Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!seismo!sundc!pitstop!sun!amdcad!ames!elroy!cit-vax!ucla-cs!berke
From: berke@CS.UCLA.EDU
Newsgroups: comp.ai.neural-nets,comp.ai
Subject: "2**n   events  using  only n units" references? (from Berke)
Message-ID: <8958@shemp.UCLA.EDU>
Date: Mon, 2-Nov-87 07:26:13 EST
Article-I.D.: shemp.8958
Posted: Mon Nov  2 07:26:13 1987
Date-Received: Fri, 6-Nov-87 22:16:26 EST
Sender: usenet@CS.UCLA.EDU
Reply-To: berke@CS.UCLA.EDU (Peter Berke)
Organization: UCLA Computer Science Department
Lines: 31
Summary: request for references.
Xref: mnetor comp.ai.neural-nets:34 comp.ai:1048


        Many connectionist researchers  have   asserted  that   a
distributed  representation provides efficient use of  resources,
encoding   2**n   patterns   in   n units.  The "2**n states  for
n units" argument is sketched below:

Replace  unit-encoding  (grandmother  cells)  with  patterns   of
activation  over n (binary) units.   Instead of representing only
n distinct "events," one with each unit, we can represent  up  to
2**n   events  using  only n units.   These patterns overlap, and
this overlap can be used to gain "associative" recall.


        Does anyone have any references to such arguments?   I've
heard  this  argument  made  verbally,  but  I don't recall exact
references in print.  Do you?  Also, is  there  a  net-convention
for  2  to-the-n?   I'm  using 2**n above, (a vestige of my early
FORTRAN experience?) which I prefer  to  2^n.   Anyone  have  any
others?

        Perhaps it would be appropriate to  "r"  a  reply  to  me
rather  than  posting  a  follow-up  to net.  If they are many or
interesting, I'll be sure to post them in one batch.

        I  would  appreciate  exact   quotes,   with   references
including  page  numbers  so  that  I  could find the, as the NLP
people say, context.

Thanks

Pete