Comparison between human participants and neural model performance on a probabilistic inference task.

A. Behavioral task design: on individual trials, human participants were asked to generate a behavioral choice based on a somatosensory cue input. Inputs were either up or down, and the mapping onto the choice (match or non-match) changed covertly across blocks of ~50 trials. Participants were told what the correct choice was on every trial (target). The mapping was probabilistic with different association levels (90%, 70% and 50%). B. Cartoon of the main neural model used in this study. The model contains a recurrently connected reservoir of rate neurons that capture the dlPFC, with long range connections to feedforward MD neurons with a winner-take-all mechanism. The model receives task relevant sensory inputs (up and down) as well as value of the mapping strategies (match and non-match). The model outputs behavioral choice (up or down) through two output neurons that linearly decode the reservoir activity. C. Mean task performance of human participants (purple) and of the neural model (green) plotted over 10 blocks of association levels, to show qualitative similarity (middle). Grey bars reflect the correct rule that was rewarded on that trial and green bars show the model response (bottom). The red trace shows the match strategy value signaled by the vmPFC (top, see Methods) which served as an input to the dlPFC RNN. These expected reward values closely follow the association levels through the experiment (black line, top). D. Comparison of model and human performance of high (90/10% predictability), low (70/30%), and non-predictive (50%) association levels. Pairwise comparisons of human responses across association levels were statistically significant for all pairs (left; two-way ANOVA; 90/10% and 70/30%, ***p E. Correlation between human and model performance with randomly sampled chunks of trials (humans: 15 trials, model: 150 trials) sampled from all trials (Pearson correlation, R2 = 0.89, ***p < 0.0001).