home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!paladin.american.edu!auvm!FAC.ANU.EDU.AU!ANDALING
- Message-ID: <9211200034.AA26550@fac.anu.edu.au>
- Newsgroups: bit.listserv.csg-l
- Date: Fri, 20 Nov 1992 11:34:36 EST
- Sender: "Control Systems Group Network (CSGnet)" <CSG-L@UIUCVMD.BITNET>
- From: Avery Andrews <andaling@FAC.ANU.EDU.AU>
- Subject: association & g-ma neurons
- Lines: 58
-
- [Avery Andrews 921120.1145]
-
- This may be pretty ignorant, but here goes anyway. It seems to me that
- a characteristic feature of classical `connectionist' systems is
- precisely that they can in a sense recognize configurations without
- having a `grandmother neuron' for each config. that is recognized.
- E.g. you can have a bunch of input & output wires, with an arbitrary
- association function connecting input to output patterns:
-
- ---
- --->| |--->
- --->| |--->
- --->| |--->
- ---
-
- For example on the input lines might be visual/auditory/olfactory
- properties, on the output lines kinaesthetic reference levels, such that
- for input combinations typically associated with the presence of
- leopards, the output levels are those associated with running away,
- etc.
-
- This sort or thing should also be possible for connecting verbal &
- nonverbal perceptions. To account for actual learning of words, it
- would have to be `programmable' very quickly (on the basis of a single
- experience), but I think there might be a preadaptation to this sort of
- things.
-
- Imagine a very simple critter whose sense inputs are like e-coli's
- (e.g. a vector), but which also has a `pain' perceptual system that
- registers damage. We imagine the sense vectors and the pain wire
- going into a black box whose output will be labelled `fear':
-
- |
- | pain
- v
- s -----
- e ----> | | fear
- n ----> | |--------->
- s ----> | |
- e -----
-
- The idea is that whenever the pain wire fires, something happens inside
- the black box such that, subsequently, the closer the sense-vector gets
- to what it was when the pain wire fired, the larger a signal comes out
- along the fear wire. This signal then is then handled in such a way
- is to typically induce behavior that will alter the sense-vector. (If
- the organisms motor system is e-coli style, the fear signal would
- contribute additively to the tumbling rate).
-
- Stick a lot of these things together in the right way, & you might be
- able to get a fast associative learner with no grandmother neurons.
-
- & of course, for the optimists, a similar picture could be drawn with
- wires labelled `pleasure', etc., where the output wire tends to do
- things to cause the sense-vector to get closer to what it was when the
- pleasure wire fired (e.g. reduce tumble-rate).
-
- Avery.Andrews@anu.edu.au
-