Hayato Idei, Keisuke Suzuki, Yuichi Yamashita
Mindfulness has established psychological benefits, such as stress reduction and emotional regulation; however, the underlying computational mechanisms, particularly in relation to mind-wandering, remain unclear. This study aimed to present a hierarchical recurrent neural network-based agent model grounded in the free energy principle to explore these processes within the framework of allostasis. The model integrates interoceptive, proprioceptive, and exteroceptive predictive processing, which minimizes the variational free energy throughout past and future contexts. Simulations of the resting state showed that inferred (mentally simulated) homeostatic context modulated attentional shifts between a 'being' mode (focused on present interoceptive perception) and a 'doing' mode (involving mentally simulated proprioceptive and exteroceptive sensorimotor behaviors). Further, we identified a meta-attentional parameter that controls these shifts by influencing cognitive attitudes toward internally generated beliefs across network modules spanning from the past to the future. By manipulating this parameter, we replicated mindfulness states ('deep-being' mode) in which attention remained focused on the present interoceptive state, despite minimal sensory prediction errors. This study offers insights into the computational dynamics of mindfulness and mind-wandering, laying the groundwork for the future exploration of consciousness and pure awareness, which is a state of consciousness devoid of content.