Abstract
Theories of top-down search guidance typically assume that guidance to a distractor is proportional to the object's similarity to a target. This relationship, however, has been demonstrated only for simple patterns; it is less clear whether it holds for realistic objects. We report a novel method for quantifying guidance by reading the subject's mind, defined here as classifying the target of a categorical search task (either teddy-bears or butterflies) based on the distractors fixated on target-absent trials. The task was standard present/absent search. Half of the subjects searched for a teddy-bear target, the other half searched for a butterfly target. Except for the targets, search displays were identical between the two groups, meaning the same distractors in the same locations. All distractors were random real-world objects selected from the Hemera collection. To quantify target-distractor similarity we used a machine learning method (AdaBoost) and new target exemplars to train a teddy-bear/butterfly classifier. Target-absent trials were then combined across the teddy-bear and butterfly groups, and the distractors selected by gaze on these trials were identified and input to the classifier. The classifier evaluated these objects in terms of color, local texture, and global shape feature similarity to the teddy-bear and butterfly classes, then assigned each object to one of these target categories. Our joint behavioral-computational method correctly classified 76% of the actual butterfly target-absent searches and 66% of the teddy-bear target-absent searches, lower than the butterfly classification rate but still significantly better than chance (50%). These results definitively prove the existence of categorical search guidance to real-world distractors; in the absence of guidance above-chance classification would not have been possible. Our method also demonstrates that these guidance signals are expressed in fixation preferences, and are large enough to read a subject's mind to decipher the target category of target-absent searches.
This work was supported by NIH grant R01-MH63748.