Computational Simulation of Wisconsin Card Sorting Task by Using Variational Recurrent Neural Network Based on the Free Energy Principle

The Wisconsin Card Sorting Task (WCST) is used to measure cognitive flexibility. In WCST, the participants are required to estimate underlying rules (called “category” in WCST) from given sensory signals. Computational modeling of the underlying cognitive mechanisms of WCST is important for elucidating flexible cognitive processing. In this study, we propose a hierarchical Recurrent Neural Network (RNN) model for explaining the underlying cognitive mechanisms of WCST, based on the free energy principle (FEP). FEP explains perception and goal-directed action as the minimization of prediction errors between predicted and real sensory signals (called free-energy minimization) and is expected to be an integrated theory of the brain. The primary characteristic of our model is that it considers free energy at future time steps, enabling it to correctly answer WCST as a goal-directed behavior based on the FEP. The simulation experiment showed that the proposed model successfully estimated the underlying categories to be estimated in WCST and correctly answer WCST. This indicates that the proposed model may provide mechanistic insights into flexible cognitive processing from the perspective of FEP.