Orthogonal representations of object category and location in object selective cortex

Abstract

Theories of object recognition often call upon the notion of invariance to account for our ability to recognize objects across many naturally occurring transformations (e.g. size and viewpoint), including changes in spatial position. To investigate this, we measured brain activity using fMRI while subjects viewed four categories of objects (faces, houses, animals, and cars) displayed to four locations in the visual field. In our analysis, we first used principal component analysis to decompose the responses of object selective areas into a set of orthogonal component activation patterns. These components were used by a linear classifier to decode the category of the object displayed to the observer. The results of the classification analysis showed that object selective areas robustly code the position of objects. We further investigated how the brain supports our ability to recognize objects across different locations. In an examination of the individual components, we found subsets that retain the ability to extract object category information across different locations. These components, which are orthogonal to components that rely on position, could be utilized to solve the problem of invariance. Notably, this represents a population response solution, which is consistent with recent modeling efforts. The orthogonality of these representations is exemplified in our analysis of FFA and PPA. In terms of the components that specified object category, the specialization of these areas was clear. The preferred category had more object category components; these components also accounted for a large proportion of the object related activity in these areas. In contrast, the representation of space in these areas was found to be roughly equivalent for the preferred and non-preferred object categories.