This is in response to Mike Tintners 10/11/2007 7:53 PM post. My
response is in all-caps.
Vladimir: ..and also why can't 3D world model be just described
abstractly,
by
> presenting the intelligent agent with bunch of objects with attached
> properties and relations between them that preserve certain
> invariants? Spacial part of world model doesn't seem to be more
> complex than general problem of knowledge arrangement, when you have
> to keep track of all kinds of properties that should (and shouldn't)
> be derived for given scene.
>
Vladimir and Edward,
I didn't really address this idea essentially common to you both,
properly.
The idea is that a network or framework of symbols/ symbolic concepts can
somehow be used to reason usefully and derive new knowledge about the
world - a network of classes and subclasses and relations between them,
all
expressed symbolically. Cyc and Nars are examples.
OK let's try and set up a rough test of how fruitful such networks/ models
can be.
Take your Cyc or similar symbolic model, which presumably will have
something like "animal - mammals - humans - primates - cats etc " and
various relations to "move - jump - sit - stand " and then "jump -
on - objects" etc etc. A vast hierarchy and network of symbolic concepts,
which among other things tell us something about various animals and the
kinds of movements they can make.
Now ask that model in effect: "OK you know that the cat can sit and jump
on
a mat. Now tell me what other items in a domestic room a cat can sit and
jump on. And create a scenario of a cat moving around a room."
I suspect that you will find that any purely symbolic system like Cyc will
be extremely limited in its capacity to deduce further knowledge about
cats
or other animals and their movements with relation to a domestic room -
and
may well have no power at all to create scenarios.
IT DEPENDS WHAT YOU MEAN BY PURELY SYMBOLIC. IN THE PAST SYMBOLIC
GENERALLY REFERRED TO SYSTEMS WITHOUT MUCH GROUNDING, SO THE SYMBOLS HAD
RELATIVELY LITTLE SEMANTIC MEANING. OFTEN SUCH SYSTEMS RELIED ON
RELATIVELY BRITTLE DEFINITIONS AND RULES OF INFERENCE.
THAT IS NOT THE APPROACH I ADVOCATE. (AND PLEASE DONT HOLD CYC UP AS A
GOOD EXAMPLE OF THE APPROACH I ADVOCATE. THERE IS A WORLD OF DIFFERENCE
BETWEEN A RELATIVELY OLD-FASHIONED AI SYSTEM LIKE CYC AND A STATE OF THE
ART AGI SYSTEM LIKE NOVAMENTE.)
THE STATE OF THE ART AGI APPROACH I FAVOR IS BASED ON (1) MASSIVE AMOUNTS
OF EXPERIENCES OF SOME SORT TO PROVIDE GROUNDING TO SYMBOLS AND (2)
FLEXIBLE RULES FOR MATCHING, INSTANTIATING, GENERALIZATION, AND
INFERENCING IN A CONTEXT-SPECIFIC WAY FROM SUCH MASSIVE EXPERIENCE, SO AS
TO ENABLE SOMETHING APPROACHING -- AND ULTIMATELY SURPASSING --
HUMAN-LEVEL INTELLIGENCE.
BUT SUCH SYSTEMS WOULD BE COMPOSED ALMOST ENTIRELY OF SYMBOLS. EVEN THE
TYPES OF SYSTEMS YOU SEEM TO BE FAVORING WOULD BE COMPOSED OF SYMBOLS.
BITS AND BYTES ARE, AFTER ALL, SYMBOLS. SO PLEASE LETS STOP KNOCKING
SYMBOLS, PER SE.
