Re: Reply to Wolf

JGCASEY wrote:
On Jul 4, 3:00 am, Wolf <ElLoboVi...@xxxxxxxxxx> wrote:

[...]

JC:
Saying that my machine is not being "conditioned" like an animal
without any possible explanation of how it could ever be "conditioned"
leaves me with nothing. So I will just go back to what I understood
learning to mean before Longley, Sizemore and Olea, even if it isn't
what animals do when they learn, that is, if it changes its behavior
for the better (with respect to some measure of better) as a result of
experience.

Wolf:
That is a very high level and abstract account, IOW, handwaving. Just
what do you mean by "changing behaviour"? Why must it be "for the
better?" (Much of our actual learning is for the worse.) What exactly
do
you mean by "experience?" How does the system "get experience?" Why
does
some experience produce behaviour changes, and other experience
doesn't?
How would you differentiate "good experience", which improves
behaviour
from "bad experience", which makes behaviour worse? Etc and so on and
so
forth. -- To frame such questions, and attempt to reframe them so
that
they make sense is an experimental (or design) setting, is to engage
in
EAB.

JC:
When something changes it behavior for the worse most people would
consider that was not learning.

Well, what most people consider to be learning is irrelevant here.

Learning means knowing something you
didn't know before.

If that's the sum total of the concept "learning" for you, it's time you reconsidered the concept.

Look, John, learning is simply a change in behaviour brought about by some interaction with the environment. That's all. "Knowing something that you didn't know before" means among other things "being able to say/write things you couldn't say/write before", for example. That's a change in behaviour. It's also why we make students write tests as proof that do know things they didn't know before.

You cease to fall off the bike or get the exam
questions wrong. But I do understand that learning can be learning
something wrong and acting on it because you are unable to see it
is a load of nonsense.

Learning is a change in behaviour. That's all.

Now you ask why does some experiences produce changes and other
experiences don't. Ultimately you must look to the mechanisms to
answer that question. Indeed if you know what those mechanisms
are you may even predict the class of behaviours possible and
the kinds of mistakes that will be made.

Well, yes, but the class of behaviours possible is fairly obvious just from the gross anatomy of the animal or machine. A robot without graspers will not be able to pick anything up, for example. Then there's the issue of "mistakes." What's a mistake in one context is the right thing to do in another. So how can you tell that from knowing the mechanisms for learning?

For example in my first efforts at AI it was a problem to read
hand written characters via a touch pad. I wrote three simple
but different methods to learn to recognize the numeric characters
0 to 9. Their actual behaviors, how they learned, the mistakes
they made, were different because of how they worked although
they all essentially did the same thing.

Cool.

So now you have a device that can be trained to recognise ten different symbols. Can the same device be trained to recognise a different set of ten symbols? If yes, you have a generalised recognition machine (or program). If not, you have at best a genetic algorithm that builds a number recognition module (which would of course be useful as part of some more complex device.) -- If the device can be trained to recognise a different set of symbols, does it still recognise the original set? Or would you have to combine several of these devices in order to build up a larger recognition set? If so, how do you prevent the device from learning new symbols, and thus forget the old ones?

Wolf:
You will become a behaviorist despite yourself.

JC:
Really, I had no problem with behaviorism, as I understood it, until
being whacked continually by the behaviorists in this group. I just
find their narrow take on the problems and restricted language set
to be counter productive when trying to imagine how to get a machine
to do intelligent things. It is much a style difference as anything
else. Any attempt to theorize about getting a machine to do something
becomes "phlogiston". But that is what an AI has to do. Invent ways
of getting a machine to do things. Those ways may not be the same way
a biological machine does it. It may be "phlogiston" to a biological
machine. But for a man made machine the proof is in the eating. Run
that program. Press that start button and see what happens.

AI is about man made machines we can understand fully that we are
trying to get to do things that if done by a human (or animal) we
might deem intelligent. It is in that sense pure behavior because
we are not trying to duplicate the actual physiology of animals.

