Mid-level features are sufficient to drive the animacy and object size organization of the ventral stream

摘要

The ventral visual stream exhibits a large-scale organization by the dimensions of animacy and real-world size (Konkle & Caramazza, 2013). While this organization could reflect a conceptual-level distinction, there are also perceptual differences between animals, big objects, and small objects, evidenced by systematic mid-level feature differences (Long et al., 2016). Here, we examined whether these mid-level feature differences are sufficient to drive the ventral stream organization by animacy and object size. Using a customized texture synthesis model, we created a stimulus set from images of big animals, big objects, small animals, and small objects. These stimuli preserve mid-level information about texture and form ("texforms"), but are unrecognizable at the basic-level. Eight observers underwent functional neuroimaging while viewing texform images and the original counterparts. For each voxel in the ventral stream, we computed animacy preferences (animals a?? objects) and real-world size preferences (small objects a?? big objects), for both original and texforms images. These preference maps were correlated with each other to assess the similarity of the animacy and real-world size topographies. Overall, we found that texforms elicited robust animacy and size preference maps across the entire ventral stream. Further, these maps exhibited a highly similar spatial topography to the maps generated when observers viewed recognizable images (Animacy: r=.76, t(7)=22.3, p 0.001; Size: r=.43, t(7)=5.7, p 0.001). In a replication experiment, we presented images above and below fixation, and again found that texforms and original images generated similar topographies (Animacy: r=.69, t(7)=9.8, p 0.001; Size: r=.36, t(7)=5.4, p=0.001). These results demonstrate that the ventral stream organization does not rely on intact basic-level recognition, and that mid-level feature differences drive ventral visual cortex in a way that is not simply inherited from retinotopy. Broadly, these data imply that mid-level perceptual information is represented along the ventral stream well beyond early visual areas.