Figure 1.

Interactions of mPFC and lPFC in the HER model. (A) The HER model is organized hierarchically, with each hierarchical level instantiating a computational motif of prediction and prediction error calculations. Information flows between hierarchical levels along top-down and bottom-up pathways, which carry information regarding likely errors in a given context (top-down), and error signals (bottom-up) derived from external feedback (base hierarchical level) conjoined with items maintained in working memory (additional hierarchical levels). Circled numbers indicate governing equations within and between hierarchical levels (see section Methods). Reciprocal connections between mPFC and lPFC correspond to a putative hierarchical rostrocaudal gradient (Badre and D'Esposito, 2009) with bottom-up and top-down pathways governing the adjustment of behavior during ongoing task performance. (B) During a trial, top-down connections serve to establish the current context and relevant rules that govern eventual behavioral responses during preparatory periods. (C) Following the generation of a response (right), bottom-up error signals are used to update outcome predictions and derive composite “proxy” outcomes that train higher-order representations of rules and contexts.