THE DISTINCTION SHOULD BE BETWEEN RELATIVELY NAKED SYMBOLS AND SYMBOL
GROUNDED IN NETWORKS OF MEANING I.E., NETWORKS OF RELATIONSHIPS SUCH AS
SENSORY PATTERNS (YES, I LIKE YOU THINK SENSORY EXPERIENCE IS GENERALLY
IMPORTANT), ASSOCIATIONS, CONDITIONAL PROBABILITIES, TEMPORAL RELATIONS,
CAUSE-AND-EFFECTS, ATTRIBUTES, FUNCTIONS, GOALS, VALUES, IMPORTANCE
WEIGHTINGS, GENERALIZATIONS, SPECIALIZATIONS, AND BEHAVIORAL SCHEMAS, ALL
IN THE CONTEXT POWERFUL INFERENCING AND AUTOMATIC LEARNING.
OF COURSE AS I SAID IN A VERY RECENT POST, GROUNDING COMES IN ALL SORTS OF
DIFFERENT TYPES AND DEGREES. SO DIFFERENT TYPES AND DEGREES OF
INTELLIGENCE CAN BE DERIVED WITH DIFFERENT TYPES OF GROUNDING. EVEN IN A
SYSTEM LIKE CYC OR WORDNET A CONCEPT WOULD NORMALLY HAVE SOME DEGREE OF
GROUNDING.
But you or I, with a visual/ sensory model of that cat and that room, will
be able to infer with reasonable success whether it can or can't jump, sit
and stand on every single object in that room - sofa, chair, bottle,
radio,
cupboard etc etc. And we will also be able to make very complex
assessments
about which parts of the objects it can or can't jump or stand on - which
parts of the sofa, for example - and assessments about which states of
objects, (well it couldn't jump or stand on a large Coke bottle if erect,
but maybe if the bottle were on its side, and almost certainly if it were
a
jeroboam on its side). And I think you'll find that our capacity to draw
inferences - from our visual and sensory model - about cats and their
movements is virtually infinite.
And we will also be able to create a virtually infinite set of scenarios
of
a cat moving in various ways from point to point around the room.
Reality check: what you guys are essentially advocating is logical systems
and logical reasoning for AGI's - now how many kinds of problems in the
real
human world is logic actually used to solve? Not that many. Oh it's an
important part of much problemsolving but only a part. How much scientific
problemsolving depends seriously on logic? Is logic going to help you
understand and have ideas about genetics or how cells work, or the brain
works, or how and why wars start? Is it going to be much use for design
problems? Does it help in telling stories? .. keep on going through the
vast
range of human and animal problemsolving (all of which remember are the
ONLY
forms of [A]GI that actually work).
That's why I asked you: give me some examples of useful new knowledge or
analogies [especially analogies] that have been derived from logical
systems
or logic, period (except about logic itself).
THIS TIME THE ANSWER DEPENDS ON WHAT YOU MEAN BY "LOGICAL." WIKIPEDIAS
BROAD DEFINITION OF LOGIC IS: THE STUDY OF THE PRINCIPLES AND CRITERIA
OF VALID INFERENCE AND DEMONSTRATION. THUS, THE TERM IS MUCH MORE BROAD
THAN THE BRITTLE FORMAL LOGICS THAT MUCH OF AI WAS HUNG UP ON FOR YEARS.
I AM NOT A BIG FAN OF TRADITIONAL FORMAL LOGIC. SINCE THE EARLY 70S I
HAVE SAID FORMAL LOGIC IS TO HUMAN THOUGH WHAT DRESSAGE IS THE MOTION OF
HORSES -- EXCEPT IN ITS SIMPLEST FORMS IT IS TOTALLY UNNATURAL. COMMON
SENSE NOTIONS, SUCH AS THE EXCEPTION THAT PROVES THE RULE INDICATES THAT
REASONING WITH BINARY TRUTH VALUES IS BRAIN-DEAD IN MANY DOMAINS.
BUT MANY FORMS OF LOGICAL REASONING ARE MUCH MORE FLEXIBLE. BAYESIAN
INFERENCING, FOR EXAMPLE, IS A TYPE OF LOGIC BECAUSE IT IS A TYPE OF
REASONING DESPITE ITS LIMITATIONS HAS SHOWN ITSELF TO BE EXTREMELY
VALUABLE. IT IS USED IN MANY SUCCESSFUL COMMERCIAL PRODUCTS. BAYESIAN
CLASSIFIERS, FOR EXAMPLE, HAVE BEEN USED TO MAKE NEW SCIENTIFIC
DISCOVERIES FROM VAST AMOUNTS OF SENSOR DATA. SO IN FACT, SOME TYPES OF
LOGIC ARE EXTREMELY VALUABLE AND DO HELP SCIENTISTS SOLVE PROBLEMS.
