Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!uflorida!haven!aplcen!jhunix!ins_atge
From: ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards)
Newsgroups: comp.ai
Subject: Re: the surrealism of dreams
Summary: PDP Dreams
Keywords: Boltzman Machines Dreams PDP
Message-ID: <1302@jhunix.HCF.JHU.EDU>
Date: 31 Mar 89 22:27:51 GMT
References: <74@opmvax.kpo.fi> <5698@cognos.UUCP>
Reply-To: ins_atge@jhunix.UUCP (Thomas G Edwards)
Organization: The Johns Hopkins University - HCF
Lines: 36

In article <5698@cognos.UUCP> rayt@cognos.UUCP (R.) writes:
>In article <74@opmvax.kpo.fi> Andy Ylikoski writes:
>>I would like to try to contribute to the discussion involving dreams
>>"not following the usual laws of nature".

A while back I mentioned that there are some analogies between dreaming
and Boltzman Machine learning in PDP.  I am sure that the author's
do not mean to claim that dreaming is Boltzman Machine learning
unclamped, but just that there are interesting similarities.

Hinton and Sejnowski point out that Crick and Mitchison (1983)
"have suggested that a form of reverse learning might occur
during REM sleep in mamals.  Their proposal was based on the
assumption that parasitic modes develop in large networks that
hinder the distributed storage and retrieval of information." They
quote Crick and Mitchision: "More or less stimulation of the
forebrain by the brain stem that will tend to stimulate the
inappropriate modes of brain activity ... and especially those which
are too prone to be set off by random noise rather than by specific
signals."
  Boltzman Machine learning uses a similar concept.  Learning is
split into two phases, a phase+ , where positive Hebbian learning
occurs with input and output units clamped to their proper values,
and a phase-, where negative Hebbian learning occurs to the 
unclamped network (thus randomly stimulating those parasitic modes
described above, and the associated nodes get de-strengthened by the
negative Hebbian learning which usually strengthens nodes which are
simultaneously stimulated, reducing the combined importance of those
"erroneous" modes).

Crick, F. and Mitchison, G. (1983).  The function of dream sleep.
  Nature, 304, 111-114.
Rumelhart, McClelland eds (1987).  Parallel Distributed Processing.
  MIT Press, 282-317.

-Thomas Edwards