home · mobile · calendar · defenses · 2002-2003 · 

Thesis Defense - Colagrosso

A Rational Theory of Skilled Performance and Practice: Modeling Long-Term Repetition Priming
Michael Colagrosso
Computer Science PhD Candidate

I present a theory to address the time course of neocortical information processing and changes in information transmission dynamics through learning. From the general theory, I distill a model of perceptual identification and the facilitation due to repetition priming, a simple, yet pervasive, form of learning where past experience with a stimulus or response causes more efficient processing on subsequent occurrences. A central conjecture to my investigation is that psychological and neuropsychological phenomena can be explained from a "rational perspective" (Anderson, 1990), which views cognition as being optimized with respect to task-related goals and the statistical structure of the environment. Together with the assumption of rationality, the theory uses dynamic belief networks to model the processing from stimuli and responses. Mental states are represented as probability distributions, and microinference is performed via Bayesian belief revision. The belief networks provide a principled mathematical theory for conceptualizing information processing, and its temporal structure captures the time course of information transmission.

I present computer simulations from thirteen experiments addressing key phenomena within repetition priming; the data and their interpretation consider the effect of subliminal and supraliminal long-term repetition priming under varying target stimulus presentation durations, delays between prime and target stimuli, experience with the stimuli, all over a range of report modalities (e.g., naming, forced choice, lexical decision). Repetition priming is fundamentally related to skilled practice, i.e., repeated performance on a task that is already somewhat familiar. The model accurately accounts for human performance provides an elegant interpretation of the operative learning mechanisms. My probabilistic approach decomposes priming into two distinct mechanisms: one adapts the prior distributions and one adapts the strength of association between inputs and outputs. These two mechanisms roughly correspond to the bias and sensitivity effects from the experimental priming literature. The claim of my model is that both types of adaptation occur in parallel, although one may be more prevalent for a particular task.

My Bayesian theory is elegant, flexible, and explicit, resulting in a far more parsimonious account of a range of human behavioral data than other proposed models.

Committee: Michael Mozer, Associate Professor (Chair)
Gregory Grudic, Assistant Professor
Clayton Lewis, Professor
James Martin, Associate Professor
Randall O'Reilly, Department of Psychology
Department of Computer Science
University of Colorado Boulder
Boulder, CO 80309-0430 USA
May 5, 2012 (14:20)