Fig. 6: Simulating context-dependent sensorimotor transformations with empirically-estimated task activations and inter-unit FC estimates.

We constructed the ENN by identifying the vertices that contained task rule, sensory stimulus, and motor response activations and by estimating the resting-state FC weights between them. a The input layer, consisting of vertices with decodable task rule and sensory stimulus activations. b Through activity flow mapping, input activations were mapped onto surface vertices in conjunction hubs. The activity flow-mapped vertices were passed through a nonlinearity, which removed any negative values. This threshold was chosen given the difficulty in interpreting predicted negative BOLD values. c The predicted conjunctive activations were then activity flow-mapped onto the motor output vertices, generating a predicted motor activation pattern. d These predicted motor activations were then tested against the actual motor response activations of other subjects using a four-fold cross-validation scheme. A decoder was trained on the predicted motor response activations and tested on the actual motor response activations of the held-out cohort (see Methods and Supplementary Fig. 1). e An equation summarizing the ENN model’s computations.