Local recurrence accounts for extended processing during occluded-object recognition
Recognizing objects from incomplete visual input often requires processing beyond the initial feedforward sweep, but the relative contributions of local recurrence and long-range top-down feedback remain unclear. We combined source-localized magnetoencephalography (MEG), time-resolved decoding, backward masking, representational Granger causality, and computational modeling to examine these mechanisms during occluded-object recognition. We characterized neural dynamics in early visual cortex (V1-3), the lateral occipital complex (LOC), and inferotemporal-parahippocampal cortex (IT-PHC). Occlusion delayed the emergence of category information and elicited a late component that was selectively disrupted by backward masking, consistent with dependence on continued processing. These temporal changes occurred without detectable changes in the relative timing of regional responses or directed interareal interactions, including no measurable occlusion-related increase in feedback among the regions examined. To assess candidate computational mechanisms, we compared three nested model variants sharing a trained backbone: feedforward, local recurrent, and local recurrent with added long-range top-down feedback. Local recurrence improved recognition under occlusion and increased model-brain correspondence during the late, mask-sensitive interval, particularly in V1-3. Adding the implemented top-down pathway provided no consistent further benefit. Together, these findings favor prolonged local recurrent processing as an account of the additional computation supporting occluded-object recognition under the conditions tested.