Computational modeling of visual motion processing neurons in the dorsal medial superior temporal area (MSTD): Functional architecture and learning mechanisms

Neurophysiological studies in monkeys suggest a role for the medial superior temporal area (MST) in the analysis of optic flow. The dorsal section of MST (MSTd) contains neurons selective for large-field planar, "spiral-space" (radial, circular, and spiral), and shear motions, but few for deformations, whereas the main input to MSTd is from small-field directional motion responses of middle temporal area (MT) neurons. Although the patterns of connectivity contributing to the tuning of MSTd neurons are still largely unknown, computation models allow exploration of different hypotheses.;Most models of MSTd obtain selectivity to motion patterns from excitatory contributions of sets of MT-like units representing several directions of motion. Based on electrophysiological studies, Duffy and Wurtz [Journal of Neurophysiology 65 (1991) 1346] proposed an alternate hypothesis: that responses of MSTd neurons may result from as few as two spatially-graded motion profiles ("planar motion gradients") combined with excitation and inhibition. Here we propose a computational model of feedforward connectivity between MT and MSTd that combines pairs of excitatory and inhibitory planar motion gradients. Evaluating our MSTd units with a set of planar, spiral-space, shear and deformation patterns, we identify parameters of planar motion gradients that lead to tuning consistent with MSTd neurons, including Gaussian tuning to spiral-space patterns, position-invariant responses, and center of motion tuning. Furthermore, we demonstrate that local motions across spiral space contain an inherent Gaussian distribution that may contribute to the Gaussian tuning of MSTd neurons.;Observed improvements in position invariance resulting from the addition of lateral interactions to the feedforward model motivated us to perform extensive learning simulations using Hebbian and anti-Hebbian mechanisms to establish excitatory and inhibitory lateral connectivity. Both learning mechanisms result in Gaussian-shaped weight profiles and improvements in position invariance for spiral-space patterns shifted to peripheral regions of the receptive field.;This research suggests both a functional architecture for area MSTd that produces selectivity to motion patterns from a minimal set of planar motion selectivity and learning mechanisms that increase the position invariance of MSTd neurons.