Abstract
Investigating the function of the various subregions of the visual system is a major goal in neuroscience. One approach is specifying to which types of stimuli they show the strongest response to, however given the variety of the visual world it is impossible to present all possible stimuli in-vivo. We follow an alternative approach to reach this goal. We trained a convolutional neural network-based model of the occipitotemporal cortex to match the behaviour of an individual brain reacting to visual input, using a large-scale functional MRI dataset (Seeliger & Sommers 2019). This model allowed us to predict voxel responses in several areas defined functionally (such as FFA, LOC, PPA) and from an anatomical atlas (such as PHC, VO) in-silico. To identify the preferred stimuli for voxels in these areas we developed an interpretability technique for convolutional neural networks, based on a generative adversarial neural network (GAN), and using gradient ascent for synthesizing naturalistic preferred images. As expected, voxels in areas V1-V3 yielded small receptive fields with a preference for gratings. Higher order areas showed mixed preference: For instance, while confirming face-selectivity in FFA and place-selectivity in PPA, both regions additionally showed preference for other visual features, such as oval shapes and vertical lines in FFA, or horizontal lines and high spatial frequency in PPA. The GAN latent vectors for the investigated subregions were highly distinct, as shown in classification tasks, underscoring the validity of the results. This approach opens the avenue towards precision functional mapping of selectivity at the level of individual voxels across the whole visual system, and may reveal previously unknown functional selectivity.