"The time course of visual and conceptual contributions to scene categorization"
Visual scene understanding is remarkably rapid. Over the past two decades, a number of features have been hypothesized as mechanisms that facilitate this ability, ranging from low-level visual features such as color and contours, to high-level conceptual features such as attributes and affordances. However, significant correlations exist across these features, making it difficult to assess the independent contributions of any given feature to scene categorization. We obtained representational dissimilarity matrices (RDMs) from a number of leading visual and conceptual models and used a whitening transformation to decorrelate them while retaining their original interpretation. Using high-density EEG, we used a combination of decoding and multivariate regression to assess feature use over time. We found that although simple visual features explained more variability in EEG signals, nearly none of it was shared with behavioral assessments. By contrast, higher-level conceptual features explained less and later variance in EEG, but nearly all of it was shared with behavior. Taken together, these results paint a picture of scene categorization being primarily a conceptual process that is reliant on previously processed visual features.