Title: Network mechanisms of dynamic feature selection for flexible visual categorizations

Abstract: Visual categorization is the pervasive cognitive function with which humans make sense of their external world. Current theories and models assume that the visual hierarchy reduces its high-dimensional input, by flexibly selecting the stimulus features that support categorization behavior. To investigate such feature selection, we quantified top-down task effects on the dynamic representation of stimulus features. We recorded the MEG as participants categorized the same scene images in different ways (in blocks of face expression, face gender, pedestrian gender, and vehicle type trials, within-participant). We show where, when and how brain networks select categorization features. Specifically, occipital MEG sources represent all stimulus features ~50-170ms post-stimulus, but with opponent response, to select or reduce the same stimulus features depending on their task-relevance. Following this 170 ms junction, occipito-ventral and dorsal sources represent only task-relevant features for behavior. Our results therefore show how systems-level network mechanisms dynamically reduce input dimensionality for behavior, by selecting task-relevant stimulus features.