Anders -- I'm not sure if you have caught up with my posts on the sleep thread yet, esp. vis-a-vis possible architectures & processing procedures. But given your comments on not being able to train and run neural networks at the same time, I thought I'd offer a comment and pose some questions.
There would seem to be several requirements here:
Do you see any problem with sleeping & dreaming being the primary
processes that allow (c) and (d) to function in the process of
transfering data from (a) to (b)?
(a) Processing of sensory input and presumably storing the
relevant stuff it in a short term pattern buffer (because
there simply isn't enough time to activate gene expression,
biochemical pathways, transport of receptor molecules and
neurotransmitters, etc. -- I.e. processes that take minutes
[In computer architectures, this is a "short-term" cached input-buffer.]
(b) Some application of that short term "eletrical" pattern into
a chemico-physico structure (i.e. long term memory) [In CA, this is writing the data to disk or tape].
(c) Some logical resolution of conflicts in the data. If someone
tells you garbage and you know it is garbage you immediately reject it. But there may be data or complex experiences that can't be immediately accepted or rejected (but you don't have the processing capacity to do that immediately). Examples that come to mind (of the immediate accept/reject type) would be noting spelling when reading or mis-spoken words when listening. Examples (of the delayed accept/reject) could include complex scientific or philosophical arguments. Now, some of the logical accept/reject may be conscious, but some of it may be unconscious as well. [In CA, an example of this that comes to mind is the 4 computers on the Space Shuttle that use majority & confirmation "logic" -- 2 of 3 identical computers must agree and 1 different computer confirms the result.]
(d) Perhaps rehersal of conclusions to strenghen the results & connections.
Do you see any problem with sleeping & dreaming being the primary processes that allow (c) and (d) to function in the process of transfering data from (a) to (b)?
Unfortunately as we are "initialized", we accept data in (c) all too unconsciously with few "reality checks". This is probably required for rapid learning in babies. It also contributes to our difficulties in "unwiring" the early stuff like religious beliefs.
It is interesting to consider that the sleep requirements decline significantly in older healthy adults. Now there could be many reasons for this but one would expect that they have well honed survival skills and have sufficiently developed their accept/reject pathways that most of the data gets processed very quickly. One possible way of testing this would be to see if sleep requirements increase in individuals who have traveled to a "strange" place. It might be difficult to account however for decreases in sleep that might be caused by a higher arousal/alertness state due to potential dangers perceived from the environmental differences.