Action prediction relies on high-level holistic and low-level kinematics priors

Our brains continuously predict what will happen next, helping us interact smoothly with others and navigate a dynamic world. A key example is biological motion perception, which relies on “priors,” or lifelong expectations about how bodies move. Yet it remains unclear how these priors shape the dynamics of neural prediction across hierarchical levels. Using a new dynamic extension to representational similarity analysis (dRSA), we measured how the brain of human participants predicted continuous dance movements recorded with magnetoencephalography. Under normal viewing conditions predictions followed a hierarchy: high-level viewpoint-invariant body motion was predicted earliest, while viewpoint-dependent body motion and low-level pixelwise motion were predicted closer to real-time. Disrupting high-level long-term priors by inverting videos selectively reduced high-level predictions. Instead, disrupting low-level short-term kinematics priors by piecewise scrambling eliminated all motion prediction, while increasing post-stimulus responses. These findings reveal how high- and low-level priors jointly shape predictive perception of naturalistic continuous input.