A.s.a. H. may have inadequate means for feeding back information/activation from upper layers in the hierarchy back down to lower layers. A.s.a. has operated mostly "feed forward"/"bottom up."
This has not been a problem with output of actions. Typically there are orders of magnitude more sensory input as compared to actuator output anyway and reflex actions are contained in the single lowest layer and can be learned there. In a good many experiments we have also concentrated on A.s.a. observing the world and modeling it rather than acting.
A.s.a. is not completely without downward activation. We have, for example, fed back utility signals to lower level concepts/cases reflecting their use and value when employed in higher layers. We have also applied feature extraction algorithms to a layer and then added any newly discovered features as cases/concepts/patterns in the next lower layer. This can be done efficiently with parallel processing.
No comments:
Post a Comment