[mod.ai] Associative Memory

hamscher@HT.AI.MIT.EDU (Walter Hamscher) (02/27/86)

From: Walter Hamscher <hamscher@MIT-HTVAX.ARPA>


   Date: 24 Feb 86 22:59:07 GMT
   One of the biggest problems AI'ers seem to be having with their machines is
   one of data access. Now, a human [or other sentient life-form :-)] has a
   large pool of experience (commonly refered to as a swamp) that he/she/it has
   access to.
   It is linked together in many obscure ways (as shown by word-association
   games) so that for any given thought (or problem) there are a vast number
   (ususally) of (not-necessarily) connected replies.
   Thinking of that swamp as a form of data-base, does the problem then boil
   down to one of finding a path-key that would let you access all of the
   cross-referances quickly?

It's not invalid but unfortunately it isn't new either.  See any paper
on Frames.  The power of a frame-organized database isn't that there
happen to be these defstructs called frames, it's in the fact that the
frames are all connected together -- it's indexing by relatedness (how
dense the connections have to be before you start to win is an open
question, but see Lenat's recent stuff on CYC in the recent issue of
AI Magazine).  For background see Minsky (A Framework For Representing
Knowledge, 1975).  See NETL (e.g.  Fahlman, Representing Real-world
Knowledge, circa 1979, MIT Press).  See Connection Machine literature
(e.g.  The Connection Machine, Hillis, 1985, MIT press).  If you want
to see the connection between AI KB's and traditional DBMS's covered
extensively, see `Proceedings of the Islamorada Workshop on Large
Scale Knowledge Base and Reasoning Systems' (Feb 85) chaired by
Michael Brodie, available (I think) from Computer Corporation of
America, Cambridge MA (617) 492-8860.