Abstract
Seminal work has shown how humans can be highly accurate in judging their direction of self-motion (heading) from optic flow, to within 1° (Van den Berg, 1992; Warren, Morris, & Kalish, 1988). Remarkably, accuracy only decreases to ~3° in displays containing as few as two moving dots, which suggests that a sparse flow field is sufficient to drive the underlying neural mechanisms. Indeed, receptive field models fit with only a few input connections do a good job at capturing single neuron data in MSTd (Mineault, Khawaja, Butts & Pack, 2012), an area of primate cortex that has been shown to be causally linked with heading perception (Gu, Deangelis, & Angelaki, 2012). This is difficult to reconcile with many biologically inspired models of heading perception that rely on full-field (e.g. radial) connection templates to integrate motion across the visual field. In the present study, we used neural modeling to investigate how sparse connectivity between areas MT and MST may shape heading perception. We found that sparse connectivity yields heading estimates more consistent with human heading judgments than a densely connected model under a range of dot density and noise conditions. The model builds on the Competitive Dynamics model (Layton & Fajen, 2016), which relies on the pattern tuning of active MSTd heading cells to recover object motion in a world-relative reference frame. We leveraged sparse connectivity to efficiently simulate large numbers of MSTd cells tuned to complex combinations of speed, direction, and disparity inputs, which allows the model to accurately estimate object motion in natural cluttered environments, not just under idealized conditions. Our findings support the intriguing possibility that the sparse connectivity structure of MSTd may influence heading and object motion perception.
Acknowledgement: ONR N00014-18-1-2283