Title: Tracking the Evolution of Learning on a Visualmotor Task
(Master of Science Thesis Defense)
Speaker: Sameer Siruguri
Abstract: How do humans acquire skills on complex visualmotor tasks? The aim of explaining the evolution of learning is to bootstrap learning on such tasks: we analyse a large sequential corpus of low-level visualmotor data and construct models of the learning exhibited during the generation of this data that can expose the subject's inadequacies and thus help correct them. The data-driven construction of such models is challenging because visualmotor performance data is
  • 1. non-stationary---characterized by periods of slow evolution of a policy punctuated by conceptual shifts in which policies change radically
  • 2. sparse---available data occupies a millionth part of the high-dimensional space it is drawn from and
  • 3. has non-Gaussian noise, which is difficult to model and eliminate.
We develop data-driven algorithms that build and track models of the policies across the conceptual shifts that are behaviorally equivalent to the subject being modeled. We present a novel method for locating changepoints in the data, i.e. points of radical policy shifts. We construct ``local'' models for stationary policies between two changepoints, which are fitted by a novel extension to locally weighted regression. These local models are rich enough to capture individual differences in the task, and are simple enough to learn in real-time. We experimentally demonstrate the effectiveness of our methods by showing the closeness of fit of model performance to human learning curves. In contrast to previous modeling work, we make no a priori assumptions about the underlying cognitive architecture required to duplicate the behavior of strategies. We also compare the performance of our methods to decision trees and demonstrate the superiority of local models for the task of learning compact representations of high-dimensional, noisy, non-stationary sequential data.