Re: Perceptual symbol systems
From: Sergio Navega (snavega_at_intelliwise.com)
Date: 08/11/04
- Next message: Allan C Cybulskie: "Re: Misc comments on categorisation, problem solving, etc (Was Re: Aaron Somon's....)"
- Previous message: Sergio Navega: "Re: Perceptual symbol systems"
- In reply to: Traveler: "Re: Perceptual symbol systems"
- Next in thread: Eray Ozkural exa: "Re: Perceptual symbol systems"
- Messages sorted by: [ date ] [ thread ]
Date: Wed, 11 Aug 2004 19:55:06 -0300
"Traveler" <traveler@nospam.com> escreveu na mensagem
news:mf9kh0lapt5jard1vsr4itf27ir7pk3376@4ax.com...
> 2In article <411a1906_1@news.athenanews.com>, "Sergio Navega"
> <snavega@intelliwise.com> wrote:
> >
> >I don't see a way to support this claim, other than the suggestion
> >that biological intelligences are neural. But that's not a good enough
> >reason to think that *only* neural systems must be used.
>
> It is a very simple claim to support, IMO, if only because the
> interconnectedness of intelligence is so great as to be intractable to
> formal means. Consider the sheer volume of sensory signals that any
> advanced intelligence must process in real time. The number of
> relationships (temporal correlations) which must be discovered among
> the signals is astronomical.
A Pentium IV may run a symbolic program and also be intractable
by formal means (provided it is someway linked to the environment,
which provides a great part of the "intractability").
> A correlation is nothing but a dependence
> (a link or connection) between two events. Considering that the number
> of correlations is much greater than the number of sensors, the only
> solution to this problem is a connectionist approach. What other
> approach can you offer that will discover and keep track of zillions
> of correlations between individual events?
It has been demonstrated that neural networks are computationally
equivalent to turing machines. They don't offer anything that a
turing machine cannot do. So our first choice must go to the
things that ease our conceptual thought.
> Now I am supposing (correct me if I am wrong) that your idea of
> intelligence may have to do with text parsing and the Turing test. If
> so, I can assure that text strings and Alan Turing have nothing to do
> with intelligence.
So do I. Text parsing or language understanding are things at the
top of the iceberg.
> [cut]
> >> This is precisely it. The neural approach frees you from all the
> >> constraints inherent in other approaches.
> >
> >Unfortunately, it adds a whole bunch of other constraints.
>
> Such as?
A phenomenon known as catastrophic interference (where new information
displaces old information, preventing a comparison of new and old
by the agent). Also the inability to generalize in a psychologically
plausible way (as demonstrated by Gary Marcus). The computational
inefficacy of the learning, at least of the traditional back-propagation
networks. A serious deficiency of neural nets in relation to symbolic
behavior, in special the processing of embedded clauses in natural
language or the mapping of analogical structures. Lots of these
deficiencies are corrected in special models (such as the LISA, of
Holyoak et al.), but that's what I mean by an implementation that
follows a theoretical guidance, and not the other way.
> [cut]
> >> If you are looking for abstractions, what could be more abstract than
> >> signals and signal processing? With the neural approach there is no
> >> need to talk about concrete objects in the world. The type of world is
> >> irrelevant, as long as it is consistent.
> >
> >Signals and signal processing are far from the abstractions one may
> >need to understand cognition.
>
> Funny you should say this because the brains of humans and animals use
> nothing but signals and signal processing. It is the lowest level of
> abstraction that is still within the domain of intelligence. But why
> should one need to understand cognition to create an intelligent
> machine. Again, the interconnectedness of human cognition is so high
> as to be intractable to our efforts at understanding it. Even if you
> could understand it (might take a trillion years), you can never keep
> up with it because it is continually changing. The only thing we can
> hope for is an understanding of the underlying mechanism/principles
> from which it emerges.
What I propose is that the principles of the basic mechanisms can be
abstracted without resorting to specific architectural idiosyncrasies.
> >For one, signals are very close to
> >sensory surfaces.
>
> Not in cell assemblies like the hippocampus and the basal ganglia. But
> I agree with you to a certain degree: Motor signals are removed from
> the sensory layers by no more than six or seven neurons. But is it not
> amazing what can be achieved with just a few layers of neurons?
However, how many networks of neurons are responsible for any specific
motor command? A single finger lifted may put at work tens of millions
of neurons.
> >Deep inside the brain the best level of analysis
> >may be different and this may have important consequences to our
> >understanding of the whole process.
>
> It is not different. It is all about temporal signal processing, as
> any competent neurobiologist will tell you. Signal processing is based
> on the premise that signals can be either simultaneous or sequential.
> It is that simple. As amazing as it may sound, these are the only two
> correlations used in the brain. The sensory layers deal with temporal
> contiguity (closely succeeding signals) whereas the hippocampus
> handles simultaneity and sequential correlations over multiple time
> scales.
What you say is equivalent to the idea that a computer uses 0/5v pulses
in all its architecture. All the richness and sophistication of a
computer is lost if one choses to stay at this level of analysis.
>
> >A conventional computer can be said to be doing analogical processing,
> >on a very low level of analysis. Or it can be said to be orchestrating a
> >dazzling flux of electrons, on an even lower level of analysis. Or then
> >it may be said that it is doing sequentially correlated logical (boolean)
> >operations, on a much higher level of analysis.
>
> As I have found in my research, logic is but a by-product of
> temporal/causal signal processing (which includes motor conflict
> detection) in the brain.
>
> >If we choose to understand
> >computers in that last sense, then we can conceive *another kind* of
> >computer implemented by, for instance, a system that controls flows of
> >water in tubes and bottles. Choosing the right level of analysis may have
> >impressive consequences on what one's imagination can achieve.
>
> You seem to be under the impression that intelligence can exist
> without sensors and sensory signals. It cannot.
I know that. Also, it is obvious that a neural system is appropriate to
process these signals. What I question is the idea that only neural
systems are able to do that.
> The lowest level of
> analysis (abstraction) one can get to while still remaining within the
> context of intelligence is that of signal processing. This is
> irrefutable, IMO.
This is as irrefutable as the idea that digital computers must use
a discrete system of signaling. It is a too basic level to be of
any use to an engineer concerned about the building of equivalent
machinery.
> >> >Given a set of rules and a symbolic pattern, one can manipulate it
> >> >and do something useful without ever needing to know what the symbols
> >> >stand for.
> >>
> >> And how is the intelligent system going to come up with the rules with
> >> which to manipulate the symbol if it has no clue as to what the symbol
> >> represents?
> >
> >It is not necessary to "know" what a symbol represents, provided one
> >has a good set of rules to use them.
>
> You're kidding me?
Not at all. Notice that I'm not suggesting that purely symbolic systems
are capable of "surviving in the natural world". That's not the case.
> > This is not to say that I support
> >symbolic (or rule-based) systems. I prefer to choose the "method" of
> >processing only after I understand what needs to be done.
>
> I can only hope that one day you will understand that what needs to be
> done is temporal signal processing. Nothing else. If I were forced to
> choose a single word to characterize intelligence, it would have to be
> *timing*.
And I would choose the word "pattern", whether static or dynamic.
Sergio Navega.
- Next message: Allan C Cybulskie: "Re: Misc comments on categorisation, problem solving, etc (Was Re: Aaron Somon's....)"
- Previous message: Sergio Navega: "Re: Perceptual symbol systems"
- In reply to: Traveler: "Re: Perceptual symbol systems"
- Next in thread: Eray Ozkural exa: "Re: Perceptual symbol systems"
- Messages sorted by: [ date ] [ thread ]
Relevant Pages
|
