Abstract
Perceptual expertise involves discrimination of stimuli at subordinate level of categorization. The question of whether perceptual expertise is mediated by general expertise or domain-specific mechanisms has been hotly debated. To decide between these two hypotheses, previous studies have asked whether objects of expertise share the same computations that are used for face recognition, for which all humans are experts. A main limitation of these studies is that human experience with faces is more extensive than with any object of expertise. Current computational models of object recognition enable us now to re-evaluate this question with models that are matched for the amount of experience with different categories of expertise. We evaluated computational similarity by measuring the learning curve and performance level of a deep convolutional neural network (DCNN) that was pre-trained for subordinate-level categorization of one category (faces), to re-learn subordinate-level categorization of a new category (birds) and compared to a DCNN that was pre-trained for basic-level categorization (objects). A general expertise hypothesis predicts that a face-trained DCNN will show faster learning and higher performance for birds than an object-trained DCNN. In contrast to this prediction, an object-trained network learned to discriminate birds much faster and reached higher performance level than a face-trained network. Another measure that was used to indicate whether objects of expertise share similar computations is the inversion effect. We therefore examined performance of a face-trained and a bird-trained DCNN for identity matching of upright and inverted objects, faces and birds. A face-trained DCNN showed an inversion effect only for faces and a bird-trained DCNN showed an inversion effect only for birds, indicating that the inversion effect is a specific within-category effect rather than shared among different categories of expertise. Taken together, these findings suggest that perceptual expertise is mediated by domain-specific rather than general expertise mechanisms.