Path: utzoo!attcan!uunet!aplcen!jhunix!ins_atge
From: ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards)
Newsgroups: comp.ai
Subject: Re: Artificial vs. ''real'' intelligence
Summary: Oh no, not more AI debate
Message-ID: <5734@jhunix.HCF.JHU.EDU>
Date: 3 Jul 90 18:19:10 GMT
References: <1990Jul2.182411.4441@king.mcs.drexel.edu>
Reply-To: ins_atge@jhunix.UUCP (Thomas G Edwards)
Organization: The Johns Hopkins University - HCF
Lines: 140

In article <1990Jul2.182411.4441@king.mcs.drexel.edu> jsmith@king.mcs.drexel.edu (Justin Smith) writes:
>
>Intelligence.}
>\medskip

>Roger Penrose has suggested that the human brain
>has properties that may enable it to carry out actions
>that are not reproducible by any computer.  This
>argument is used to imply that attempts to simulate the
>reasoning cabability of the human mind mechanically are
>essentially {\it futile}. His argument makes use of human
>consciousness.  I contend that one can come to the same
>conclusion without appealing to human consciousness.
>
>The basic idea is:
>\item {1.} that the human brain is a {\it physical
>object}.
>\item {2.} Physical objects have the potential for
>performing activities that are not reproducible by a
>computer.

>{\it Recursive functions} are  essentially functions that
>can be computed by executing a {\it computer program} of
>some kind.
>
>{\it Physical functions} are functions whose evaluation is
>the result of observing some {\it physical process}. 
>An example of this is the number of ticks on a geiger
>counter per minute as a function of time.

Good enough...but in the real world, one needs a physical
device to run a computer programs.  Thus all computer programs,
when executed on a real computer, become physical functions.
This means that claims based on a computer's inability to perform
physical functions are flawed.


>In fact,
>quantum-mechanical phenomena suggest {\it precisely
>this}.  Quantum mechanics contains many manifestations
>of ``random'' phenomena  --- basically contending that
>certain physical phenomena can only be analyzed {\it
>statistically}.  One can interpret ``random'' as meaning
>``not computable'' rather than ``entirely devoid of
>meaning''.

There are more than one interpretations of the meanings of
quantum-mechanical functions.  Some people want them to be
the fingers of God.  Other people see them as the most
"chaotic" functions, being utterly unpredictable.

Whether QM functions can be computed by a Turing Machine has not
to my knowledge been explicitly proven, and probably will never
be.  

One thing for sure is that QM functions, being physical functions,
can be determined by physical devices.  Computers and brains are
both physical devices.  

>The human brain, being physical, has a {\it natural
>tendancy} to make use of {\it physical functions} rather
>than recursive functions in its computations. 
>Over the course of evolution (and we have to include the
>evolution of the reptilian and mammalian as well as the
>human brain) any physical functions that gave rise to
>useful information {\it were utilized}. 

Let us assume we have a genetic algorithm program running on a
physical computer.  It too will "evolve" utilizing whatever 
computational resources the programmer gives it.  This may include
an external Gieger Counter hooked up to the machine if you
insist on having QM functions neccessary for intelligence.

>The human brain wasn't designed by engineers who have an
>interest in {\it filtering out} physical phenomena that
>cause it to {\it depart} from strict turing-machine
>computations (i.e., the effects of random thermal noise). 
>This is the only reasonable policy to follow in designing
>computers
> --- no engineer (nor anyone else, for that matter) 
>knows enough physics to ``program'' physical phenomena
>{\it fully}.  By this I mean:  if ``random'' atomic
>transitions turn out to really {\it mean something} we
>don't know {\it what} they mean, or how to {\it exploit}
>this ``information'' to solve problems.
   
    How do you reconcile the above statement with the below
     statement?

>The brain, on the other hand, has tens of millions of
>years of ``experience'' at attempting to survive by any
>means at its disposal, and it appears {\it likely} that
>it makes use of physical computations that are {\it not}
>Turing-computable.


    If there is information yielded by QM functions, it can
be determined by learning functions such as genetic algorithms,
symbolic machine learning methods, or neural network
functions such as backpropagation....I can't see how one
can argue there is "hidden information" in QM functions which can
only be interpreted by human evolution and not by any other
learning system.

    Further, there are computer programs which use stochastic
properties to make decisions (i.e. Simulated Annealing).

I see absolutely NO PROOF that QM functions provide any useful
information to an intelligent system which can be utilized.
I see plenty of evidence that QM functions can be used like any
other "random" function to provide probability spectra for
stochastic decisions.  I don't see proof why QM functions
provide any advantage over chaotic functions with similar
probability spectra.

Sorry to be antagonistic, but I don't see why people can't accept the
fact that brain is a physical computing device, as
a digital computer is a physical computing device.
The difference is that the brain relies on parallel non-linear
computational methods on a scale we are 5 or more orders
of magnitude away from, and has complex learning and organizational
of sorts that connectionists are not even dreaming of yet.

And a final note...just because a computer is digital does not
mean it cannot perform parallel analogue equations.  It might
be limited by the "quanta" of it's least significant bit,
but so too are chemical reactions in brain limited by the
"quanta" of chemical molecules, and electric phenomena in brain
limited by "quanta" of a single electron.  Real valued functions
in the real world have the same quantification problems that
real values have on computers (though there are alot more significant
bits in the real world :-).  

On a side note, I just completed training a neural net to
recognize valid targets from IR focal plane arrays.  All I can
say is that the network learned alot more about categorizing
valid targets from invalid targets than I did (I didn't even look
at most of the data).

-Thomas Edwards