Abstract
Rapid categorization paradigms, characterized by short presentation times and speeded behavioral responses, highlight both the speed and ease with which our visual system categorizes natural scenes. Existing computational models of perceptual categorization have been developed in the context of simple parameterized stimulus spaces. How these models generalize to natural scenes remains an open question. Here we consider a simple class of categorization models, which assumes that different categorization tasks correspond to different decision boundaries which curve the same perceptual space. We compute a simple measure of discriminability using a rudimentary visual representation (Oliva & Torralba, 2001) and a linear classifier to derive a decision value for individual images. Decision values thus reflect the expected difficulty to categorize individual stimuli in a task-dependent manner. In the first two experiments we validate the model’s main assumptions: In experiment 1, we demonstrate that the model predicts variations in accuracy and reaction time at the level of individual images using a natural/man-made categorization task. In experiment 2, we show that the model is consistent with variations in behavioral responses across tasks using published data from different groups (Greene & Oliva, 2009; Loschky & Larson, 2010; Kadar & Ben Shahar, 2012). Next, we show that the model provides a parsimonious explanation for the so-called superordinate advantage whereby superordinate categorization is faster and more accurate than basic categorization (Joubert et al., 2007). In experiment 3, we use the model to sample stimuli to design a reverse experiment, making participants’ superordinate categorization slower and less accurate than basic categorization. In experiment 4, we extend the model to reproduce distributions of reaction times and explain previously reported latency effects (Joubert et al., 2007). Overall, our results suggest that a simple model of visual categorization provides a parsimonious explanation for several published results and reported phenomena.
Meeting abstract presented at VSS 2014