Social training reconfigures prediction errors to shape Self-Other boundaries

Abstract

Selectively attributing beliefs to specific agents is core to reasoning about other people and imagining oneself in different states. Evidence suggests humans might achieve this by simulating each other's computations in agent-specific neural circuits, but it is not known how circuits become agent-specific. Here we investigate whether agent-specificity adapts to social context. We train subjects on social learning tasks, manipulating the frequency with which self and other see the same information. Training alters the agent-specificity of prediction error (PE) circuits for at least 24 h, modulating the extent to which another agent's PE is experienced as one's own and influencing perspective-taking in an independent task. Ventromedial prefrontal myelin density, indexed by magnetisation transfer, correlates with the strength of this adaptation. We describe a frontotemporal learning network, which exploits relationships between different agents' computations. Our findings suggest that Self-Other boundaries are learnable variables, shaped by the statistical structure of social experience.

Fig. 1. Experimental design.

a Three-day experimental timeline. On day 1 subjects played an intertemporal choice task followed by a visual perspective-taking task (see Fig. 3a). On day 2 subjects were trained on the false belief task shown in b, with two different social contexts. Here, the Lo-Share context is represented by the female avatar and the Hi-Share context is represented by the male avatar. Subjects then played the visual perspective-taking task again to measure transfer effects. On day 3 subjects were tested on the false belief task with concurrent fMRI. This time there were no statistical differences between the two social contexts. b Trial structure of the probabilistic false belief task. The middle row shows what the subject sees on the computer display. Each trial comprises a Bernoulli outcome on the bottom half of the display (pink or yellow), and an image on the top half of the display, which indicates whether the trial is ‘privileged’ (blue), ‘shared’ (purple), or ‘decoy’ (red). Subjects were intermittently probed to report their estimate of the Bernoulli parameter, p (Self-probe), or their estimate of the other agent’s false belief about p (Other-probe). c An example pair of random walks used to generate a trial sequence for the false belief task. The trial sequence is designed to induce uncorrelated beliefs in Self and Other.

Fig. 2. Behavioural training induces sustained changes in Self-Other distinction ability.

a Performance in FBT, split by session (train or test), context (Lo-Share or Hi-Share) and probe trial (Self or Other). In both sessions, there was a main effect of context with worse performance in the Hi-Share context. Large white circles show the behaviour predicted by the winning models. n = 40 independent subjects. Repeated measures ANOVA testing for main effect of condition in training session F(1, 39) = 12.64, p = 0.001. Repeated measures ANOVA testing for main effect of condition in testing session F(1, 39) = 6.78, p = 0.013. b Correlation between model-derived Self- and Other-attributed beliefs in different trial bins, split by session and context. In both sessions, there was a main effect of context with a higher Self-Other correlation in the Hi-Share context (pink line) than in the Lo-Share context (purple line). This finding was invariant to the number of trial bins used in the analysis (see Supplementary Fig. 5) and was driven largely by the λ parameter (Supplementary Figs. 2 and 3). Repeated measures ANOVA testing for main effect of condition in training session F(1, 39) = 21.7, p < 0.001. Repeated measures ANOVA testing for main effect of condition in testing session F(1, 39) = 3.8, p < 0.001. c Generative performance of the best-fitting model for the Hi-Share context in the test session for two exemplar subjects. Self and Other probe trials are intermixed in the order they were presented to the subjects. d Parameter identifiability of the most complex winning model (Hi-Share context test session). Each cell shows a Spearman correlation coefficient derived from correlating subjects’ true parameter estimates with recovered parameter estimates. A subscript ‘s’ indicates that the parameter is specific to Self updates. A subscript ‘o’ indicates that the parameter is specific to Other updates. A subscript ‘1’ indicates that the learning rate is specific to ‘privileged’ and ‘decoy’ trials. A subscript ‘2’ indicates that the learning rate is specific to ‘shared’ trials. All error bars denote s.e.m. Source data are provided as a Source Data file.

Fig. 3. Behavioural training transfers to a perspective-taking task.

a Trial structure of the perspective-taking task. Subjects were first presented with a perspective (Self or Other), followed by a target pattern and number and then a visual scene. Subjects had three seconds to report whether or not the scene contained the target number of target patterns, if adopting the relevant perspective. b Incongruency effect in the perspective-taking task, averaged over Lo-Share and Hi-Share. Mean drift rate parameters were higher on congruent compared to incongruent trials at both baseline and transfer phases. n = 46 independent subjects. Baseline: Paired two-sided t-test: t(45) = 7.6, p < 0.001. Transfer: Paired two-sided t-test: t(45) = 6.7, p < 0.001. c Transfer effect in the perspective-taking task. ‘Corrected drift rate change’ (see Methods) was significantly higher on trials with the Lo-Share agent than trials with the Hi-Share agent. n = 46 independent subjects. Repeated measures ANOVA testing for main effect of avatar: F(1, 45) = 7.4, p = 0.009. Error bars denote s.e.m. Source data are provided as a Source Data file.

