Re: Aaron Sloman's "The Irrelevance of Turing Machines to AI" article

From: David Longley (David_at_longley.demon.co.uk)
Date: 08/04/04 

Date: Wed, 4 Aug 2004 16:15:16 +0100

In article <fa69ae35.0408040449.4c054480@posting.google.com>, Eray
Ozkural exa <erayo@bilkent.edu.tr> writes
>Hello Wolf,
>
>Thanks for the interesting examples, but...
>
>Wolf Kirchmeir <wwolfkir@sympatico.ca> wrote in message
>news:<Y%rOc.6647$pc.408371@news20.bellglobal.com>...
>> If by "transfer of learning" you mean what happens when a
>> teacher-student relationship works, well, the only clue I have is that
>> the student learns what (s)he does. For example, if a student memorises
>> and recalls definitions of terms, then the student can memorise and
>> recall definitions of terms. But that does not mean that the student can
>> understand a text using those terms. Most of them can't even rephrase a
>> definiton "in your own words."
>
>If the exam questions are so much harder than what they practiced for,
>they would have to fill in the gap with thinking. That's not always
>possible due to finite speed of processing.

This is hopeless. You ignore the empirical research even when it's
presented and summarised for you.
<http://groups.google.com/groups?selm=GciKo7KGWmCBFwOc@longley.demon.co.uk>.
What we see you doing throughout this post, is exactly what you've been
doing in your other posts - you just make up some a priori clap-trap
that fits your folk psychological prejudices regardless of the empirical
evidence.

Pointing out that you don't see how silly such behaviour is invariably
tends to be a waste of time, as *if* you could see it for what it is
(something which would require you to be held accountable for your
behaviour in the way that Wolf says he is), I reckon you'd be far too
embarrassed to write the way that you currently do about these matters.
The difference between you and Wolf is experience, ie feedback on (the
outcome of) your own behaviour.

This general failure of yours is no doubt one of the reasons why you
react the way that you do to criticism. You can't see what's wrong with
the way you write about these matters because you lack the
experience/formal education to know otherwise - one can see that by the
extremely silly things you say about "Skinner's rather simplistic
learning theories".

Look at how you write about these matters - you say that when practice
doesn't suffice, the disparity or difference between the training and
"much harder" test materials must be made up for by "thinking"! Think
about the choice of words above, what does it remind you of? What is
this "thinking" which is so limited by the "finite speed of processing"?
How do you measure it and how do you refer to it except through
behaviour and extensional analysis? What differentiates a good from a
better "thinker"? What might be going on where one thinks one finds
someone who generalises (transfers) well? Why might this not be a sound
way of talking about what's really going on?

None of your intensional talk does anything more than obscure the
important questions, it just translates one folk psychological
vernacular into another and in the case of cognitive neuroscience
science and computational neuroscience this has become a black art.

>
>However, transfer of learning is not always that challenging as in an
>exam. For instance, I can recognize things if they are garbled (like a
>text), or adapt previously learnt strategies to completely novel
>situations. I can get creative. If your students can't, well, I'm
>sorry for them :)

You can't generalise either. You just can't see it.

>
>The quest to make our *machines* as flexible requires a good deal of
>theory. It needs, for instance, an incremental theory of learning that
>can recollect past models and refine them in creative ways. Although
>some of the statistical classification algorithms can do this, I
>suspect it gets much more complicated when the data isn't presented in
>the tabular attribute-value format. That's one of the problems with
>traditional neural networks approaches as Minsky has pointed out.
>Thinking very basic, you first have to do some sort of segmentation,
>centering, filtering, etc., on the image, before you can feed it to a
>statistical classifier. It is not obvious at all that this
>preprocessing or even the problem formulation itself is trivial.
>
>[We can quickly formulate classification problems, why do we do that?
>What are the criteria? What is the method?]
>
>That was a minor example from machine learning. Humans accomplish
>transfer of learning every day, every hour. Every human performance is
>adaptive, and several performances require creativity of some sort.
>Simply performing whatever knowledge has been inscribed, without
>modification, is insufficient. The world is too complex to enumerate
>like that. (as in our rather dumb chess programs) Instead, dynamic
>models will be required.
>
>When a human learns how to use a phone, he can, with a little effort,
>use every kind of phone, cell phones, car phones, little phones, big
>phones, ancient phones. I think Skinner's rather simplistic learning
>theories would have a hard time explaining all this adaptivity, and
>that is precisely why we need much more detailed learning theories
>covering reinforcement learning, supervised learning and unsupervised
>learning. (In your post, you seem to be telling me that transfer of
>learning does not exist in practice, maybe because it contradicts with
>your behaviorist leanings? What could be more obvious than adaptivity
>of humans?)
>
>Needless to say, learning also occurs virtually. We can learn in
>thought experiments we call "imagination", a good example of this
>conscious mode of learning is doing mathematics in which the agent can
>discover new theorems by himself and learn useful mental strategies
>for arriving at abstract results. (For instance, he can spot a belief
>about a proposition which turned out to be wrong when he put it to
>test. That is certainly learning)
>
>The transfer of learning problem shows us that the three-way
>separation of learning above is insufficient. When I have to use a
>new, strange phone, I do unsupervised learning, supervised learning
>and reinforcement learning all at once, and perhaps, more. [If I can't
>figure it at once, I might have to think about it, or ask somebody
>else. etc.] We need better theories of learning that can explain the
>flexibility, the control mechanisms (what is needed to drive these
>complex actions), etc.
>
>Regards,
>
>--
>Eray Ozkural
The above way of talking is hopeless. 

-- 
David Longley

Relevant Pages

  • Re: What Equipment Does a Beginner Start With
    ... has no basis in providing the best learning. ... the first initial months of learning hinders a student to learn? ... wanted to play a kit, ... I fell in love w/the Drum Nazi and was grateful he busted my ass. ...
    (rec.music.makers.percussion)
  • Re: What Equipment Does a Beginner Start With
    ... has no basis in providing the best learning. ... the first initial months of learning hinders a student to learn? ... wanted to play a kit, ... I fell in love w/the Drum Nazi and was grateful he busted my ass. ...
    (rec.music.makers.percussion)
  • Re: Aaron Slomans "The Irrelevance of Turing Machines to AI" article
    ... > the student learns what he does. ... transfer of learning is not always that challenging as in an ... Humans accomplish ... use every kind of phone, cell phones, car phones, little phones, big ...
    (sci.cognitive)
  • Re: Educate Me - What Key is This?
    ... I say if ya just wanna putt then something slow and clunky ... Different teachers, same student. ... Tab isn't something you dedicate time to learning, ...
    (alt.guitar.bass)
  • Re: Dynamically initiallizing an array
    ... Arrays are fixed in size at the point of definition ... >the ability to understand how peoples minds work whilst learning. ... >Student to teacher: What do the angle brackets do? ... This is the syntax for template arguments. ...
    (alt.comp.lang.learn.c-cpp)