The behavioral and neural mechanisms underlying the tracking of expertise

Abstract

Evaluating the abilities of others is fundamental for successful economic and social behavior. We investigated the computational and neurobiological basis of ability tracking by designing an fMRI task that required participants to use and update estimates of both people and algorithms' expertise through observation of their predictions. Behaviorally, we find a model-based algorithm characterized subject predictions better than several alternative models. Notably, when the agent's prediction was concordant rather than discordant with the subject's own likely prediction, participants credited people more than algorithms for correct predictions and penalized them less for incorrect predictions. Neurally, many components of the mentalizing network-medial prefrontal cortex, anterior cingulate gyrus, temporoparietal junction, and precuneus-represented or updated expertise beliefs about both people and algorithms. Moreover, activity in lateral orbitofrontal and medial prefrontal cortex reflected behavioral differences in learning about people and algorithms. These findings provide basic insights into the neural basis of social learning.

Figure 1

Experimental Task (A) Experimental task and timeline are shown. Participants were presented with either a picture of a human face (condition 1), a 2D fractal image symbolizing an algorithm (condition 2), or a hypothetical asset (condition 3). In conditions 1 and 2, subjects had to either bet for or against the agent. After a brief delay, they observed the agent’s choice: a prediction about whether the hypothetical asset would increase or decrease in value. Following a jittered fixation period, feedback was presented indicating whether the asset went up or down and whether the subject made or lost $1 for correct or incorrect predictions, respectively. In condition 3, the subject had to predict whether the asset would go up or down, and then received immediate feedback. ISI, interstimulus interval; RT, reaction time. (B) The task was divided into four blocks of 55 trials each. In each block, the subject observed the predictions of three agents (either two people and one algorithm, or the reverse). The true performance level of an agent is shown above each stimulus. Assignment of specific faces and fractal images to the corresponding predictions was pseudorandomly generated and counterbalanced across subjects.

Figure 2

Task Parameters and Behavioral Analyses (A) True probabilities and model estimates of correct performance for the eight agents (four people and four algorithms) that subjects observed during the experiment are shown for one subject. For half of the subjects, blue represents people, and red represents algorithms; this was reversed for the other half, as indicated by parentheses. (B) Underlying probability that the asset’s value would increase and corresponding model estimates are plotted across trials. (C) Predictions of the best-fitting behavioral model are plotted against the true choice frequencies for all trials (top) and for predictions about people (bottom left) and algorithms (bottom right). Circles indicate means. Error bars represent ±SEM. (D) In the left panels, regression coefficients for correct and incorrect agent predictions of past trials are plotted but divided into correct trials with which subjects agree (Agree and Correct), disagree (Disagree and Correct), and incorrect trials with which subjects agree (Agree and Incorrect) and disagree (Disagree and Incorrect). In the rightmost panel, mean coefficients reflecting the overall influence of outcomes across trials n-1 to n-5 for correct trials with which subjects agree, compared to disagree (Agree Correct − Disagree Correct), and mean coefficients for incorrect trials with which subjects agree, compared to disagree (Agree Incorrect − Disagree Incorrect), are plotted separately for people (blue) and algorithms (red). Note that inverse coefficients are computed for incorrect trials such that the y axis indicates positive effects for correct trials and negative effects for incorrect trials. ^∗p < 0.05; ^∗∗p < 0.01. See also Figure S2.

Figure 3

Expected Value and Reward Prediction Errors (A) Z-statistic map of the chosen option’s expected reward value at decision time is presented. (B) The same is shown for rPEs at feedback time. Maps are thresholded at Z > 3.1, p < 0.001, uncorrected for display purposes and are overlaid onto an average of subjects’ T1-weighted structural maps. Activations range from red (minimum) to yellow (maximum) Z-statistic values.

Figure 4

BOLD Effects of Ability Estimates (A) Left view shows a sagittal slice through the Z-statistic map (p = 0.05, cluster corrected across the whole brain) for ability belief, as predicted by the sequential model, independent of agent type (people or algorithms), at decision time. Right view shows the time course of the effect of expertise from independently identified rmPFC ROIs (circled), plotted separately for people (cyan) and algorithms (orange) across the entire trial. Dark lines indicate mean effects; shadows show ±SEM. (B) Left view is a sagittal slice through Z-statistic map (p = 0.05 whole-brain cluster corrected) relating to individual differences in the effect of expertise and the fit to behavior of the sequential model. In the right view, a scatterplot of the percent signal change elicited by expertise in independently identified rmPFC ROIs (circled) is plotted against the model fit (less negative numbers indicate better fit) for people (cyan) and algorithms (orange) separately. Activations range from red (minimum) to yellow (maximum) Z-statistic values.

Figure 5

BOLD Effects of Simulation-Based aPEs (A) Left view shows Z-statistic maps (p = 0.05 cluster corrected) for the simulation-based aPE predicted by the sequential model, independent of agent type (people or algorithms), at the time of the observed agent’s choice. Right view shows the time course of the effect of this aPE in rTPJ (circled) plotted separately for people (blue) and algorithms (red) across the entire trial. Z-statistic map and time course are displayed according to the same conventions used in Figure 4. (B) Left view is a sagittal slice through Z-statistic map (p < 0.001 uncorrected for display purposes) relating to individual differences in the effect of simulation-based aPEs and the fit to behavior of the sequential model across people and algorithms. In the right view, a scatterplot of the percent signal change elicited by aPEs in independently identified rTPJ ROIs is plotted against the model fit for people (blue) and algorithms (red) separately.

Figure 6

BOLD Effects of Evidence-Based aPEs (A) Left view shows Z-statistic maps (p = 0.05 cluster corrected) for the second aPE predicted by the sequential model, independent of agent type, at the time of feedback. In the right view, a time course of the effect in rdlPFC (circled) is plotted across the trial separately for people (green) and algorithms (magenta). Z-statistical map and time course are displayed according to the same conventions used in Figure 4. (B) Left view shows a Z-statistic map resulting from a between-subjects analysis of intersubject differences in relative behavioral fit (log likelihood) of the sequential and pure simulation models and the BOLD effect of evidence-based aPEs (p = 0.05 cluster corrected). In the right view, the percent signal change elicited by aPEs in independently identified rdlPFC ROIs (circled) is plotted against the relative model fit between sequential and simulation models (positive values indicate better fit of sequential compared to simulation model) for people (green) and algorithms (magenta) separately.

Figure 7

mPFC and lOFC Reflect Behavioral Differences in Ability Learning (A) Z-statistic maps (p = 0.05, cluster corrected) relating to the interaction between outcome type and agent type revealed in behavior (see Figure 2D) are shown. Z-statistic maps represent the following contrast between unsigned prediction errors at feedback: ((Agree and Correct − Disagree and Correct) − (Agree and Incorrect − Disagree and Incorrect)) × people − ((Agree and Correct − Disagree and Correct) − (Agree and Incorrect − Disagree and Incorrect)) × algorithms. Z-statistitcal map and time course are displayed according to the same conventions used in Figure 4. (B) Percent (%) signal change elicited by unsigned prediction errors for correct and incorrect agent predictions, when subjects would have likely agreed compared to disagreed, is plotted separately for people (blue) and algorithms (red). Plots on the far right show the same divided into the four outcome types separately: AC, DC, AI, and DI. See also Figure S4.