Neural representation strength of predicted category features biases decision behavior

Abstract Theories of prediction-for-perception propose that the brain predicts the information contents of upcoming stimuli to facilitate their perceptual categorization. A mechanistic understanding should therefore address where, when, and how the brain predicts the stimulus features that change behavior. However, typical approaches do not address these predicted stimulus features. Instead, multivariate classifiers are trained to contrast the bottom-up patterns of neural activity between two stimulus categories. These classifiers then quantify top-down predictions as reactivations of the category contrast. However, a category-contrast cannot quantify the features reactivated for each category–which might be from either category, or both. To study the predicted category-features, we randomly sampled features of stimuli that afford two categorical perceptions and trained multivariate classifiers to discriminate the features specific to each. In a cueing design, we show where, when and how trial-by-trial category-feature reactivation strength directly biases decision behavior, transforming our conceptual and mechanistic understanding of prediction-for-perception.