Consciousness and metarepresentation: a computational sketch.
Neural networks : the official journal of the International Neural Network Society November 1, 2007 Axel Cleeremans, Bert Timmermans, Antoine Pasquali
Consciousness involves not only being aware of something but also being aware that one is aware. This paper explores computational mechanisms that could implement this idea, starting from the observation that connectionist networks acquire knowledge that remains embedded in their task performance—knowledge in the network rather than knowledge for the network. The authors present simulations where a second-order network observes the internal states of a first-order network trained on a categorization task. The second-order network either reproduces those states or uses them as cues for a secondary task, achieving a limited form of metarepresentation. This mechanism offers a beginning for understanding mental attitudes and suggests that consciousness involves learned knowledge of one's own internal representations through interaction with self, world, and others.