Figure 3

Method summary, network architecture, training, and cross-validation for the primary dataset. (a) Concurrent EEG and fMRI data were obtained. Utilizing structural MRI, a four-concentric head model was constructed and employed to reconstruct preparatory EEG signals at (b) regions localized by fMRI, showing heightened activation during Incongruent versus Congruent conditions. (c) The table includes detailed statistics that outline the specific brain regions identified through fMRI analysis. (d) The time-series of the fMRI-guided EEG sources (i.e., Dipoles) were concatenated (Dipoles × Time: 19 × 726) to generate input for the convolutional neural network (CNN). The CNN contains only two layers with trainable parameters: 1D Convolutional and Dense layers. The EEG signals were partitioned into training (60%), cross-validation (30%), and test (10%) sets. The training and cross-validation sets were utilized to train the CNN for discriminating Hit from Error trials. (e) The training rate diminishes exponentially in staircase fashion over Epochs. The dotted black line represents the underlying exponential function, while the green line represents the actual training rate per epoch. (f) Progression of training and cross-validation accuracies over epochs. (g) Progression of training and cross-validation performance losses over epochs.