Humans can flexibly predict physical events by internally simulating object dynamics in the brain, a capacity lacking in current AI systems. Using a ball collision paradigm with visual occlusion combined with multimodal neuroimaging (fMRI/MEG), we uncover a spatiotemporally organized neural architecture for physical simulation. fMRI reveals hierarchical spatial segregation: higher order cortical regions encode relational physical variables, distinct from object-specific features encoded in the lower-order sensorimotor regions. MEG uncovers two temporally distinct neural processes: a real-time simulation tracking the evolving state of the object, in alignment with collision dynamics, and an early predictive signal anticipating collision occurrence (~700 ms before contact). We propose that this early predictive control mechanism dynamically allocates cognitive resources during simulation. This is formalized by a Dynamic Resource Intuitive Physics Engine (IPE) model, which captures both behavioral data and the dual neural timescales by optimizing accuracy-cost tradeoffs. Crucially, this framework predicts early encoding of another adaptive control variable separate from collision occurrence, as evidenced by both MEG and pupillary responses. These findings reveal how the brain achieves efficient physical inference through hierarchically organized predictive control that dynamically allocates cognitive resources.
Long, L., Wang, Q., Yang, Q., Chang, L.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 1
- Comments 0