Re: More about A comment

[ Follow Ups ] [ Post Followup ] [ Salon ] [ FAQ ]

Posted by Yon Bard on January 25, 1999 at 16:56:59:

In Reply to: Re: A comment posted by Nikos Salingaros on January 25, 1999 at 11:25:02:

: The answer to Tom Cole's inquiry is contained in the
: material I supplied initially, which was rejected
: as too far out.

: Let us consider a computer that judges the quality
: of a carpet. We have to program certain rules,
: like the checklist given in this salon; perhaps
: made to include several times the number of criteria,
: with the present checklist representing the rough
: initial judgement, and sublists that establish the
: fine degree of judgements. If we do our job correctly
: then we may come up with a list of lists, each acting
: on a different level of fine judgement, and sublists
: that expand the fine points of a single criterion.
: Such a computer will, I think, provide a relatively
: correct judgment, in the sense of accurately
: finding the "dancing" quality, or "life" in a good
: carpet design. If it doesn't, we can certainly
: refine the program by adding further and finer
: criteria until its performance is satisfactory.

: Those among the readers who know some computer
: science will immediately recognize the definition
: of a neural network. An artificial neural
: network can be trained to recognize some desired
: qualities through the use of programmed rules.

: My claim is that the human mind acts in precisely
: the same way. A knowledgeable carpet dealer trains
: his or her mind over years, building up a list of
: rules exactly as described above, which are
: programmed in the human neural system of learning.
: The response of the computer (mind) is
: instantaneous. It is still, however, dependent
: on some set of rules, and that's what we are
: discussing now.

: Nikos Salingaros

I would like to comment about your statement that if we get enough rules our job is done. Not necessarily. In statistics we can fit a number of observations by a specific type of curve if it has enough free parameters (e.g., n observed values of y vs. x can usually be fitted precisely with a polynomial of degree n-1) even if the fitted curve has no relation to the true relation between the variables (e.g., a sinusoidal wave or an exponential growth. The simplest case is drawing a straight line through two points-you can always do it, even if the true relationship is not at all linear). So, the fitted curve emulates the given observations perfectly, but has no predictive value whatsoever. My point is that just because a 'sufficient number of rules' appears to work is no proof of their validity or relevance.

Regards, Yon

Follow Ups:

Post a Followup




Optional Link URL:
Link Title:
Optional Image URL:

[ Follow Ups ] [ Post Followup ] [ Salon ] [ FAQ ]