FURTHERMORE, IF YOU HAVE READ DOUG HOFSTADTERS COPYCAT, WHICH MAKES
CONTEXT SPECIFIC ANALOGIES, YOU REALIZE IT USES A FLEXIBLE SIMILARITY
SYSTEM, CALLED SLIPNET, THAT CAUSES SIMILARITY MEASURES TO BE TIGHTENED OR
LOOSENED IN A CONTEXT-DEPENDENT WAY. THIS ALLOWS COPYCAT TO HANDLED THE
DIS-SIMILARITES IN THE CORRESPONDING THINGS THAT ARE BEING COMPARED TO
MAKE AN ANALOGY.
NARS OR A NARS-LIKE SYSTEM COULD EASILY BE USED TO REPLACE HOFSTADTERS
SLIPNET, AND COULD ARGUABLY HAVE SIGNIFICANT ADVANTAGES OVER SLIPNET,
SUCH AS MAKING COPYCATS ANALOGY DRAWING PROGRAM MORE GENERALIY APPLICABLE
TO A WORLD KNOWLEDGE BASE. SO LOGIC OF THE TYPE FOUND IN NARS COULD
ACTUALLY BE USEFUL IN THE VERY FIELD OF DRAWING ANALOGIES THAT THE ABOVE
TEXT IMPLIES IT IS USELESS FOR.
IN RECENT YEARS THERE HAS BEEN A LOT OF WORK IN DESIGNING SYSTEMS THAT
AUTOMATICALLY LEARN APPROPRIATE PROBABILISTIC LOGICS. ONE OF THESE IS
NOVAMENTES PROBABALISTIC LOGIC NETWORKS, OR PLN, WHICH BEN GOERTZEL
REFERRED TO IN HIS POST OF 10/10/2007 4:45 AM ON THIS LIST. I DONT YET
KNOW HOW WELL ANY OF THESE SYSTEMS WORK, BUT THEY HOLD THE PROMISE OF
ALLOWING LOGIC TO DELIVER ALL OF THE VERY THINGS YOU SAY LOGIC CANNOT
DELIVER IN LARGE WORLD-KNOWLEDGE-COMPUTING AGIS.
SO, PLEASE LETS STOP KNOCKING LOGIC.
New knowledge - especially new science - comes primarily from new
observation of the world, not from logically working through old
knowledge.
Artificial general intelligence - the ability to develop new, unprogrammed
solutions to problems - depends on sensory models and observations.
Let me be brutally challenging here : the reason you guys are attached to
purely symbolic models of the world is not because you have any real
evidence of their being productive (for AGI), but because they're what you
know how to do. Hence Vlad's "why can't 3D world model be just described
abstractly.." He doesn't know - he just hopes - that it can. Logically.
What you need here is not logic but - ahem - evidence {sensory stuff].
BRUTAL CHALLENGE ACCEPTED.
(AGAIN, PURELY SYMBOLIC COVERS ANY DIGITAL SYSTEM, EVEN THE TYPE YOU
SEEM TO FAVOR.)
ACTUALLY, EVER SINCE I DID MY READING LIST UNDER MINSKY IN 1969-70, MY
GUIDING PHILOSOPHY HAS BEEN THE GIST OF K-LINE THEORY I.E., THAT ONE
REASONS ABOUT NEW SITUATIONS BY EVOKING MEMORIES OF PAST SIMILAR
SITUATIONS. SO I HAVE BEEN IN FAVOR OF EXPERIENTIAL REASONING FOR OVER
37 YEARS. AND I HAVE NEVER BEEN A BIG FAN OF FORMAL LOGIC FOR THE REASONS
STATED ABOVE.
