In the last couple of decades, categorization experiments using natural images have started to unravel the way our brain represents and categorizes visual information. Unfortunately, natural stimuli lack any clear parameterization, and therefore many of those experiments are difficult to control, analyze, and interpret. This led investigators to interpret their results using abstract and general theories that lack quantitative predictions. Moreover, these interpretations are sometimes inconsistent with predictions from perceptual categorization models, since it is unclear how these models can be applied to natural stimuli. In this thesis we bridge the gap between categorization models and natural scenes categorization experiments, and show that rapid visual categorization results are a simple outcome of discriminability between two categories. Towards this, we derive a simple measure of natural scene discriminability using a decision boundary, learned over a rudimentary visual representation. We then measure the difficulty of classifying an individual image, by measuring its distance to the decision boundary. Empirical distributions of discriminability values are used to predict human accuracy and reaction times across tasks and stimuli sets. We then extend the model to predict full reaction time distributions. Using a combination of simulations and psychophysics experiments, we validate the model assumptions, show that it is consistent with a large set of published results, and predict novel categorization effects. Our work greatly improves on the current status in several aspects: First, the theory generates quantitative and testable predictions of human's speed and accuracy in natural scene categorization tasks, both at the single image level and the category level; second, it allows to control and design experiments based on measurable visual information; third, it provides a simple, elegant, and intuitive interpretation of novel and existing results based on discriminability between targets and distractors.
Sofer, Imri,
"Setting boundaries in space: A model of rapid visual categorization of natural images"
(2014).
Cognitive Sciences Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://doi.org/10.7301/Z06971ZN
