Not all of one's mental contents have been exhausted when you have identified all of the concepts an individual possesses. (See, for instance, Essays on Nonconceptual Content, Y. H. Gunther, MIT Press, 2003) Experiments with Asa H help us explore both the conceptual and nonconceptual contents of mind (both human minds and artificial ones).
Reflexive reactions to bright light, striking of the shin, etc. are nonconceptual in humans and a robot may have these as well. Preprocessors may adjust visual (and other) sensitivities (sensor "autoscaling") for example.
In Asa H concepts are formed from input through vector clustering. A concept is a cluster.
Clusters which have not been identified (associated) with a human word/label may be concepts for Asa H but may not correspond to any concept humans use to describe the world.
Any new input vector which isn't sufficiently similar to an existing concept will be stored in memory and may later be incorporated into the formation of a new cluster. Prior to that time it might be considered to be a "protoconcept."
More abstract concepts reside higher up in the Asa H hierarchical memory network. More abstract concepts require a deeper hierarchical network. Shallow networks might be able to learn the same content (input-output mapping function) that deeper networks learn but they would not be able to contain the same concepts.
No comments:
Post a Comment