From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!swrinde!cs.utexas.edu!news Sun Dec  1 13:05:39 EST 1991
Article 1652 of comp.ai.philosophy:
Xref: newshub.ccs.yorku.ca sci.philosophy.tech:1163 comp.ai.philosophy:1652
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!swrinde!cs.utexas.edu!news
>From: turpin@cs.utexas.edu (Russell Turpin)
Newsgroups: sci.philosophy.tech,comp.ai.philosophy
Subject: Re: Zeleny's argument lives!
Summary: Dr Frankenstein!  It moves!
Message-ID: <kj5t8lINNrv@cs.utexas.edu>
Date: 27 Nov 91 01:29:57 GMT
References: <1991Nov24.124945.5834@husc3.harvard.edu>
Followup-To: sci.philosophy.tech
Organization: U Texas Dept of Computer Sciences, Austin TX
Lines: 47

-----
MZ:
>>> [Assume we can build a machine that exhibits our behavior and competence].
>>> However, at any given time, by assessing its construction, we may comprehend
>>> all causal factors that influence its behavior (to the extent that this is a
>>> machine constructed by ourselves, I assume that we can do so, retracing, if
>>> necessary, the modifications imposed on the initial configuration by the
>>> learning process).

MW:
>> You assume wrongly. Even now, the behavior of only moderately complex
>> ANNs is often impossible to understand by looking only at the modified
>> weights (or even the history of those weight changes). I strongly suspect
>> that the more sophisticated and complex ANNs to come will be even harder
>> to analyse by looking at the encodings. Your argument dies right here.

In article <1991Nov26.113244.5920@husc3.harvard.edu> zeleny@zariski.harvard.edu (Mikhail Zeleny) writes:
> Are they more powerful then Turing Machines?  If so, explain how, and
> automatically refute Church's Thesis.  If not, you lose.

This is simplistic.  What does it mean to "comprehend all causal
factors that influence [a machine's] behavior"?  Comprehension
comes in a variety of flavors, and it is far from clear what
precisely is needed for Mr Zeleny's argument to carry through. 

Suppose what is needed is the ability to predict an arbitrary
machine's response to an arbitrary input, ie, that what is needed
is an *effective* method to determine, for *any* Turing machine,
whether OR NOT an arbitrary string is accepted by the machine in
question.  This, of course, is equivalent to the halting problem,
and is undecidable. 

Even for classes of machines much weaker than Turing machines,
there are many unsolvable questions.  For example, there is no
effective method to determine whether one CFL is contained in
another.  

I am sure that Mr Zeleny knows these unsolvability results.  What
he has not made clear, however, is exactly what kind of
"comprehension" is required for his argument.  (At least, I am not
able to determine this.)  It is very easy to slip into the realm
of the unsolvable, and to assume that someting has an effective
decision procedure when it doesn't.  (Many working computer
scientests can attest to this by their own experience!)  Has Mr
Zeleny done this without realizing it?  

Russell


