Thanks Bill for the Eric Baum reference.
Deep thinker that I am, I've just read the book review on Amazon and
that has orientated me to some of the key ideas in the book (I hope!) so
I'm happy to start speculating without having actually read the book.
(See the review below.)
It seems that Baum is arguing that biological minds are amazingly quick
at making sense of the world because, as a result of evolution, the
structure of the brain is set up with inbuilt limitations/assumptions based
on likely possibilities in the real world - thus cutting out vast areas for
speculative but ultimately fruitless computation - but presumably limiting
biological minds' ability to understand phenomena that go beyond
common sense that has been structurally summarised by evolved
shortcuts. (That must be why non-Newtonian phisics always makes my
brain hurt!)
I'm sure that most people on the list who are heavily into developing
AGIs will have traversed this ground before. But I wondered......
(By the way....what follows is most likely not of any interest to people
well versed in this issue..what I'm doing is feeding back to the list my
understanding of this issue in the hope that somebody who knows all this
stuff can can tell me if I'm on the right track...so I'm really hoping I can
learn something from both my own cogitations and from the feedback
others can offer someone still very much in the AGI sandbox.)
So here we go......On the face of it, any AGI that is not designed with all
these short cuts and assumptions in place has a huge amount of
catching up to do to develop (or learn) efficient rules of thumb
(heuristics?). Given the flexibility of AGIs and their advantages of
computation speed and accuracy, the 4000 million years of evolutionary
learning could perhaps be recapitulated in rather less time. But how
much less? Would it only take I million years? 100,000 years, 100
years? I'm sure, Ben that you won't want to be sitting around traiing a
baby Novamente for that long.
Perhaps AGI's need to be structured so that their minds can do two
things:
- absorb rules of thumb from observations of other players in the world
around them (like children picking up ways of thinking from grown ups
around them) or utilise rules of thumb that are donated to it via
data dumps.
- be prepared to and be capable of challenging absorbed rules of thumb
and be able to revert to a systematic, relatively unbiased
exploration of an issue when rules of thumb turn up anomalous results
or when the AGI simply feels curious to go beyond the current rules
of thumb
Maybe all the databases of common sense relationships that Cyc is
developing and the Wordnet database etc. can be considered to be huge
sets of inherited rules of thumb ie. they are not derived from the
experience of the AGI.
The biggest problem for an AGI starting to learn seems to me to be able
to simply get to first base whereby the AGI can make *any* sense of its
basic sensory input. It seems to me that this is the AGIs hardest task if it
doesn't have any built in rules of thumb to orientate it.
Maybe an AGI does have to see the world through the lens of inherited
rules of thumb in it's first hours and even years in order to boost it's
competence at interpreting the world around it and then it can go about
replacing inherited rules of thumb with its own grounded self-generated
rules of thumb?
Maybe it needs to have an inbuilt program a bit like an optical character
recognition program that takes each class of incoming data and sifts it
into pre-recognised categies of data - ie. patterns can be letters,
numbers, colours, shapes, spacial orientation (up, down, left right,
forward, back etc,). Once the AGI is used to dealing with these preset
categories it could be fed more abiguous data where it has to perhaps
invent new categories of its own.
Presumably this is all very obvious, but from comments Ben has made
over a fair length of time, it seems he's very reluctant to fill an AGI's
head full of downloaded data/rules of thum or whatever. Ben, the
language you use suggests that you'd be happy to start with none of this
downloaded stuff. But it seems to me that an new Novamente would
struggle really badly, perhaps floundering endlessly in its effort to
interpret incoming data unless it's primed to make some good guesses
and to have some preset notions of what to do with this incoming data.
It seems to me that a new-born Novamente needs to be able to use lots
of preset rules related to its first learning environment so that of the data
coming in, a very large amount of it already makes sense at some level
so that the AGI can apply most of it's brain power to resolving a few very
simple ambiguities - like we do when solving a jigsaw puzzle. It seems
to me the key learning experience comes from successfully mastering
these very minor areas of ambiguity thus starting to build up some
personally grounded understanding - which can be added to
(exponentially?) as the AGI retests the validity of its understanding
based on inherited rules of thumb and as it builds a more and more
complex picture of what's around it - at each level gaining mastery
through resolving minor ambiguities at the new level of understanding.
If this model is right then perhaps it shouldn't matter if the AGI has been
given a humungous pile of downloaded data/rules of thumb? It would
just call on data in the databanks as these seem to be have some useful
connection to the data/rules of thumb that the AGI has mastered.
Initially the AGI would understand so little that virtually all of the data in it
storages would be just so much noise. It would only be able to work it's
way into the data as it mastered some initial concepts and concept
labels. So in that sense an infant AGI wouldn't be burdened with having
too much downloaded ungrounded data - because most of that data
would be efectively invisible to it. Isn't this pretty much like a child that
has grown up in a house with a huge library, the contents of which only
make sense very slowly as the child builds level after level and area
after area of base knowledge?
Anyway enough for now. If anyone has time for a babe in the sand box
I'd love to know what you think of these musings!
Cheers, Philip
---------------
What Is Thought?
by Eric B. Baum (Author)
Publisher: MIT Press; (January 1, 2004)
ISBN: 0262025485
Review: In What Is Thought? Eric Baum proposes a computational
explanation of thought. Just as Erwin Schr?ger in his classic 1944 work
What Is Life? argued ten years before the discovery of DNA that life
must be explainable at a fundamental level by physics and chemistry,
Baum contends that the present-day inability of computer science to
explain thought and meaning is no reason to doubt there can be such an
explanation. Baum argues that the complexity of mind is the outcome of
evolution, which has built thought processes that act unlike the standard
algorithms of computer science and that to understand the mind we need
to understand these thought processes and the evolutionary process that
produced them in computational terms. Baum proposes that underlying
mind is a complex but compact program that corresponds to the
underlying structure of the world. He argues further that the mind is
essentially programmed by DNA. We learn more rapidly than computer
scientists have so far been able to explain because the DNA code has
programmed the mind to deal only with meaningful possibilities. Thus
the mind understands by exploiting semantics, or meaning, for the
purposes of computation; constraints are built in so that although there
are myriad possibilities, only a few make sense. Evolution discovered
corresponding subroutines or shortcuts to speed up its processes and to
construct creatures whose survival depends on making the right choice
quickly. Baum argues that the structure and nature of thought, meaning,
sensation, and consciousness therefore arise naturally from the evolution
of programs that exploit the compact structure of the world.
To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
