Conference on Cognitive Computational Neuroscience

Although several types of spatially-aggregated neural functional selectivities have been reported in the inferior temporal (IT) cortex of humans and monkeys, such as face, place, and body selectivities, broad swaths of IT have yet to be similarly characterized. Here, we present the first steps of using Topographic Deep Artificial Neural Networks (TDANNs) as hypothesis generators of not-yet-detected spatially-aggregated IT functional selectivities. To isolate the shared selectivities across a population of TDANNs, we applied hyperalignment to the IT layer of ten TDANNs. We then analyzed the shared underlying functional representations to identify eleven predicted neuronal functional selectivity clusters. After mapping these clusters back to the spatial IT maps in each TDANN, we find that face-selective units -- which spatially aggregate in TDANNs -- are strongly loaded on one of these functional clusters. On visual inspection, the other functional clusters appear to be selective for scenes, animal bodies, and mid-level object properties. Topographic ANNs, when analyzed in this manner, could be used to predict novel spatially-aggregated selectivities shared by all brains and to predict the spatial relationships between those functional aggregates. Both types of predictions could then be tested via targeted fMRI experiments.