Lauren Williams¹ (), Xin Yang, Joshua Langfus², Yarrow Dunham³, Justin Halberda¹; ¹Johns Hopkins University, ²University of California, San Francisco, ³Yale University

How early in visual processing are faces marked for category membership like race and gender? We addressed this question using change detection paradigms adapted from visual working memory literature, where one face from an array of faces changes to another. Crucially, the two faces could come from the same or different gender or racial categories. We predicted that, if gender and race are coded in VWM, participants would be faster to find the changing face when it crosses one of these boundaries. With both accuracy (Study 1) and response time (Studies 2 and 3) measures, we found that people spontaneously categorize faces based on race and gender in VWM, detecting changes faster and more accurately when they cross category boundaries. This pattern persisted in both standard change detection (Study 1) and flicker change detection (Study 2) paradigms, and occurs even when participants complete only a single trial (preventing shifts to higher-level strategies; Study 3). To rule out the possibility that the positive effects that we observed can be explained by lower-level perceptual features or higher-level subjective features associated with faces, we incorporated a large set of such features into our analysis (luminance, feature shapes, affect, attractiveness, etc.). Including these features in our primary analysis did not diminish the predictive power of crossing a category-boundary; thus, the effects in working memory that we report are not reducible to these rich sets of features. Importantly, we find this in a working memory task that does not require conscious attention to categories (providing evidence of spontaneity), that does not appear to be reducible to effects of lower-level perceptual similarities (providing evidence of categorization over and above perceptual similarity), and in a task in which attention to categories is not task-relevant and cannot in general improve performance if employed as a deliberate strategy.