Fig. 4. Representations of PEs adapt with behavioural training.

a Clusters of voxels where BOLD signal covaried with |PE|, either Self-attributed (PE^s) or Other-attributed (PE^o), from a searchlight analysis, using two one-sided t-contrasts on n = 40 independent subjects. Clusters were defined with a cluster-forming threshold of p < 0.001. Only clusters large enough to survive FWE-correction at p < 0.05 are displayed (top). BOLD signal was extracted from these clusters and patterns of PE-related activity were compared for Self- and Other-attributed signals for the Lo-Share and Hi-Share contexts. Exemplar trial patterns are shown from a single subject. The ‘signal change’ refers to the difference in BOLD signal between a trial with a large PE and a trial with a small PE (see Methods). Patterns for PE^self and PE^other are more similar for the Hi-Share context than the Lo-Share context (bottom). b Decoding performance (cross-entropy below chance) when classifying PE activity patterns as Self- or Other-attributed. Classification accuracy was significantly higher in the Lo-Share context than the Hi-Share context. n = 40 independent subjects. Paired two-sided t-test: t(39) = 2.1, p = 0.041. c Decoding performance when predicting |PE^other| after training on |PE^self| and vice versa. Cross-decoding accuracy (Fisher Z-transformed correlation) was significantly higher in the Hi-Share context than the Lo-Share context. n = 40 independent subjects. Paired two-sided t-test: t(39) = 2.75, p = 0.009. d The difference in cross-decoding performance shown in c, is positively correlated with the ratio of leak to learning rate (parameters derived from the Hi-Share context). This indicates a relationship between neural Self-Other mergence and behavioural Self-Other mergence. n = 40 independent subjects. Pearson correlation: r = 0.41, p = 0.009. e Cluster of voxels where myelin-related MT covaried with the difference in cross-decoding shown in c, overlaid on white matter. Clusters were defined with a cluster-forming threshold of p < 0.001. Only clusters that were large enough to survive FWE-correction at p < 0.05 are displayed. All error bars denote s.e.m. Source data are provided as a Source Data file.

Fig. 5. Probability of sharing information between Self and Other is tracked in vmPFC and temporal pole.

a Example regressors from a learning model that tracks the probability of encountering a ‘shared’ trial, P(share). The red line indicates P(share) whilst the black line indicates the PE magnitude from this model. The sharp boundary, at trial 444, indicates that this particular subject did the Lo-Share task, followed by Hi-Share. Left: Regressors generated with a learning rate (η) of 0.01. Right: Regressors generated with η of 0.025. b Left: P(share), at η = 0.01, is tracked in bilateral vmPFC. Blue shading shows the mask used for small-volume correction. Right: P(share), at η = 0.025, is tracked in left temporal pole. Clusters were defined with a cluster-forming threshold of p < 0.001. Only clusters that were large enough to survive FWE-correction at p < 0.01 are displayed (additional FWE correction was applied to account for testing multiple η values). One-sided t-contrasts on n = 40 independent subjects. t-statistics are shown in the colour bars. c Left: Mean contrast estimates in vmPFC cluster are positively correlated with the difference in Self-Other cross-decoding between the Hi- and Lo-Share conditions, after excluding two subjects with extreme contrast estimates (empty circles). Pearson correlation: r = 0.35 p = 0.032. Right: Mean contrast estimates in the temporal cluster are not correlated with the difference in Self-Other cross-decoding between the Hi- and Lo-Share conditions. Pearson correlation: r = −0.05 p = 0.74. Source data are provided as a Source Data file.

Fig. 6. Temporal discounting propensity can be predicted by Self-Other distinction.

a Choice behaviour in the intertemporal choice task. Each scatter point shows the proportion of trials, averaged over n = 40 independent subjects, where the larger later option was chosen. The inset bar plot shows the relative fits of one- and two-parameter hyperbolic discounting models in terms of integrated Bayesian information criterion (iBIC). The two-parameter hyperbolic discounting curve, using the mean parameters across subjects, is shown as a blue dashed line. Error bars denote s.e.m. b Performance of five logistic regression models predicting subjects’ temporal discounting propensities, using various predictors, with leave-one-subject-out cross-validation. Data is presented as median cross-entropy ± interquartile range, across n = 40 independent subjects. Lower cross-entropy denotes more accurate prediction. Discounting propensity could be predicted, above chance, using the leak parameter (λ): p = 0.005. Prediction accuracy was significantly better with λ than with any other behavioural parameter from the false belief task (learning rate p = 0.002, memory decay p = 0.004, temperature p = 0.001). Prediction accuracy (p < 0.001) was significantly improved (p = 0.016) when Self-Other cross-decoding accuracies from the fMRI analysis were included as additional predictors in a new logistic regression model. All p values were derived from two-sided permutation tests (see Methods). Vertical bars show chance level prediction accuracy for each logistic regression model. Source data are provided as a Source Data file.