Stability and flexibility in the Wisconsin Card Sorting Task emerge from goal-directed inference and planning under the free-energy principle
In the authors' words
In goal-directed behavior, the brain has to balance stability and flexibility. Stability means keeping the current recognition of environmental context, whereas flexibility means switching to a new one. The Wisconsin Card Sorting Task (WCST) is a standard task for studying this balance. In this task, stability is keeping the category estimate after correct feedback, and flexibility is switching it after incorrect feedback. Computational models have been proposed to elucidate the mechanism behind this balance. However, most models are given a variable for each category in advance, and every model treats correct feedback as desirable from the start. These models cover neither the category learning process nor the goal-directed process that adjusts the card selection to get correct feedback. To explain both processes under one principle, we modeled the WCST with a hierarchical recurrent neural network based on the Free Energy Principle (FEP). The model first infers the environment from observations and then plans the card selection to reach the goal. In both steps, the model minimizes prediction error based on the FEP. For inference, the error lies between the predicted signals and the received signals. For planning, it lies between the predicted feedback and the correct feedback. As a result, the latent states formed category-related clusters during learning. In the testing phase, the model kept the estimate after correct feedback and switched it after incorrect feedback. Removing the goal-directed process degraded the keeping and the switching. This implies that the two processes and this balance emerge from the FEP.
Appeared: Thursday, November 19. bioRxiv. Preprint, not yet peer-reviewed.