Abstract
<jats:p>Humans can rapidly extract summary statistics, such as the mean and variance of visual features, to efficiently represent complex visual environments despite limits in attention and working memory. Although ensemble perception is central to visual cognition, two fundamental questions remain unresolved. First, it is unclear whether ensemble statistics for features represented at different levels of the visual hierarchy rely on a common neural system or on separate feature-specific systems. Second, it remains unknown whether different summary statistics, such as mean and variance, rely on shared or dissociable neural mechanisms. Here, we used fMRI and multivariate pattern analysis to examine the neural representation of ensemble mean and variance across three visual features spanning the processing hierarchy: orientation (low-level), shape (mid-level), and animacy (high-level) (N = 24; two fMRI sessions). By combining whole-brain searchlight and ROI-based approaches, we found a graded division of labour between ventral and dorsal visual pathways. Although mean and variance ensemble statistics were distributed across the visual cortex, mean decoding was stronger in ventral regions, whereas variance decoding was stronger in dorsal regions. Ensemble mean representations followed a posterior-anterior gradient within the ventral visual pathway, consistent with increasing abstraction from orientation to shape and animacy, and showed little anatomical overlap suggesting largely feature-specific. By contrast, ensemble variance was weighted toward dorsal parietal and frontoparietal regions, especially superior parietal cortex and intraparietal sulcus, decoding clusters largely overlap across features and generalized robustly across orientation, shape, and animacy. Together, these findings provide a more nuanced account of ensemble perception, showing that feature-specific and feature-independent neural codes can coexist across visual cortex and help reconcile previously conflicting evidence.</jats:p>