Unraveling Representations in Scene-selective Brain Regions Using Scene-Parsing Deep Neural Networks

Abstract Visual scene perception is mediated by a set of cortical regions that respond preferentially to images scenes, including the occipital place area (OPA) and parahippocampal (PPA). However, differential contribution OPA PPA remains an open research question. In this study, we take deep neural network (DNN)-based computational approach investigate differences in function. first step, search for model predicts fMRI responses scenes well. We find DNNs trained predict components (e.g., wall, ceiling, floor) explain higher variance uniquely than DNN category bathroom, kitchen, office). This result robust across several architectures. On basis, then determine whether particular predicted differentially account unique PPA. explained navigation-related floor component compared wall ceiling components. contrast, are better combination floor, is, together contain structure texture scene. sensitivity suggests functions processing. Moreover, our results further highlight potential proposed as general tool investigation basis human perception.