Topographic deep artificial neural network as a model of primate ventral visual stream

Abstract

Face processing in visual cortex has been widely studied and emphasized for its importance in primate survival and social communication. Monkey inferior temporal (IT) cortex contains neurons that respond preferentially to faces and cluster into several regions ("face patches") that together are referred to as the IT face process-ing network. While recent work has demonstrated that deep artificial neural net-works (ANNs) optimized for object categorization are strong predictors of neuronal responses at corresponding levels of the primate ventral visual stream (V1, V2, V4, and IT), those models do not explain the spatial organization of those neurons in general or the organization of the IT face processing network in particular. In this work, we test whether a new class of ANNs can naturally reproduce the core phenomena of the IT face network, including the rich spatial topography of face-selective neurons. Specifically, we designed and successfully trained topographic deep artificial neural networks (TDANNs) to solve a real-world object recognition task and to also minimize a proxy for neuronal wiring costs within each of the two highest IT levels (cIT and aIT). We report that layers of the trained TDANNs corresponding to cIT and aIT cortex contain clusters of face-selective units and reproduce core phenomenology of the face-patch system, such as connectivity between clusters and the emergence of viewpoint invariance. We also found that the model IT face network emerged over a range of naturalistic experience during training, but not for highly unnatural experience. Taken together, these results argue that the functional organization of the ventral stream might be explained by the need for the visual system to perform general, real world object categorization while also minimizing wiring costs over evolutionary and/or post-natal developmental time scales.