home · mobile · calendar · defenses · 1997-1998 · 

Thesis Defense - Mathis

A Computational Theory of Consciousness in Cognition
Donald Mathis
Computer Science PhD Candidate

Issues of awareness and consciousness are becoming increasingly important in studies of psychological phenomena such as perceptual priming, implicit memory, automatization, and the disorders resulting from brain damage, yet the role of consciousness in these phenomena remains unclear. This thesis develops a computational theory of the processes underlying consciousness, which accounts for these data under a single unifying framework. Simulation models of five experiments are described showing that the theory is consistent with data across several domains.

Motivated by findings in cognitive psychology, neuropsychology and neuroscience, the theory consists of three basic claims.

  1. The human cognitive architecture consists of a set of functionally specialized modules, each of which functions as an associative memory in its domain. The modules are extensively interconnected, and each computes an output that represents the most appropriate interpretation of its inputs that is consistent with a set of domain constraints. Any cognitive task is achieved through the coordinated operation of many modules.

  2. Each module computes its output in a two-stage process consisting of a fast, input-output mapping step followed by a slower, incremental relaxation search step. The mapping step produces and output quickly, but this output may not satisfy domain constraints. The relaxation step then gradually transforms the initial output state into one which maximally satisfies the domain constraints, ending in a stable state. This operation results in a speed/accuracy tradeoff in each module.

  3. Stable states in the outputs of modules correspond to the contents of consciousness. Unlike most theories, consciousness in our theory is fully distributed, because stable states in any module may reach consciousness, without requiring the involvement of any special brain area or special consciousness system. The characteristics of consciousness in cognition are accounted for by the properties of stable states in this system of modules.

A connectionist implementation of the theory is used to model data from five cognitive psychology experiments across several domains. Experiments by Greenwald and colleagues (Greenwald and Draine, 1995; Greenwald, Draine, & Abrams, 1996) demonstrate basic conscious and unconscious perceptual priming effects, and find that speeded decisions increase priming effects. An experiment by Marcel (1980) shows a correlation between consciousness and selection in the processing of semantically ambiguous words. Debner and Jacoby (1994) describe experiments suggesting that conscious and unconscious processes operate independently, for which our theory accounts without making such an assumption. Fehrer and Raab's (1962) experiment shows a dissociation between awareness and responses in metacontrast. And an experiment by Levy and Pashler (1996) suggests a way in which conscious decisions may affect ongoing processes in higher cognition. Our theory offers accounts for all of these phenomena within a single, unified framework.

The question of the computational utility of consciousness in cognition is then addressed, and it is shown through simulations that the mechanisms proposed to underlie consciousness serve an important computational function within the system. The theory suggests that, rather than existing simply to implement consciousness, these mechanisms boost cognitive performance by allowing the system to generalize better to new problems within a domain, and to stay "on track" when only degraded or incomplete information is available.

Committee: Michael Mozer, Associate Professor (Chair)
Clayton Lewis, Professor
Satinder Singh, Assistant Professor
Paul Smolensky, 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)