FIGURE 2.

Neural dynamics supporting threat learning and updating. (a) Specification of the DCM model space in terms of: (i) task‐independent effective connectivity (grey, dashed lines) (A‐matrix); (ii) modulatory connections (B‐matrix) (blue), including threat signals in both acquisition and reversal task phases; and (iii) direct inputs to the system (C‐matrix) comprising visual (all CS events) and auditory (US) stimuli (red). (b) Three models were estimated for each subject (see text for details). The difference between models arises from the specification of contextual modulators (threat signals, all trials, or both). Bayesian Model Selection showed that the exclusive modulation by threat signals (Model 1) best explained the fMRI data (as accessed by the highest exceedance probability). (c) Results from Bayesian Model Reduction (BMR) on second level Parametric Empirical Bayes analysis of trial‐independent (fixed) connections across individuals. Results showed a positive modulation from the dACC to the AIC, a negative modulation from the AIC to the dACC, and local effects within both regions. (d) BMR results showed a significant modulatory effect of threat signals on patterns of effective connectivity during acquisition and reversal (left). Results highlight very consistent modulatory effects of task phase (from acquisition to reversal) on AIC → dACC (top, right) and AIC self‐connections (bottom, right)