Fig. 2: Overview of experimental procedure.

The behavioral data of 17 healthy controls (H-subjects)6,7 were used as training data. Ten recurrent neural network (RNN) models, referred to as model subejcts (M-subjects), were generated for each H-subject. After we checked whether each M-subject successfully reproduced the behavioral features of the SoA task in the corresponding H-subject test data, four lesion experiments were simulated through M-subjects, including the hypothesis in the current study. We evaluated the similarity in SoA judgment between behavioral data in patients with schizophrenia and the lesioned M-subjects.