Skip to page navigation menu Skip entire header
Brown University
Skip 13 subheader links

Sensorimotor Learning of Depth Estimation for Perception and Action


Many actions that are important for successful behavior require information about an object’s threedimensional (3D) shape. Human observers estimate 3D shape based on patterns of stimulation on the retina. This retinal image is determined by many different sources of information (cues) that co-vary with 3D shape, like texture gradients and binocular disparities. However, cue information is inherently ambiguous with respect to 3D shape properties because an infinite number of potential structures may yield the exact same 2D retinal image. A major focus of research over the last 30 years has been to characterize normative models for how an observer should combine cues to afford successful sensorimotor interactions with their environments. These existing theories are largely deficient in specifying how cue combination may be impacted by learning processes. In particular, it remains unclear how learning of 3D shape operates, and what the consequences of 3D learning are for an observer’s actions. This dissertation attains two goals. First, we develop and evaluate a novel theory about how sensorimotor learning mechanisms simultaneously operate both 3D perceptual learning and learning of action-related variables in response to prediction errors about object shape. To do so, we conducted a psychophysical study and supplemented our findings with the derivation of a computational model of sensorimotor learning. Second, we compared two existing normative models of shape perception in their descriptive capacity to predict human cue combination: Maximum Likelihood Estimation (MLE) and Intrinsic Constraint. We psychophysically measured cue combination in human observers and derived predictions about task performance based on the computational specifications of either theory, allowing us to implement a quantitative comparison between theories for the first time. Taken together, this work provides evidence supporting the notion that human observers combine cues to maximize the sensitivity of their shape estimates with respect to the distal object’s true shape and that observers are constantly learning both 3D shape and action-related variables in parallel to reduce action errors.
Thesis (Ph. D.)--Brown University, 2022


Wilmott, James Paul, "Sensorimotor Learning of Depth Estimation for Perception and Action" (2022). Cognitive, Linguistic, and Psychological Sciences Theses and Dissertations. Brown Digital Repository. Brown University Library.