Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!europa.chnt.gtegsc.com!news.sprintlink.net!simtel!harbinger.cc.monash.edu.au!yarrina.connect.com.au!labtam!labtam!chris
From: chris@labtam.labtam.oz.au (Chris Taylor)
Subject: Re: Can neural nets think
Message-ID: <chris.804405697@labtam>
Organization: Labtam Australia Pty. Ltd., Melbourne, Australia
References: <3sem09$o0v@Owl.nstn.ca>    <182254011wnr@tropheus.demon.co.uk> <0jvfkPK00YUr85Omot@andrew.cmu.edu> <886840599wnr@tropheus.demon.co.uk> <Tarmo.Kaldma-2706951257100001@macbeth.hh.se> <3spcuu$cot@popeye.jsc.nasa.gov>
Date: Thu, 29 Jun 1995 06:01:37 GMT
Lines: 45

Dave Forrest <forrest@rsoc.rockwell.com> writes:

>Tarmo.Kaldma@cca.hh.se (Tarmo Kaldma) wrote:
>>
>> Natural (biological) brains are programmed.  Training is used only for
>> tuning. There is no reasons why the same approach cannot be used for
>> artifical ones.
>> 

>This is an interesting point.  I would agree that we are born with
>"programmed" synapses, etc.  But wouldn't you say that this 
>"programming" is actually a result of the "tuning" that has occurred 
>over some 2000 generations of human evolution?   Taking that one step
>further, it seems that it would suggest a training approach for a 
>"Brain-sized" NN.  i.e. you start with a "primitive brain" that you
>can train and that is relatively simple-minded.  You then "evolve" 
>the brain for several generations, giving it a "life-time worth of
>training" in each generation.


I imagine the strategy would be to have a sequence of learning-stages
that you ramp up gradually as the evolving ANN gets sufficiently competent.

A big problem is laying out a sequence of suitable learning stages.
It should introduce relevent features of human learning ability
and 'instinct' at a rate that the evolution can cope with.

Instinct primarily being an evolved model of the environment to
which the ANN can refer to. Without this the ANN would tend to just
evolve learning skills.

I think it is probably inefficient to evolve an ANN that 'just thinks'
with the expectation that it can learn an arbitrary environment.
Evolving a referance model of the target environment should take pressure
off learning skills.


The computation time would be massive.
Emulating all those generations of evolution on all those individuals with
all those neurons.
You would probably prefer a programmable hardware framework for the 'brain'
to make testing of individuals quicker. Developing that would be a major
effort itself and would probably need to be evolved gradually too.
And there is no guarentee that our current ANN models are suitable
starting points.