Um, if we can understand them fully, they are probably nor worth building.

Imagine how far mathematics and physics would have progressed if they
had limited themselves to a restricted set of ways to think about a
problem. Sorry Einstein, stick to Euclidean geometry, space curving
is so much phlogiston. Sorry we can't have the square root of -1
that is stupid, so much phlogiston.

That's a bad analogy, because it mistakes the subject matter for the method.

And what makes the radical behaviorist think phlogiston was such
a bad theory at the time just because it wasn't exactly right?
Theories lead to experiments which lead to new theories. I have
a theory how to get a machine to do X. Run the theory and see how
close or far the theory was from explaining one way it might be
done.

Huh?????

To quote from a book on the philosophy of Natural Science by Carl
G. Hempel: "The transition from data to theory requires creative
imagination. Scientific hypotheses and theories are not derived
from observed facts, but invented in order to account for them."
"... the scientist must give free rein to his imagination, and
the course of his creative thinking may be influenced even by a
scientifically questionable notion." It goes on about even in
disciplines such as mathematics, where results are validated
exclusively by deductive reasoning, imagination and free invention
play an important role.

Yup. So exercise your creative imagination. Speculate on what a conditionable mechanism would have to do with the data it receives so that data that did not produce a given response initially will do so after the machine "has learned what to do with the data."

Who knows what way out crazy ways real brains work. Maybe in time
we will untangle it all. But it will not be done by behaviorists
who simply catalogue behaviors it will be done by others who bring
other areas of expertise to the problem. They may not be restricted
to observing actual behaviors, they may be able to predict them
from first principles. They may do it without an exhaustive
analysis of complex animal behaviors which may in fact hide the
simple principles evolved.

Behaviourists do not simply catalogue behaviours. They study changes in behaviour resulting from carefully observed (and usually controlled) interactions with the environment. In doing this, they discover new behaviours, and/or that what seem simple behaviours are in fact complexes of simpler behaviours. Actually, AFAIK, there is no good catalogue of behaviours, let alone a classification into different levels of complexity. If there were, your job would be a lot easier, believe me.

In other words you cannot derive a solution as to the mechanisms
required from the data alone. And it is a heuristic that if you
cannot solve a complicated problem you might be better able to
solve a simpler although perhaps similar problem.

Yep, and that's why I say you should try modelling a bacterium. Or if that's too simple for your ambitions, try a flatworm. C. elegans will do nicely. It has 100 neurons, and IIRC all its neural connections have been mapped. A good starting point for studying how conditioning works in a real neural net, and hence learning what general principles of "mechanism" must be used in designing a machine that can do what C. elegans can do. C. elegans can be taught quite a few tricks, as it happens. Read up on it.

Rather than worry too much about the complicated analysis of
classical conditioning you described we might start with a simpler
but similar behavior in a machine. At least one that functionally
produces usable results if not strictly (or not at all) biological.

Damn it, John, there is no simpler learning behaviour than classical conditioning. And what I describe is not at all complicated IMO.

Changing the intensity/frequency/etc of a given response (eg, becoming conditioned to cold by taking cold baths) is AIUI more complicated because it requires feedback. Learning the difference between two inputs along the same channel (monkey face vs human face) is also more complicated, because it too requires feedback. But feedback must be other channels than the primary one, the response to which you are trying to modify. In a sense, you are trying to set up classical conditioning within the system, by making one or more parts of the system the secondary stimulus/stimuli for another part's response. Or so it seems to me. (I don't know how plausible a speculation this is.)

What possible use would a classical conditioning inspired behavior
be in a robot? How would we implement it? And so on ...

Well, the same use as in training dogs, I think. A robot that could be trained as a dog is trained would be a damn useful machine. Assuming it had a large enough repertoire of built-in behaviours, that is.

--
Wolf
'Just because it's true doesn't mean it's the right answer.'
.

Quantcast