Conference on Cognitive Computational Neuroscience

Humans are remarkably efficient at adapting to different contexts by exerting varying levels of cognitive control. Yet, it remains unclear how humans regulate cognitive control to respond to various contextual demands. In this study, we show that a recurrent neural network model trained to perform five different categorization tasks in an interleaved manner, spontaneously learned to adjust cognitive control in response to different switching rates (i.e., context). Specifically, our results indicate that the model learned to reduce separation of the task representations in task space when trained on more, versus less, task switches, resulting in smaller task switch costs. Finally, we also found that the same modulation of switch costs can be observed when varying the task switching probability during testing (as opposed to training), which emphasizes the importance of a confound-free test phase when studying learned control in behavior.