Abstract
How is the representation of complex visual objects organized in inferotemporal (IT) cortex, the large brain region responsible for object recognition? Areas selective for a few categories such as faces, bodies, and scenes have been found, but the vast majority of IT is “wild,” lacking any known specialization, leading to uncertainty over whether any general principle governs IT organization. We recorded responses of IT cells in macaque monkeys to a set of objects, and built a low dimensional object space to describe these objects using a deep network Alexnet, by performing principal component analysis on the responses of fc6 layer. We found that responses of single IT cells could be well-modeled by projection onto specific axes of this object space. Remarkably, cells were spatially clustered into networks according to their preferred axes, and this clustering was determined by the topography of object space. Furthermore, this topography was replicated across at least three hierarchical stages of increasing view invariance. Finally, pooling cells across this newly identified object-topic map allowed reconstruction of general objects. Taken together, these results provide a unified picture of IT organization in which category-selective regions are part of a topographic continuum.