Abstract
We present a novel account of the origin of domain-selective regions in ventral temporal cortex. We use a deep convolutional neural network model of early visual cortex as input to a multi-layer recurrent map-like model of ventral temporal cortex, and train this model to jointly recognize faces, objects, and scenes. We implement spatially-restricted receptive fields within and between VTC map layers, and restrict feedforward connectivity to be excitatory. Learning in the model results in the development of smooth topographic domain-selectivity for faces, objects, and scenes, especially in more “anterior” layers of the network. We confirm the functional significance of domain-selectivity using searchlight and lesion analyses. By contrast, the network does not develop topographic domain representations without the excitatory restriction on feedforward connectivity. Further, we implement a more biologically detailed version of the model in which neurons can be only excitatory (E) or inhibitory (I) in their influence on other neurons (Dale’s Law), with only E units projecting feedforward connections. Using two overlaid maps of E and I neurons per VTC area, we again find topographic domain organization in VTC layers. Moreover, E and I units develop column-like responses with overlaid selectivity profiles. In contrast to classical self-organizing map models, we find that spatially broader inhibition is not needed to explain topographic organization. Rather, broadening inhibition relative to excitation gives rise to finer-grained patterning of multiple domain-selective regions whose spatial profiles can be further tuned by the receptive field size. Finally, whereas previous work has simulated topography by explicitly encouraging an exponential decay of pairwise unit response correlation as a function of unit distance (Lee et. al, 2020), this result emerges naturally from learning in our model. In sum, we show that biological constraints on network connectivity can produce topographic domain-selectivity in a distributed neural network without innate domain-specificity.