Abstract
Natural scenes are comprised of collections of objects with specific types of objects tending to occur in certain classes of scenes. We hypothesize that the visual system might exploit these co-occurrence statistics in order to classify scenes more efficiently. If this is true, then a model that captures the distribution of objects in natural scenes should provide good predictions of visual cortical activity during natural vision. To construct such a model we adapted a recent probabilistic algorithm known as Latent Dirichlet Allocation (LDA). Given thousands of object-labeled natural images, LDA analyzes the label co-occurrences and ‘learns’ the distribution of objects in various scene classes. (The number of scene classes is a free parameter determined by the modeler; the specific scene classes are latent states learned by LDA). We then presented thousands of natural scenes to subjects while recording fMRI-BOLD activity in retinotopic and object-selective visual cortex. Afterwards, we estimated the scene-specific selectivity of all recorded voxels by regressing the BOLD responses evoked by each image onto the scene classes provided by LDA. We find that specific regions within lateral and ventral occipital-temporal areas are selective for various specific classes of natural scenes. (This is consistent with previous results; Naselaris et al. 2009). Selectivities, as determined by our method, are generally consistent with selectivities defined by standard functional localizers. However, we also find scene selectivity in many areas that are not identified by standard functional localizers. Finally, in order to determine how many distinct scene classes might be represented in anterior visual cortex, we varied the number of scene classes learned by LDA. The best model descriptions were obtained when the model learned 8-18 scene classes. In summary, our results suggest that regions within occipital-temporal visual cortex represent the distribution of objects in certain specific categories of natural scenes.