Neural coding of object contour shape in primate posterior inferotemporal cortex

Primates possess an impressive ability to effortlessly recognize the shapes of a nearly infinite number of visual objects, even across dramatic changes in size and position. The fact that human performance far exceeds the most sophisticated machine vision algorithms attests to the complexity of the underlying neural mechanisms, though they are not yet well understood. The goal of the thesis research presented here is to quantitatively characterize the neural code for a critical aspect of object form, 2D boundary shape, within macaque monkey posterior inferotemporal cortex (IT), an important high-level stage of neural object processing.; Object shape information is processed in the ventral stream of primate visual cortex, a hierarchical network that transforms the retina's simple, pixel-like object representation into one that is more appropriate for efficient, invariant recognition. Although the earliest processing stages are well understood at a general, computational level, studies of higher-level stages have produced mostly simple phenomenological descriptions of neural response selectivity. Here, we measure posterior IT single-unit responses to large, parametric sets of abstract shape stimuli, and we characterize the resulting neural selectivity with explicit models of the transformation between shape stimuli and IT unit responses. We show that IT neural responses are tuned for the shapes and relative positions of multiple contour regions (parts) within stimulus objects, consistent with a selective combination of a small number of single-part selective inputs from afferent area V4. We found that distinct groups of IT neurons combine information about multiple parts using linear summation---producing responses that are ambiguous with respect to the number and type of parts present---and using nonlinear (conjunctive) summation, producing responses that explicitly encode the similarity of a stimulus pattern to an optimal multi-part configuration. Finally, we show that IT responses exhibit a transition in time following stimulus onset---from early linear responses to later nonlinear responses---indicating that IT shape selectivity is generated via dynamic processing mechanisms more complex than a strictly feedforward network. These results advance the study of neural object coding by producing the first general, quantitative characterization of the principles underlying object shape selectivity in IT.