Redundancy reduction in motor control

Research in machine learning and neuroscience has made remarkable progress by investigating statistical redundancy in representations of natural environments, but to date much of this work has focused on sensory information like images and sounds. This dissertation explores the notions of redundancy and efficiency in the motor domain, where several different forms of independence exist. The dissertation begins by discussing redundancy at a conceptual level and presents relevant background material. Next, three main branches of original research are described. The first branch consists of a novel control framework for integrating low-bandwidth sensory updates with model uncertainty and action selection for navigating complex, multi-task environments. The second branch of research applies existing machine learning techniques to movement information and explores the mismatch between these methods for extracting independent components and the forms of redundancy that exist in the motor domain. The third branch of work analyzes full-body, goal-directed reaching movements gathered in a novel laboratory experiment, using explicitly measured information about the goal of each movement to uncover patterns in the movement dynamics. Each branch of research explores redundancy reduction in movement from a different perspective, building up a sort of catalog of the types of information present in movements. Redundancy is discussed throughout as an an important aspect of movement in the natural world. The dissertation concludes by summarizing the contributions of these three branches of work, and discussing promising areas for future work spurred by these investigations. More detailed models of voluntary movements hold promise not only for better treatments, improved prosthetics, smoother animations, and more fluid robots, but also as an avenue for scientific insight into the very foundations of cognition.