Fig. 7. Systematic alteration of ENN model architecture verifies validity of “full S-R model” results.

a We first benchmarked the motor response decoding accuracy for each hand separately using a standard cross-validation scheme on motor activation patterns for each hand (tested across subjects). This standard motor decoding was done independently of modeling sensorimotor transformations. b The full stimulus-response model, taking stimulus and context input activations to predicting motor response patterns in motor cortex. c The ENN model after entirely removing the conjunction hubs. d The ENN model, where we randomly sampled regions in the hidden layer (conjunction hubs) 1000 times and estimated task performance. e The ENN model after removing the nonlinearity (ReLU) function in the hidden layer. f The ENN model after lesioning connections from the task context input activations. g The ENN model, where we shuffled the connectivity patterns from the stimulus and context layers 1000 times. h Benchmarking the performances of all model architectures. Accuracy distributions (n = 1000) were obtained by running multiple iterations of the same cross-validation scheme (leave-4-out cross-validation scheme and randomly sampling within the training set; see Methods for clarification). Statistical testing was performed using a one-sided permutation test for each model separately (n = 1000 shuffled labels). For each iteration, we calculated a p value, and then averaged all p values. Boxplot maxima/minima reflect the 95% confidence interval, the box bounds define the 1st and 3rd quartiles of the distribution, and the center line indicates the median. Grey distributions indicate the null distribution generated from permutation tests (permuting labels 1000 times). (*** = p < 0.001; ** = p < 0.01; * = p < 0.01). i Summary statistics of model performances. Reported accuracy is the mean across the distribution; red indicates statistically significant decoding. Source data are provided as a Source Data file.