Re: Cool visual illusion


Wolf Kirchmeir wrote:
> feedbackdroids wrote:
> > Wolf Kirchmeir wrote:
> [...]
> >>Try building a circuit/network that exhibits classical Pavlovian
> >>conditioning, for example. How do you do that? I don't know, I'm not an
> >>EE. But it looks to me like such a circuit/network is good deal more
> >>complex than anything built sofar (eg, the letter-recognition networks.)
> >>How do you have the network "associate" two inputs? Ie, how do you
> >>program the response system so that after a sufficient number of dual
> >>inputs from radically different sensors the network exhibits the same
> >>respone(s) for either input in isolation?
> >
> > I thought this might be your answer, but what you describe is no real
> > help. For one thing, an ant is not a symbol and language-processing
> > human.
>
> True, but it is one smart cookie compared to the most sophisticated
> robots that have been built sofar. They are also trainable.



What you've done with your answers is to show exactly what has been
pointed out many times. Beh provides some help at the highest-most
levels of the List, but not much towards building specific AI
solutions. We already know this.

If you think, as you do, that

> That's why I assert that making a bacterium-level learner is apparently
> beyond us at this time.

then what will help? That's the real question. Techniques of "classical
conditioning" have already been shown to not get very far. So, more of
the same will magically solve the problem? This approach can solve some
simple problems, is all, and has been tried many times already.



>
> > In addition, people have already been doing what you describe
> > since the time of Grey Walter and his tortoises, in 1948 or so. CORA =
> > conditioned reflex analogue. Simple variations on Brooks' subsumption
> > also do this. Most AI'ers would dismiss this level of intelligence as a
> > toy. People already have "many" schemes to do association,
> > letter-processing, etc. It's not that hard.
>
> These machine don't come close to implementing conditioning (for
> example.) Letter processing is not AI - it's just another form of
> pattern recognition (not pattern _learning_, as I understand it). And so on.
>
> > It is well illustrated from past work that simple learners will NOT do
> > the real job of AI. The search-space [problem-space] is just too
> > enormous for simple naiive learning approaches to handle. Many people
> > have already tried this route, and some like Curt and Bill.M still try
> > it.
>
> The devices you have alluded to are not "simple learners". They aren't
> even at the level of bacteria, which can be trained to change their
> responses to temperature gradients, pH gradients, etc. IOW, the people
> who built these machines did not (and do not) have a clear understanding
> of what they mean by learning, let alone what they mean by intelligence.
>
> They devices might be useful in building a learning machine though. EG,
> you might be able to use several CORA modules, plus assorted circuitry
> and software to implement a single conditioned reflex.
>
> > What evolution has given living organisms is all manner of hardwired
> > and/or modifiable sensory pre-processors which drastically limit the
> > search-space. This is why such organisms are able to survive, not just
> > because they can learn, which many do not even do to any great extent.
>
> Ah, I think I see a blind spot. I don't know of any organism that
> doesn't learn. That's one of the features of an organism: its ability to
> modify its responses to its environment. To say that what a worm does is
> not "learning to any great extent" is only possible if you don't in fact
> know what a worm can learn. It's is not simple learner: I don't know of
> any device or program that comes close to being able to do what a worm
> can do.
>
> I'm sure that you will accept that learning = changing behaviour. But I
> suspect that you haven't assimilated the full import of that equation,
> and are still hung up on symbol processing etc. I suspect further that
> you give too much weight to the kind of learning that's hard for humans
> to do, and don't give enough weight to the learning that "comes
> naturally" (which a lot of people don't even want to call learning, in
> my experience.) I know it took me a long time to get rid of the bias
> that all educated people have: that what they did in school and college
> is what constitutes the bulk of their learning, and/or is the model of
> learning.
>
> I'll give you a simple example: some snails can be "habituated" to
> continue moving forward into a hostile (say slightly too acid)
> environment. This is clearly a type of learning: its normal behaviour is
> to move away from lower pH. Yet the snail can be trained to move towards it.
>
> Clearly, some aspects of its physiology must change to allow it to
> change its behaviour. Does this change occur in the sensors of pH? In
> the internal chemical responses to pH, eg, some change in enzyme
> production? In the neurons that trigger avoidance when low pH is sensed?
> Or what? I don't know. I'm sure one could find out; I suspect that
> someone has already found out. Now suppose that the change in behaviour
> is mediated by a change in the chemistry of the sensors. Does that
> discovery change your concept of "learning?" Why/not?
>
> > I, for one, spent several years working with back-prop nets - naiive
> > generic learners - and they are just too limited in what they can
> > solve.
>
> Precisely. That's my point. They aren't even as smart as bacteria, as
> far as I can tell (the last time I read a technical description was
> about 5 years ago, so my impressions are getting rather hazy.) OTOH, I
> suspect a bacterium might be simulated by a network of such networks
> (and probably a few other types of NN, too.)
>
> > The one thing that behaviorists fail to comprehend, and factor in to
> > any great degree when arguing nature-vs-nurture, is the ENORMOUS extent
> > to which the innate [genetically-specified] mechanisms in the nervous
> > system reduce the problems that the learning systems have to deal with.
> > For every organism, genetics specifies overall organization, types of
> > sensory preprocessors, general environmental situations the orgranism
> > can effectively deal with, types of learning mechanism [if present],
> > specific architecture of said learning mechanisms, available behavioral
> > repertoires, on and on.
>
> I think them big numbers are a major discombobulation factor for you.
> They don't faze me at all. And I don't think nothing you say above
> invalidates EAB's contribution to the enterprise. If anything, it
> sharpens an understanding of what EAB can contribute: a better concept
> of how the "problem space" as you call it, is reduced to manageable
> proportions. Because that's what EAB studies: the "problem space" a she
> is actually encountered by the organism.
>
> Alos, EAB IMO has shown that, for most of us, our concepts of "learning"
> are not subtle and complex enough.
>
> > Without all this underlying infrastructure, simple naiive learning gets
> > you nowheres. There is a HUGE amount more to the systems than what you
> > indicate. Basically, what you describe is a trivial approach which
> > already has been shown for decades to not get us all that far. It's
> > little more than reactive-robotics with a learning module.
>
> Well, I beg to disagree. What you seem to think of as "simple naive
> learning" is neither. That's why I invited you to consider classical
> conditioning: I expected you to notice that it actually is rather more
> complex than "reactive robotics with a learning module." Glen then gave
> you a load more data to show just how complex it really is.
>
> That's why I assert that making a bacterium-level learner is apparently
> beyond us at this time.
>
> >>The behaviour will have to built-in, and hardwired to appear when input
> >>A is sensed in the environment. Sofar, so good: that's a simple task.
> >>Now add sensors that provide different inputs, B, C, etc, to which the
> >>system does not respond. What changes in the response system must occur
> >>so that it will (eventually) respond to B in the same way as it responds
> >>to A? Um, if we're simulating it, I suppose an array that counts input
> >>instances would be part of the system; but how you decide what a
> >>"sufficient number" of instances of A is? Hardwire it (ie, store that
> >>number in an unwritable array)? What do you do to ensure that the system
> >>will associate Input C with Input B, so that a secondary Pavlovian
> >>conditioning can occur? And what do you do to allow for the fact that
> >>when a conditioned behaviour is extinguished, the number of
> >>presentations to retrain it changes? And so on.
> >>
> >>It looks to me that if you think of the system's desired behaviours (and
> >>the desired changes in them) in EAB terms, you get a pretty strict
> >>specification of "what the client wants." Too strict, perhaps?
> >>
> >>You will object that this is a long way from implementing Artifical
> >>Intelligence. Quite so. But the reason for that is not that we don't
> >>know how to build an intelligent structure, but that we don't know, let
> >>alone agree on, what's meant by intelligent behaviour.
> >>
> >
> >
> >
> > This last is BS. This is not the limitation with implementing AI.
>
> Do you mean my claim that "we don't know, let alone agree on what's
> meant by intelligent behaviour"?
>
> >>So, leave the structure alone for the time being. Concentrate on
> >>intelligent behaviour, and see whether you can specify that precisely
> >>enough that you might be able to implement it on a machine.
> >
> > You can write software along this same approach using what are called
> > "stub"-routines. Stub-routines don't actually do anything, except to
> > simply output apparently-correct return values. People sometimes use
> > these to satisfy the customer that they have something to show, when
> > they still have most of the implementation work yet to do.
> >
> > It looks like it does something, but it just returns dummy values that
> > look like they mean something. I also use these when writing code, in
> > order to do high-level integration of code modules. Then, I go back and
> > write the actual code inside the modules. Having a return value
> > outputted involves writing 1 or 2 lines of code, whereas writing the
> > actual functional code might take 100s or 1000s of lines.
>
> OK, I used to call them "empty routines", it seems terminology has
> standardised since those long ago days. But I really don't see how that
> relates to my claim about understanding intelligent behaviour, nor my
> opinion that we can't as yet specify intelligent behaviour well enough
> to know what we are actually trying to do.
>
> HTH

