Re: Is universal artificial neural network possible?
- From: Jonathan <jonathan.attias@xxxxxxxxxxxx>
- Date: Tue, 06 May 2008 13:02:51 GMT
On May 6, 4:52 am, "Kenneth P. Turvey" <kt-use...@xxxxxxxxxxxxxxxxxx>
wrote:
On Tue, 29 Apr 2008 10:40:53 +0000, Dmitry A. Kazakov wrote:
What I was trying to get at here was more a simple description of the
problem complexity. We only know of one system with human competitive
intelligence, that is humans themselves.
Actually, I would challenge that too, for the reason that we don't know
what is essential for being either human or intelligent. Using one
ill-defined thing in order to clarify another, brings nothing.
In this case it actually does bring something. We agree that humans are
intelligent beasts. That means that such a thing can be created.
Second, since we know how long it has taken humans to evolve, we can
develop at least a general idea of the likely complexity of the problem.
We also have a good idea how
long it took to come up with a program for these systems.
But they weren't programmed. Programming is an engineering activity,
which evolution (at least in theory) is not. Even if we considered the
hypothesis of creative design, the comparison would still be invalid
because
I'm not even considering creative design (no evidence for it). The
comparison is valid since using evolution to solve the problem is
certainly one way of doing it.
1. the programmer teams are different
2. we have no idea how we, intelligent things, program other intelligent
things. We just don't, which is incidentally the whole problem...
3. even less we know about how our alleged masters did.
All three of these seem to assume intelligent design of some sort. We
got from not having a program to having one through a series of finite
improvements. This sounds very much like programming to me. It doesn't
require any goal oriented entities.
Right now we
have no evidence to indicate that we could do much better in developing
the software for a human competitive system. We could argue that
evolution is slow, but that doesn't really put an upper bound on the
complexity of the problem of developing such a program.
It certainly puts some time constraint, statistically. However that is
probably irrelevant as the event already happened - we call themselves
intelligent. A real constraint for Turing-complete systems could exist
if our brain used some incomputable elements.
Ok, to me this is silly. Just my opinion. If it was not computable,
then our brain clearly couldn't compute it.
OK, that would be a Turing test. I have a problem with it, because it
does not confirms anything as intelligent. It rather does that the
tester is not enough intelligent to denounce the respondent. I.e. if
this is a fitness function then for another problem.
OK, you and I simply view the world differently. I'm going to have to
side with Turing on this. If you can't tell if it apart from a human
then it must be as smart as a human at least in the dimensions of
interaction. I can't tell exactly what's going on in your mind either,
but I assume you are intelligent since your interaction with me leads me
to that conclusion.
Now, that said, I would be more comfortable if we broke into a number
of dimensions and recognize that we are really trying to optimize many
different things, but this is still an optimization problem.
Maybe we could state it as an optimization problem if we knew more about
what intelligence is, but we didn't so far. We also know nothing about
the complexity of the problem if stated in this form. Evolution is
solving a completely different problem and the best solutions found
(bacteria, insects etc) aren't any intelligent.
We can state it as an optimization problem already. We do it all the
time. That's what standardized testing is all about. We do it in many
other ways too. When you got your drivers license you passed a small
portion of a human intelligence test. We could very easily put together
100 of these and come up with a metric for what it means to have roughly
human level intelligence. Not only would it not be hard, most of the
work would already have been done.
We may not have a good scientific definition of what intelligence is, but
we have a very good working definition of what it is.
--
Kenneth P. Turvey <kt-use...@xxxxxxxxxxxxxxxxxx>
One thing we have to consider is that the Human brain (any brain), and
the "intelligence" it conveys does not exist in isolation. Rather, it
is an integral part of the whole body and has evolved to work with
that body.
If we transplant a human brain into a dog (don't try this at home
kids) then we wouldn't expect that dog to be able to talk or even to
walk. The "hardware" just isn't made to fit the "software". In fact,
the result probably wouldn't appear to be very intelligent at all,
just a blubbering wreck.
Side note: This example is actually not extreme enough because there
are likely to be enough similarities between human and dog physiology
that the brain could actually adapt and show some signs of intelligent
behaviour. So, if it helps the argument, consider a human brain
working a house fly where the eyes are constructed in an entirely
different way to mammals, there are too many legs plus there's those
flappy wing things on the back.
So, although our brains have a "general intelligence" aspect to them,
that can only be measured or observed within the context of system for
feeding it with sensory information (text in the case of the classic
Turing text, sight, sound, smell, touch, etc. for humans) and a means
to express some behaviour (text on a vdu or speech and muscle
activity).
In the absense of a true, measurable definition of "intelligence",
Turing tests are about the best we have - using one intelligent being
to "vote" on the level of intelligence of another being. Is an
earthworm intelligent? What about a frog? I sometimes even wonder
about some of my neighbours!
So we could devise 100 (1000?) Turing tests to cover (all) aspects of
percieved intelligence as a means for measurement but they have to be
tailored to the being as a whole.
Passing a driving test in a standard car might be a good test for a
humanoid robot. If it could climb in and out of a car, understand the
examiner, communicate back, control the car, perform the desired
manoeuvres and take into account everything else going on in the
environment then it's certainly showing a lot of intelligent
behaviour. In fact, that one test probably involves many other smaller
tests.
On the other hand, maybe it could just be clever programming for that
specific task rather than general intelligence. Could it also make a
good cup of tea for instance? More to the point, if it can't, is that
because of a lack of intelligence or a lack of the right hardware to
pick up the delicate teabags and pour from the kettle? My dog can't
make tea but I'd definitely call it intelligent.
So the tests have to be applicable to the capabilities of the being
under test.
Could we develop a universal intelligence "module" (neural network or
otherwise)? I don't see why not but if it's going to be "plugged in"
to an artifical being (robot or "deep thought" type computer) and be
immediately useful then it will have to be pre-programmed (as a
starting point) from an identical module that has already been trained
on the same kind of hardware and for the same purpose. In this sense,
"training" could mean evolution, incremental improvement, optimisation
of fitness for purpose or any other applicable technique.
And there's the rub. How could the training occur and how long would
it take?
[ comp.ai is moderated ... your article may take a while to appear. ]
.
- References:
- Re: Is universal artificial neural network possible?
- From: Kenneth P. Turvey
- Re: Is universal artificial neural network possible?
- Prev by Date: Re: Is universal artificial neural network possible?
- Next by Date: CFP: EXTENDED Deadline KASW'08: "Knowledge Acquisition from the Social Web"
- Previous by thread: Re: Is universal artificial neural network possible?
- Next by thread: Re: Is universal artificial neural network possible?
- Index(es):
Relevant Pages
|
