Fig. 6. Part-whole relations in deep networks.

a Schematic showing the perceptual representation of objects with a break introduced either at natural or unnatural part cuts. b Part processing index across layers of the VGG-16 network (blue) and VGG-16 with random weights (brown). The dashed line represents the effect size estimated from human visual search on the same stimuli³⁰. c Schematic showing how the same object can be broken into either natural or unnatural parts. The natural part advantage is calculated as the difference in correlation between part-sum models trained to predict whole-object dissimilarities using the parts (see he “Methods” section). d Natural part advantage across layers of the VGG-16 network (blue) and VGG-16 with random weights (brown). The dashed line represents the effect size estimated from human visual search on the same stimuli³¹. e Perceptual representation of hierarchical stimuli. The left and middle images differ only in global shape whereas the middle and right images differ only at the local level. According to the global advantage effect, a change in global shape is more salient than a change in local shape. f Global advantage index across layers of the VGG-16 network (blue) and VGG-16 with random weights (brown). The dashed line represents the effect size estimated from human visual search on the same stimuli⁶³.