John,
It is certainly clear that mental imagery plays a role in human
thinking, but this role does appear to vary from person to person, both
in extent and in nature. Take a look at Hadamard's old book "The
Psychology of Mathematical Invention" for a fascinating discussion of
the different sorts of mental imagery pursued by different people
(visual, acoustic, verbal, etc.). I myself use a lot of visual and
auditory imagery in my own thinking, but I know others who do not (at
least not at the conscious level).
In the Novamente design this is dealt with via a currently unimplemented
aspect of the design called the "internal simulation world." This is a
very non-human-brain-like approach, but I think it's an interesting and
ultimately very powerful one. What it means is that NM will actually
have, internally, a private 3D world-simulation, complete with a simple
physics engine. It can use this internal sim to experiment with
hypothetical actions in hypothetical situations, but also to draw
various abstract sketches and movies that don't correspond to any
real-world phenomena.
We haven't implemented this part yet due to the familiar lack of
adequate human resources, but I think it will be a valuable addition to
NM's cognitive arsenal. For the sim world, we would use the
CrystalSpace engine that we are now using (in the AGISim project) to
give NM a sim-world to use for embodiment and interaction with humans...
I don't really see mental imagery as a critical missing link btw the
symbolic and the subsymbolic. In NM, there is interaction & translation
between symbolic and subsymbolic knowledge without need for mental
imagery. However, in some cases mental imagery can provide insights
that would be hard to come by otherwise.
-- Ben G
John Scanlon wrote:
My philosophy of AI has never been logic-based or neural-based. I
did explore neural nets during the neural-net mania of the nineties.
I did a lot of reading, and experimented with some with feedforward
nets I wrote using simulated annealing and backpropagation (which
never did work very well). Neural nets seem to have potential as
one tool among several types of incremental learning algorithms,
including genetic algorithms and statistical methods, but in
themselves, they are no more than that -- useful tools, but not
the solution.
Language, which includes logic, is a way of representing ideas simply
and crudely. Good for communication and internal reasoning -- "if I
do this then this will happen, unless state X is the case, which means
that this other thing will happen," etc. My project uses an
artificial language (Jinnteera) for both these things, and the
language is integral to the whole thing. But it does not function
as the core knowledge-representation scheme.
So this brings us to what I've been calling the missing piece.
Artificial neural nets (as they currently exist) can function as
general-learning algorithms, but they don't represent knowledge of the
real spatiotemporal world well. They are too low-level for handling
what in human intelligence is thought of as mental imagery. Yes, in
the brain, it is all neural based, but in a non-massively-parallel von
Neuman computer system (even a PDP system), building a
100-billion-node neural net is computationally intractable (is that
the right word?). It has to be done differently.
The missing piece lies between low-level learning algorithms and
highest-level logical-linguistic knowledge representation. When a
human translator, at the U.N., for example, translates between Chinese
and English, he (or she) does it infinitely more effectively than any
translation software could do it, because there is an intermediate
knowledge representation that is neither Chinese nor English, but that
can be readily translated to or from either language by a fluent
speaker. The intermediate knowledge representation is non-linguistic
-- it consists of mental models constructed of sensorimotor patterns
representing a 3-D temporal world.
This sounds very vague and abstract, but I'm working on making it
concrete, in my system (Gnoljinn) -- developing the data structures in
code for implementing this knowledge-representation scheme. There's
been some talk here recently about 3-D vision systems, and this points
roughly in the direction I'm going in. Gnoljinn uses a single sensory
modality right now -- vision -- and will be restricted to it for a
good while, because, while it might be useful to have other sensory
modalities, none of them are absolutely necessary for higher
intelligence, and it's best to keep things as simple as possible
starting out.
I seriously wonder if I can do this project myself, or whether I need
to try to find some collaborators.
Yan King Yin wrote:
John Scanlon wrote:
> [...]
> Logical deduction or inference is not thought. It is mechanical
symbol manipulation that can can be programmed into any scientific
pocket calculator.
> [...]
Hi John,
I admire your attitude for attacking the core AI issues =)
One is either neural-based or logic-based, using a crude
dichotomy. So your approach is closer to neural-based? Mine is
closer to the logic-based end of the spectrum.
You did not have a real argument against logical AI. What you
said was just some sentiments about the ill-defined concept of
"thought". You may want to take some time to express an argument
why logic-based AI is doomed. In fact, both Ben's and my system
have certain "neural" characteristics, eg being graphical, having
numerical truth values, etc.
In the end we may all end up somewhere between logic and neural...
------------------------------------------------------------------------
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303
