Programming a kinder, gentler conscious HAL

All this Neurdon hullabaloo over memristors and Kurzweilian futurism has got me thinking about the inevitable media question concerning all this: Will our RoboSlave Bots learn to love us in a somewhat creepy, Haley Joel Osment “Artificial Intelligence: AI” kind of way? In other words, will humans be able to one day produce conscious, silicon-based offspring? There are obviously a cornucopia of contingencies when discussing artificial sentience, however, I am going to not-so-subtly sidestep all the philosophical snafus and approach the problem from a modeler’s POV.

My first foray into piecing together a computational perspective on consciousness led me to Bernard Baars’ global workspace theory (GWT). In Baars’ theory, conscious experience is best represented using a theater stage metaphor. Working memory serves as a stage for actors to perform (the contents of conscious experience), where the spotlight is seen as a focal point of attention on the stage. The audience, then, behaves
as a system of unconscious networks which offer their expertise as to what actor should remain in the spotlight. These conscious experiences act serially whereas unconscious actions act in parallel, illustrating the competitive nature of neuronal networks (and biological systems in general for that matter).

http://www.neurdon.com/wp-content/uploads/2009/07/baars-franklin_wm2003diagram-300x216.jpg 300 216

An overview of Baars' Global Workspace Theory (from Baars & Franklin, 2003).

At the heart of Baars’ theory is the idea that conscious problem-priming depends upon unconscious problem-solving which loops back to conscious solution display. What does this mean? Unconscious systems are working hard to return conscious answers based on what you did consciously in the past. In neural network language, this is often filed under the “pattern recognition” category. The theater image, then, is meant to convey direction of information flow in the brain. I find the theater metaphor to be somewhat inadequate for the massively vertical subcortical-cortical connections along with horizontal cortical-cortical activity yet it works as a very basic metaphor. Baars’ theater model of conscious-unconscious interaction with competition and cooperation incorporated into neural networks is in essence the belief that “consciousness creates global access.”

Sure, this is a great psychological account of consciousness, but what about the computational side of things? GWT has undergone several stages of growth since the foundational cognitive groundwork of Bernard Baars two decades ago. Some of these advances include Stan Franklin’s more abstract, module-based software application of Baars’ theory, Murray Shanahan’s robotic implementation with biological analogues, as well as Stanislas Dehaene’s physiologically-based cortical models of cognitive function. Much of the recent work in GWT is based in actual brain area research using more realistic spiking network models, thus I will be referring to this more specific form of GWT as the global neuronal workspace (GNW).

In future posts I will go into more depth with each of these GNW computational models and discuss their contribution towards ushering in the era of HAL. Each of these models raises questions about what a cortical column is actually computing, whether embodiment is a mandate, how does consciousness “emerge” in such a system, and what would the phenomenological properties of a GNW-based simulation or robot entail? To keep you preoccupied until the robot takeover, go watch “Moon” starring Sam Rockwell. Trust me on this one.