Abstract
Prior work has identified regions of high-level visual cortex selectively responsive to faces, places, bodies, and words. However, this largely hypothesis-driven work cannot reveal how prominent these category selectivities are in the overall functional organization of visual cortex, or what other unhypothesized selectivities exist. Further, standard voxel-wise tests cannot detect selective neural populations that coexist with functionally distinct populations within voxels. To overcome these limitations, we applied data-driven voxel decomposition analyses and generalized canonical correlation analysis to identify a robust set of canonical response profiles consistent across subjects in a recently-released public data set of fMRI responses in eight participants to thousands of complex photographic stimuli (Allen et al 2021). Because these analyses permit many degrees of freedom, our strategy is to freely explore only four of the participants, register our hypotheses, and test them on the held-out participants. To date, the first four participants reveal components in the ventral pathway clearly selective for people, scenes, and words, replicating prior results, as well as an intriguing novel component that appears to respond selectively to images of food. Although accounts of this “food” component in terms of low- or mid-level visual properties remain possible, it does not emerge from similar analyses of V1/V2 or from activations to these stimuli in early layers of Alexnet. Analyses of lateral visual cortex reveal components apparently selective for implied motion and social groups, along with other novel components. We find no evidence of components selectively responsive to animals or tools, or to “stubby” or “elongated” shapes. The hypotheses emerging from these analyses about neural selectivities (and lacks thereof) will be refined, registered, and then tested in the held-out participants. We expect our data-driven analyses to powerfully validate some but not all previously reported selectivities and to identify novel selectively-responsive neural populations in high-level visual cortex.