Explaining the heterogeneity and topography in inferior temporal cortex with deep neural networks (360G-Wellcome-206521_Z_17_Z)

Object recognition is accomplished by the ventral visual pathway, also known as inferior temporal (IT) cortex. IT contains regions that respond preferentially to faces, color, and places. However, we still do not understand the principles underlying the emergence and organisation of these regions. Currently, there are no computational models that take into account the heterogeneity of IT, including the spatial clustering of neurons with similar selectivity and the spatial organisation of these clusters on the cortical surface. Therefore, we would like to answer two questions regarding IT: 1) Does the heterogeneity of IT simply emerge from exposure to many images, or are specialized mechanisms and/or innate constraints involved in its development? 2) What are the organisational principles reflected in IT topography? We will determine whether the heterogeneity of populations of units, like that empirically observed in IT, can emerge spontaneously in deep neural networks (DNNs) without specialized architecture, or whether we need an architecture that specifically accounts for it. We will then build a DNN that incorporates the spatial information about selective populations of units and assigns these populations to topographical maps. We will test the models’ predictions using electrophysiology and fMRI data.