home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!dtix!darwin.sura.net!mips!swrinde!cs.utexas.edu!rutgers!igor.rutgers.edu!planchet.rutgers.edu!nanotech
- From: hsr4@vax.ox.ac.uk (Auld Sprurklie)
- Newsgroups: sci.nanotech
- Subject: Re: Multiple Experiences
- Message-ID: <Aug.21.18.27.43.1992.20255@planchet.rutgers.edu>
- Date: 21 Aug 92 22:27:44 GMT
- Sender: nanotech@planchet.rutgers.edu
- Lines: 88
- Approved: nanotech@aramis.rutgers.edu
-
- In article <Aug.19.20.18.58.1992.28197@planchet.rutgers.edu>, sinster@cse.ucsc.edu (Darren Senn) writes:
- > In article <Aug.15.10.19.21.1992.10779@planchet.rutgers.edu> hsr4@vax.ox.ac.uk (Auld Sprurklie) writes:
- ""From what I'd read so far, it appeared to me that the memories -are-
- ""associated with the means used to store them (don't different transfer
- ""functions apply to storage of different types of memory, some transfer
- ""functions being unsuitable for learning in certain cases ?), [...]
- "
- " The transfer function is generally fixed when the network is first
- " created: it's not learned. So the transfer function is really part of
- " the net's structure, not the memories. Certainly, the exact values of
- " the weights and thresholds depend on the network's transfer functions,
- " but since only the weights and thresholds change, it's somewhat
- " meaningless to associate the memories (a purely learned phenomenon)
- " with the transfer functions (a static phenomenon).
-
- I know that at present the transfer function is fixed at creation, but I'll
- have to go back and re-read my references; my understanding was that different
- transfer functions were suited to different applications - else why have
- different transfer functions; surely one would suffice ? If the function
- introduces non-linearities into the network in order to facilitate the learning
- of information (the creation of memories), then I would assume that while the
- function is associated with the network structure it is also related, however
- obscurely, to the type(s) of memory being created.
-
- Or am I way off beam ?
-
- [bits deleted]
-
- " So in order to copy a whole skill, you must discover all the memories
- " that compose that skill and its subskills, and teach those memories to
- " the recipient network. Once you have the memory, teaching it is simple.
- " Finding the memories that compose the skill is a different matter
- " entirely. :(
-
- Yes, that was one of my points earlier - how do you determine which memories
- or skills lie where, and how do you obtain them if everything is inter-
- dependent ? In addition, if you need to update an existing skill (perhaps one
- which was incompletely learned at an earlier age) how best to install that new
- version of the skill so that it fits seamlessly ?
-
- " [...]
- ""[...] The obvious
- ""question which follows on is: how do you decide which dimple or group
- ""of dimples corresponds to a particular skill/memory ? And if we still
- ""take the parallel dump from clones to primary as a possibility, would
- ""it be possible to differentiate the new skills/memories from the set
- ""supplied to form the clone(s) in the first place ?
- "
- " Detecting which dimple/group corresponds to a particular memory is simple.
- " Assert the memory onto the network (that means provide the network's
- " input _and_ output) and calculate the optimal set of weights for that
- " memory (there's a generalized formula for this). Euclidean distance
- [...]
-
- I think I wondered in an earlier post about the necessity for presenting an
- original input to a network in order to elicit its whereabouts (although that
- was for the purpose of 'forgetting' it in order to replace it).
-
- [...]
-
- " The danger, when teaching a new memory to a network, is that old memories
- " might be forgotten, or just become harder to reach. Imaging a stretched
- " sheet of silk with a number of marbles glued to certain positions. These
- " marbles will deform the silk. If you drop a glass bead, somewhere, it'll
- " roll around until it runs into a marble. This marble is the memory
- " associated with the input corresponding to the bead's original placement.
- " If you put a bowling ball on this sheet of silk, then a large number of
- " marbles will never be hit by the bead. The bowling ball represents a
- " memory that is remarkably easy for a particular network to learn, or
- " a memory that has a large input space (imagine teaching a network to
- " recognize individual faces -- the marbles -- and then training that same
- " net to recognize Caucasians -- the bowling ball).
-
- By large input space, I assume you mean having been presented much more often
- than other inputs (such as the ability to recognise faces the right way up,
- but not when oriented upside down - since they are not often encountered in
- that orientation).
-
- This may be a blind alley, but I seem to recall that the older one becomes,
- the easier it appears to be to recall earlier memories - as if they are
- coming to the surface (poor analogy, I know). Perhaps the ability to be
- cloned successfully will be age-dependent; skills learned long ago might be
- easier to locate in an older invididual than in a younger one.
-
- (This assumes that the memories of yore which suddenly seem to resurface are
- in fact being correctly recollected!).
-
- Peter "I can't recall just now - ask me in twenty years time" Brooks
-