper questions whether this is sufficient, especially when explicit group labels are absent. The authors propose that people might also infer "latent" social structures by considering how individuals relate to each other, not just how they relate to oneself. The core problem is understanding the neural mechanisms underlying this more sophisticated form of social group learning and how it influences our decisions about who to align with.

Methods

The researchers employed a multi-pronged approach combining computational modeling and fMRI.

  • Computational Models: They developed three models to capture how participants might form social representations:
  • Dyadic Similarity: Assumes people rely on pairwise similarity between themselves and each individual agent.
  • Feature Similarity-over-Agents: A second-order similarity measure, considering how agents' preferences align with each other and with the participant.
  • Latent Structure Learning: A more complex model that infers group assignments based on the overall pattern of agent choices, using a Bayesian approach (specifically, the Chinese restaurant process).
  • fMRI Study: 42 participants completed a task where they learned about the political stances of three novel agents (A, B, and C) on various issues. After this learning phase, they were asked to choose an ally for a "mystery" political issue. The fMRI data was analyzed to identify brain regions whose activity correlated with the outputs of each computational model (parametric modulators). They also tested whether the neural signal in these regions could predict behavioral choices beyond the model predictions.

Datasets

The study utilized a custom dataset of political issues sourced from ISideWith.com. These issues were selected for having significant agreement/disagreement discrepancies. Face stimuli were drawn from the Chicago Face Database (CFD), with careful selection to ensure gender-matching and control for prototypicality and attractiveness. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Neural Dissociation: The study provides compelling evidence for distinct neural substrates for different social learning strategies. Dyadic similarity and feature similarity-over-agents recruit the pregenual anterior cingulate cortex (pgACC), a region known for self-referential processing. In contrast, inferring latent social group structure relies on the right anterior insula (rAI).
  • Latent Structure is Key for Allyship: The rAI's activity, reflecting latent structure learning, was a better predictor of participants' ally choices than the pgACC's activity, which reflected dyadic similarity. This suggests that understanding the broader social network is more critical for forming alliances than simply matching individual preferences.
  • Anterior Insula's Role in Social Structure: The findings highlight the anterior insula's role not just in processing internal states (like pain or emotion) but also in dynamically updating representations of social relationships and group memberships, even in novel social contexts.

Key Results

  • Brain Regions:
    • Dyadic similarity and feature similarity-over-agents correlated with activity in the pgACC.
  • Latent structure learning correlated with activity in the right anterior insula (rAI), extending into the inferior frontal gyrus. This rAI cluster did not overlap with regions identified by the other models.
  • Predictive Power: Variability in the neural signal from the rAI significantly improved predictions of ally-choice behavior, above and beyond model predictions. The pgACC signal did not offer this predictive power.
  • Model Specificity: Cross-validation showed that the latent structure learning model explained significantly more variance in the rAI compared to the other models. Conversely, for the pgACC, no single model was distinctly superior.

Limitations and Future Work

  • Generalizability: The study used novel agents and political issues. While this allowed for the examination of learning processes, future work could explore how these mechanisms apply to pre-existing social groups and more complex social dynamics.
  • Choice of Behavior: The authors acknowledge that the specific behavior tested (ally choice) might influence which brain regions are most relevant. They suggest that other social decision-making tasks might engage the pgACC more.
  • Model Complexity: While the latent structure learning model was more predictive, the paper doesn't delve into the computational efficiency or cognitive load associated with these different strategies.
  • Future Directions: The authors emphasize the need for further research into how social latent structures and coalitions shape social choices, particularly in understanding and mitigating bias and discrimination. They also suggest exploring how these mechanisms might extend to other species.

Summary updated

Abstract

Humans form social coalitions in every society, yet we know little about how we learn and represent social group boundaries. Here we derive predictions from a computational model of latent structure learning to move beyond explicit category labels and interpersonal, or <i>dyadic</i>, similarity as the sole inputs to social group representations. Using a model-based analysis of functional neuroimaging data, we find that separate areas correlate with dyadic similarity and latent structure learning. Trial-by-trial estimates of 'allyship' based on dyadic similarity between participants and each agent recruited medial prefrontal cortex/pregenual anterior cingulate (pgACC). Latent social group structure-based allyship estimates, in contrast, recruited right anterior insula (rAI). Variability in the brain signal from rAI improved prediction of variability in ally-choice behavior, whereas variability from the pgACC did not. These results provide novel insights into the psychological and neural mechanisms by which people learn to distinguish 'us' from 'them.'
BibTeX
@inproceedings{lau2020social,
  author = {Tatiana Lau and Samuel J. Gershman and Mina Cikara},
  title = {{Social structure learning in human anterior insula}},
  booktitle = {eLife},
  year = {2020},
  doi = {10.7554/elife.53162},
}
Figure