From: Crosby_M (CrosbyM@po1.cpi.bls.gov)
Date: Sat Dec 28 1996 - 15:07:05 MST
On Fri, 27 Dec 1996, Nicholas Bostrom wrote:
<BTW, Eugene, I heard about a suggestion made a number of years ago,
that perhaps dreaming is a manifestation of antihebbian learning. The
system is shaked into a random state and then allowed to settle into
an attractor. Then an antihebbian learning rule is applied. When this
process is repeated, it tends to eliminate spurious attractors,
thereby increasing the storage capacity of the network from 0.138N to
about .78N . Do you know whether this beautiful idea has been followed
up?>
Excuse me (I'm not Eugene), and I have no recent knowledge about this,
but a little history might be interesting.
Perhaps you're not referring to anything so ancient as Francis Crick
and Graeme Mitchison's early 80's work on dreaming and neural nets
(e.g., 830714 Nature), but I believe these guys are the originators of
the notions of fantasy, obsession and hallucination in neural nets.
'Fantasies' occur in neural nets when the net "knows too much and
begins to relate everything to everything." 'Obsessions' occur when no
matter what input you provide, the net responds with the same output.
'Hallucinations' can happen in nets with feedback loops where they
respond to inputs that would normally not provoke a response.
As for 'unlearning', Hopfield and colleagues also looked at this way
back when (e.g., same July 1983 issue of Nature mentioned above); so,
there obviously has been much research since then, especially in
neurophysiology circles (J. Allan Hobson, in particular, in regards to
unlearning and dreams).
Mark Crosby
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 14:35:56 MST