Tasks and behavior.

A) Example probability learning task (similar to Study 1, 2). Top left: Four example trials. Bottom: Participants are presented with a sequence of two stimuli (here illustrated with blue and yellow dots). The generative probability p(blue) changed unpredictably (see gray line), the blue line represents the estimates of an ideal observer. Top right: Model posterior over p(blue) at the trial indicated by the red line. Note that the stimuli shown here are just for illustration, actual stimuli were geometric patterns or tones in different studies. B) Reward learning task (Study 4). Top left: Two example trials. At each trial subjects choose between two options and observe the outcome (reward) of the selected option. Bottom: The latent mean reward levels (light orange and purple) underlying the two options could change unpredictably between 3 reward levels (30, 50, 70). The observation (dark orange or purple dots) for the chosen option is drawn from a Gaussian distribution. The dark orange and purple lines represent the estimates of an ideal observer. Top right: Example posterior probability following an observation. C) All studies contained occasional reports about the latent parameters. Reported confidence levels sorted in 6 bins of ideal estimates. Dots and error bars indicate the mean and SEM. The dashed line is a linear fit. Note that for the probability learning studies (Studies 1-3) subjective confidence (lowest and highest response bins set to 0 and 1 arbitrarily) and ideal confidence (in log SD unit) are on different scales. D) Reported probability estimates sorted in 6 bins of ideal estimates. Dots and error bars indicate the mean and SEM. The solid line indicates the identity. Probability reports are not available for Study 3 and 4.

Effect maps are largely invariant across studies.

A/D) Group-level z-statistic maps for confidence/surprise in the four studies. B/E) Brain regions where effects significantly correlated with confidence/surprise across studies. The colors correspond to the number of studies in which a region showed significant clusters (thresholded at p<0.001 and corrected for multiple comparisons across voxels at the cluster level with pFWE <0.05). C/F) Correlations between effect maps across studies. Underlined correlation coefficients were significant in a comparison to a null model (all p= 0.001 at 1000 rotations)

Learning-related effect maps can be predicted from receptor/transporter distributions.

A) Analysis framework: Is there a linear relationship between a learning-related effect map and PET-derived density distributions of 20 receptors/transports in the cortex? B) Cross-validated R2 for a model including all 20 receptor/transporter distributions, by study and variable combination. The light part of each bar corresponds to the average performance of a spatial autocorrelation-preserving null model. The null model was not applicable to study 2, due to incomplete cortical coverage of the fMRI data. Error bars reflect the SEM. Asterisks denote significant models, compared to the null, in a paired t-test (FDR-corrected p <0.05, one-sided). n.a.: not applicable, n.s.: not significant. C) Mean cross-validated explained variance in ratio to the mean amount of variance that can be explained in a held out subject’s fMRI effect map by the other effect maps. Error bars reflect the bootstrapped SEM. Asterisks denote study/variable combinations in which the receptors/transporters explained a significantly larger fraction of variance than in the language network. Non-parametric permutation test on group level ratios (one-sided, FDR-corrected). n.s.: not significant.

Contribution of each receptor/transporter to explaining the learning-related effect maps.

Dominance analysis assigns a proportion of the model fit to each independent variable. For comparison across studies and learning-related variables, each contribution is normalized by the total model fit, and expressed as a percentage. The top row in A) (confidence) and B) (surprise) depicts the mean percent contribution across all subjects and studies. Error bars reflect the SEM across all subjects. The heat maps illustrate the mean dominance results by study.

Mean results by study for a linear regression analysis comparing explicit subject reports and ideal observer estimates

Correlations between confidence and surprise maps.

Comparison of average model fit for different model complexities

Results from Mann–Whitney U test on subject level explained variance ratios

Parameter estimates across receptors/transporters

Pearson’s correlations between pairs of receptor/transporter distributions, grouped by neuromodulator.

Results of the dominance analysis for a model including the linear and quadratic predictors, revealing the contribution of each receptor and transporter to the model fit.

The percent contribution of each receptor reflects the variables dominance normalized by the model fit (R2). The top row in depicts the mean percent contribution across all subjects from all probability studies. Error bars reflect the SEM. The heat maps illustrate the mean dominance results by study.