"Discovering the Visual Structure of Our World"
Identifying the visual structure of our environment is critical for navigation, tracking, prediction, and pursuit. This structure is informed by relative motion of entities that make up this environment. However, how exactly we use visual motion information to infer the environment's structure remains unclear. In this talk I will take a normative perspective, and assume that humans perform Bayesian structure inference with a relatively simple structured motion model. First, I will demonstrate that this model explains human object tracking performance in two experiments better than alternative heuristics. Second, I will show that it also does so in a human structure discrimination task. In both cases, the model operates "offline" by making decisions after having observed a whole trial's visual motion information at once. Third, I will modify the model to perform human-like "online" inference by updating the inferred structure with every new piece of incoming visual motion information. This model turns out to replicate a range of previous structure motion experiments qualitatively, as well as provides a better quantitative fit to one of our previous experiments. Lastly, I will describe a neural network implementation of this model, that suggests possible network motifs required to perform such inference, and can be tested in future experiments.
A32 Lecture Hall or via Zoom: https://princeton.zoom.us/j/96233413571?pwd=QjhIUHFiL3BWREpCbUJkLzFtZVpLZz