Abstract
A recent computational model (Ullman, Vidal-Naquet, & Sali, 2002) proposes that viewpoint-dependent features of intermediate complexity (IC) are more informative about an object's category than either highly complex or very simple features of the object. Consequently, such IC features should be optimal for object categorization. We tested the psychological reality of this model by assessing the influence of the information delivered by a feature (mutual information - MI) on categorization speed and on ERPs. Participants categorized IC features of three object categories (faces, cars and various-class images) of varying MI levels. Both RT and accuracy of performance increased as a function of MI for both car and face features but not for the various-class-defined features. Faces were classified better than cars and MI value influenced differently the two classes, possibly pointing to the effect of different visual experiences with faces and cars. The face-sensitive N170 component was larger to face features than to car features and various-class features and delayed relative to full faces and objects. Moreover, only for faces the N170 amplitude increased as a function of features' MI. We conclude that IC features play a dominant role in object categorization and that MI has a psychological and neural basis. IC features extraction by maximization of information may serve as a new model for the detection of primitives evident in N170 and will thus allow a better understanding of the processes underlying face detection.