Abstract
We reported previously that responses of individual neurons in macaque monkey inferotemporal cortex (IT) convey information about 3D medial axis shape (Hung et al., Neuron, 2012). Specifically, many IT neurons signal configurations of medial axis elements (connected torsos and limbs) in terms of 3D position, orientation, curvature, and connectivity. We hypothesized that these neurons provide an explicit, efficient shape code for elongated, branching objects such as vertebrate animals. Here, we analyzed the strength of IT neural population responses to projecting limbs (medial axis elements that have a termination on one end). Our dataset comprised spiking responses of 111 IT neurons, each tested with 400–600 3D medial axis shapes presented on a computer monitor using shading and binocular disparity as cues for shape-in-depth. These shapes were initially random but evolved through multiple generations based on a genetic algorithm driven by the neuron's responses. Thus, our sampling strategy converged toward high response shapes in later generations. We characterized projecting limbs in terms of their object-centered 3D position, 3D orientation, curvature, and surface shape. We binned the position/orientation/curvature/surface space into a multi-dimensional matrix. For each stimulus, we summed the neural response into the bins occupied by that stimulus. We used plots and statistical tests to analyze anisotropies in the resulting matrix. One major trend was over-representation of vertical limbs, that is, limbs projecting downwards or upwards. We hypothesize that this over-representation of vertical limbs reflects the prevalence and/or ecological significance of vertical projections in the natural world.
Meeting abstract presented at VSS 2017