🚀 go-pugleaf

In article <875704714snz@longley.demon.co.uk>,
David@longley.demon.co.uk (David Longley) wrote:

>[...]
>What  I  want to draw attention to is the fact that even  if  the 
>work in ANN research proved successful, ie that such systems were 
>engineered,  the results would be what we would generally  regard 
>as superstitious behaviour. 

  Of course, since learning is an inductive process.  But induction
does lead to the discovery of deduction and there is no reason to
think that machines cannot be trained to achieve a high degree of
scientific rigor beyond that which is possible in humans.  I say this
because machines will not initially be encumbered by the genetic
programming and social baggage that humans have to carry.  Besides,
uncommon discipline is not needed for our coming learning machines to
be immensely useful.  Most tasks for which machines will be used do
not require any skills beyond that which is obtainable through simple
conditioning.  Examples are driving a vehicle, gardening, farming,
patient care, factory work, etc.  My point goes further and I will
argue that if our future machines can be trained to accomplish these
tasks, there is no reason to suppose that they can also be trained to
do your job and mine more effectively and efficiently.  I believe this
will happen in our lifetime.

>To produce intelligent behaviour one needs to inculcate extensive 
>skills which are situationally apposite, and I think those skills 
>will  turn  out to be procedural but embedded  -  which  probably 
>means, not as intelligent as we imagine <g>. That's what I  meant 
>when  I  said that AI just *is* science - perhaps I  should  have 
>said, our scientific knowledge base. 

  Your definition of intelligence is somewhat different than that of
most people, I believe.  This is fine.  The rigorous scientific
intelligence you are looking for is needed by or is useful to only a
small minority of the population.  Even then, I would argue that if
humans can be trained to achieve a high level of discipline in certain
domains, conditionable machines will achieve an even higher level.

>I  once  thought that the CYC project had grasped that -  but  it 
>hasn't  -  it's in pursuit of the original  McCarthy-Hayes  -goal 
>which  I  think is misguided to the extent that 1)  it  seeks  to 
>include folk notions and 2) is practically undeliverable even  if 
>it didn't. 

  I know I'll get flamed for this by the old guard, but the CYC
project is a waste of time, IMO.  Some people simply underestimate the
complexity of highly interconnected knowledge.  This type of knowledge
cannot be hand-fed to a machine, it can only be learned inductively.
And to do that one must first have a good general theory of learning.

>What these folk are getting close to realising is that there is a 
>need for good scientific literature databases, improved  versions 
>of INDEX MEDICUS, PSYCHLIT etc. 

  Certainly but this has very little to do with intelligence.
Creating and maintaining such databases is a tedious and boring
enterprise, one which is best left in the hands of trained or
conditioned learning machines.  We humans should just have fun and sip
margaritas on the beach while being given a massage by one of our
mechanical masseurs.  :-)

>I wish you luck in your R&D though.

  Thanks.  I think you have a somewhat pessimistic view of the power
of inductive learning.  Sure, it is a trial and error processes that
can easily give rise to superstitions, as you pointed out, but this
trial and error process should not be underestimated.  It can also be
used to send men to the moon and invent intelligent, conditionable
robots.

Regards,

Louis Savain,

President, Marengo Media, Inc.