BUT I SEEK TO AVOID BEING NARROW MINDED. I THINK THERE ARE MANY DIFFERENT
POSSIBLE TYPES AND DEGREES OF EXPERIENCE, THERE ARE MANY DIFFERENT WAYS IT
CAN BE REPRESENTED, ALTHOUGH SOME REPRESENTATIONS ARE MUCH MORE CAPABLE
AND EFFICIENT THAN OTHERS. THERE ARE MANY DIFFERENT DEGREES AND TYPES OF
INTELLIGENCE. NOT ALL AGIS NEED VISUAL MODELS, OR EVEN SENSORY MODELS OF
PHYSICAL REALITY. AGIS USED FOR SOME LIMITED DOMAINS MAY NOT EVEN NEED
MODELS OF 3-DIMENSIONAL PHYSICAL SPACE -- SUCH AS THE HYPOTHETICAL
PROGRAM-LEARNING AGI IN MY EARLIER POST OF TODAY. (ALTHOUGH IT WOULD
ALMOST CERTAINLY DEVELOP OR START WITH A GENERAL MODEL OF N-DIMENSIONAL
SPACES.)
I BELIEVE THE CONCEPT OF TURING EQUIVALENCE SHOULD OPEN OUR MINDS TO THE
FACT THAT MOST THINGS IN COMPUTATION CAN BE DONE MANY DIFFERENT WAYS.
ALTHOUGH SOME WAYS ARE MUCH LESS EFFICIENT THAN OTHERS AS TO BE
PRACTICALLY USELESS, AND ALTHOUGH SOME WAYS MAY LACK ESSENTIAL
CHARACTERISTICS THAT LIMIT EVEN THEIR THEORETICAL CAPABILITIES.
AS MUCH AS YOU MAY KNOCK OLD FASHIONED AI SYSTEMS, THEY ACCOMPLISHED A
HELL OF A LOT WITH FLY-BRAIN LEVEL HARDWARE. THUS, RATHER THAN DISMISS
THE TYPES OF REPRESENTATIONS AND REASONING THEY USED AS USELESS, I WOULD
SEEK TO UNDERSTAND BOTH THEIR STRENGTHS AND WEAKNESSES. BEN GOERTZELS
NOVAMENTE EMBRACES USING THE EFFICIENCY OF SOME MORE NARROW FORMS OF AI IN
DOMAINS OR TASKS WHERE THEY ARE MORE EFFICIENT (SUCH AS LOW LEVEL VISION,
OR FOR DIFFERENT TYPES OF MENTAL FUNCTIONS), BUT HE SEEKS TO HAVE SUCH
DIFFERENT AIS RELATIVELY TIGHTLY INTEGRATED, SUCH AS BY HAVING THE SYSTEM
HAVE SELF AWARENESS OF THEIR INDIVIDUAL CHARACTERISTICS. WITH SUCH SELF
AWARENESS AN INTELLIGENT AGI MIGHT WELL OPTIMIZE REPRESENTATIONS FOR
DIFFERENT DOMAINS OR DIFFERENT LEVELS OF ACCESS.
LIKE NOVAMENTE, I HAVE FAVORED A FORM OF REPRESENTATION WHICH IS MORE LIKE
A SEMANTIC NET. BUT ONE CAN REPRESENT A SET OF LOGICAL STATEMENTS IN
SEMANTIC NET FORM. I THINK WITH ENOUGH LOGICAL STATEMENTS IN A GENERAL,
FLEXIBLE, PROBABILISTIC LOGIC ONE SHOULD BE ABLE TO THEORETICALLY
REPRESENT MOST FORMS OF EXPERIENCE THAT ARE RELEVANT TO AN AGI --
INCLUDING THE VERY TYPE OF VISUAL SENSORY MODELING YOU SEEM TO BE
ADVOCATING.