.

Relevant Pages

  • Re: Cool visual illusion
    ... > conditioning" have already been shown to not get very far. ... >> of what they mean by learning, let alone what they mean by intelligence. ... That's one of the features of an organism: ... >> relates to my claim about understanding intelligent behaviour, ...
    (comp.ai.philosophy)
  • Re: What is conditioning?
    ... Kirchmeir's argument about innate walking ... There are all kinds of innate learning ... that you would consider is an example of conditioning ... A "simple" stimulus like a bell or odor also requires ...
    (comp.ai.philosophy)
  • What is conditioning?
    ... Kirchmeir's argument about innate walking ... There are all kinds of innate learning ... that you would consider is an example of conditioning ... A "simple" stimulus like a bell or odor also requires ...
    (comp.ai.philosophy)
  • Re: Cool visual illusion
    ... The search-space is just too enormous for simple naiive learning approaches to handle. ... Does this change occur in the sensors of pH? ... That's why I invited you to consider classical conditioning: I expected you to notice that it actually is rather more complex than "reactive robotics with a learning module." ... intelligent behaviour, and see whether you can specify that precisely enough that you might be able to implement it on a machine. ...
    (comp.ai.philosophy)
  • Re: Reply to Wolf
    ... learning to mean before Longley, Sizemore and Olea, even if it isn't ... So now you have a device that can be trained to recognise ten different symbols. ... the scientist must give free rein to his imagination, ... 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. ...
    (comp.ai.philosophy)