Fig. 5: Identifying conjunction hubs: brain areas (vertices) that contain task-relevant conjunctions of sensory stimulus and task rule activations.

a The Guided Activation Theory states that there exist a specific set of association (or hidden) areas that integrate sensory stimulus and task context activations to select appropriate motor response activations. In an ANN where task rules and sensory stimulus activations serve as inputs, the ANN’s hidden layers integrate rule and stimulus activations providing a computational framework that is analogous to the role suggested for association areas in the Guided Activation Theory. b We therefore used the representational similarity matrix (RSM) of the ANN’s hidden layers as a blueprint to identify analogous conjunctive activations in empirical data. c We constructed RSMs for each brain parcel (using the vertices within each parcel as features). We evaluated the correspondence between the representational geometry of the ANN’s hidden layers and each brain parcel’s representational geometry. Correspondence was assessed by taking the correlation of the upper triangle of the ANN and empirical RSMs. d The representational similarity of ANN hidden units and each brain parcel. e We showed the top 10 regions with highest similarity to the ANN hidden units. f The full ENN architecture for the C-PRO task. We identified the vertices that contained task-relevant rule, sensory stimulus, conjunctive, and motor output activations.