Edward W. Porter
Porter & Associates
24 String Bridge S12
Exeter, NH 03833
(617) 494-1722
Fax (617) 494-1822
[EMAIL PROTECTED]
-----Original Message-----
From: Mike Tintner [mailto:[EMAIL PROTECTED]
Sent: Thursday, October 11, 2007 7:53 PM
To: [email protected]
Subject: Re: [agi] Do the inference rules.. P.S.
Vladimir: ..and also why can't 3D world model be just described
abstractly,
by
> presenting the intelligent agent with bunch of objects with attached
> properties and relations between them that preserve certain
> invariants? Spacial part of world model doesn't seem to be more
> complex than general problem of knowledge arrangement, when you have
> to keep track of all kinds of properties that should (and shouldn't)
> be derived for given scene.
>
Vladimir and Edward,
I didn't really address this idea essentially common to you both,
properly.
The idea is that a network or framework of symbols/ symbolic concepts can
somehow be used to reason usefully and derive new knowledge about the
world - a network of classes and subclasses and relations between them,
all
expressed symbolically. Cyc and Nars are examples.
OK let's try and set up a rough test of how fruitful such networks/ models
can be.
Take your Cyc or similar symbolic model, which presumably will have
something like "animal - mammals - humans - primates - cats etc " and
various relations to "move - jump - sit - stand " and then "jump -
on - objects" etc etc. A vast hierarchy and network of symbolic concepts,
which among other things tell us something about various animals and the
kinds of movements they can make.
Now ask that model in effect: "OK you know that the cat can sit and jump
on
a mat. Now tell me what other items in a domestic room a cat can sit and
jump on. And create a scenario of a cat moving around a room."
I suspect that you will find that any purely symbolic system like Cyc will
be extremely limited in its capacity to deduce further knowledge about
cats
or other animals and their movements with relation to a domestic room -
and
may well have no power at all to create scenarios.
But you or I, with a visual/ sensory model of that cat and that room, will
be able to infer with reasonable success whether it can or can't jump, sit
and stand on every single object in that room - sofa, chair, bottle,
radio,
cupboard etc etc. And we will also be able to make very complex
assessments
about which parts of the objects it can or can't jump or stand on - which
parts of the sofa, for example - and assessments about which states of
objects, (well it couldn't jump or stand on a large Coke bottle if erect,
but maybe if the bottle were on its side, and almost certainly if it were
a
jeroboam on its side). And I think you'll find that our capacity to draw
inferences - from our visual and sensory model - about cats and their
movements is virtually infinite.
And we will also be able to create a virtually infinite set of scenarios
of
a cat moving in various ways from point to point around the room.
Reality check: what you guys are essentially advocating is logical systems
and logical reasoning for AGI's - now how many kinds of problems in the
real
human world is logic actually used to solve? Not that many. Oh it's an
important part of much problemsolving but only a part. How much scientific
problemsolving depends seriously on logic? Is logic going to help you
understand and have ideas about genetics or how cells work, or the brain
works, or how and why wars start? Is it going to be much use for design
problems? Does it help in telling stories? .. keep on going through the
vast
range of human and animal problemsolving (all of which remember are the
ONLY
forms of [A]GI that actually work).
That's why I asked you: give me some examples of useful new knowledge or
analogies [especially analogies] that have been derived from logical
systems
or logic, period (except about logic itself).
New knowledge - especially new science - comes primarily from new
observation of the world, not from logically working through old
knowledge.
Artificial general intelligence - the ability to develop new, unprogrammed
solutions to problems - depends on sensory models and observations.
Let me be brutally challenging here : the reason you guys are attached to
purely symbolic models of the world is not because you have any real
evidence of their being productive (for AGI), but because they're what you
know how to do. Hence Vlad's "why can't 3D world model be just described
abstractly.." He doesn't know - he just hopes - that it can. Logically.
What you need here is not logic but - ahem - evidence {sensory stuff].
-----
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/?&;
-----
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/?member_id=8660244&id_secret=52711895-465380
