Computational Models Recapitulate Key Signatures of Face, Body and Scene Processing in the FFA, EBA, and PPA

Abstract

Deep convolutional neural network (DNN)- based models have emerged as our leading hypotheses of human vision. We recently described DNN-based encoding models of the fusiform face area (FFA), the extrastriate body area (EBA) and the parahippocampal place area (PPA), that can predict the response to new images with very high accuracy. But do these models also explain the key experimental results from previous studies? Many stimuli in these studies were highly manipulated (e.g. isolated face parts in re-arranged spatial positions), far outside the domain of the natural images, and thus could provide strong tests of generalization. Further, these stimuli were designed to test, and taken as evidence for, classic “word model” hypotheses about visual representation (such as “holistic” face processing). Here we asked whether our current best encoding models directly replicate the main findings in prior published papers. To do this, we identified 20 influential papers that localized and reported response magnitudes of the FFA, PPA, and EBA. We tested the key experimental conditions directly on our encoding models of these regions without retraining the model. Our models could recapitulate all the key univariate and multivariate signatures of neural face, body, and scene processing described in those publications, including results previously taken to demonstrate holistic face processing, real-world size effects, sensitivity to animacy, and eccentricity bias. These findings show that our models generalize even outside their training domain. They also provide a computationally precise basis for findings previously described only with word models and show that these phenomena can emerge without any built-in, domain-specific biases or world knowledge apart from what can be gleaned from hierarchical computations. This approach is made possible because of our functional region-of-interest level of computational modeling and paves the way to efficiently test novel hypotheses completely in-silico in future work.