I think one can now present a convincing case why any symbolic/linguistic
approach to AGI, that is not backed by imaginative simulation, simply will
not work. For example, any attempt to build an AGI with a purely symbolic
database of knowledge mined from the Net or other texts, is doomed.
This is obviously something I have long argued, but it has been difficult to
find a truly focussed argument with sufficiently general application and
power.
The basic argument:
Language depends on
1) "General Activity Language" - a core, very extensive vocabulary of words
for basic kinds of movements, which we all acquire normally very early.
These words/movements are essential for moving about, and manipulating the
world - and understanding how the world moves. They are also essential for
General Intelligence, because they apply to all activities, and are central
to the acquisition of new physical activities.
2) Our movement words, (like, in fact, all words), are general, open-ended
concepts which cover, in this case, vast, "all-encompassing" ranges of
specific, possible movements. In order to interpret them, we continually
have to decide which one of a vast range, is appropriate in a given
environment - for example, just which direction and angle we are going to
decide appropriate to "reach out" - horizontally, vertically, at 45% , 60%,
75%, etc. etc.
3). It is, if not absolutely impossible, utterly impractical, and absurdly
complicated, to instantiate a movement-word by any kind of symbolic
process - by, for example, first trying to symbolically label each and every
one of a range of possible movements.
The only practical way - and the ideal way - is to decide the specific
movement, by an imaginative/ sensorimotor simulation. Exactly what this
should entail is open to discussion, (and getting much discussion
elsewhere), but for the sake of focussing our minds here, let's think of it
, if only provisionally, as some kind of visual mapping process.
4)The same basic argument can be extended to every area of language. I am
focussing on this particular area because it is not only fundamental to any
worldview, but can be treated very concretely, and from a more or less
mathematical and robotics POV.
The argument in detail:
1) General Activity Language - it is acknowledged that we rapidly acquire a
certain vocabulary of basic words. What I'm focussing on here is that we
especially acquire a core of hundreds of basic movement words, such as:
"reach, push, pull, hit, punch, throw, kick, wave, catch, handle, grab, put,
move, enter, exit, slip, slide, remove, connect, disconnect, fit, step,
stride, walk, run, climb, jump, hop, leap, press, lift, raise, lower, drop,
pick up, fall, slip, knock, tap, shake, rock, roll, scratch, settle,
unsettle, slap, slop, fix, propel, repel, rope, stick, withdraw, touch,
finger, point, hold, snatch, thrust, scrape, grip, grasp, grope, back,
support, circle, rotate,
These can be considered as basic level concepts, which, like "dog", "cat,"
"bird," "chair," are the easiest to visualise - in this case as movements.
We also acquire a range of superordinate movement concepts, involving much
more general, and not so immediately obvious to visualise, categories of
movement, like:
"come, go, make, start, stop, give, take, use, do, be, get, dance, play,
heat, cool, add, subtract, travel, journey, advance, retreat"
(These can be compared to similar superordinate, not so
obvious-to-visualise, concepts such as: "animal," "furniture," etc.)
We also acquire a rich range of subordinate concepts, involving more
specific types of movements, some of which may belong to specific
activities, like:
"hammer, nail, screw, chop, slice, net, bat, elbow, head-butt, pin, clip,
vacuum, catapult, glue, brick, "
We also acquire a whole set of prepositions which give direction to those
movements, such as:
"in, into, on, onto, out, towards, away from, up, down, through, around,
inside, outside, over. under, along, underneath, about "
An AGI POV allows us to appreciate that this core vocabulary is a brilliant
invention of the human mind, although no doubt, animals share many of the
same concepts. These are general movements which can be applied to any
physical activity. They can be, and are, used to acquire new physical
activities/ skills. Look at the instruction manuals for virtually any
activity, and you will find that extensive use is made of these basic words.
A how-to-cook or a how-to-play-a-sport manual will liberally tell you to
"move", "put," "take," "go," "add" etc, and won't be couched in entirely
activity-specific words, like "play a forehand/backhand/ drop shot," or
"execute a pas-de-deux". (Any AGI must have this vocabulary to succeed).
2) Our movement concepts are, like all our concepts, general and open-ended.
They cover vast ranges of possible specific movements, typically
"all-encompassing." For example, concepts like "reach," "push," "pull" can
have a full, 360% range of possible trajectories -
http://www.mediafire.com/imageview.php?quickkey=ad4mmmqgdom&thumb=4
and the same range again for every plane in 3-D space. (All suggested
improvements to my geometrically crude exposition here are welcome). There
may be then many other variables. The hands can be shaped in many different
ways. The arms can take many shapes, not just straight ones, and twisted in
many different positions. The body can be adjusted or bent in many ways to
support the movements. And the movements can follow many different paths,
not just straight ones.
Obviously, the possible ranges of movement will vary according to the limbs
used. But no matter what limbs, the possible ranges are still vast - think
of the virtually infinite shapes the hand can take to "handle" objects.
[And new configurations for any given movement can always be, and often are
being, invented. "He reached out his arm in a V-shape." "John Cleese's silly
walk starts from a crouched goose-step."]
The prepositions of direction offer a similarly vast range of
possibilities - there are a vast number of different directions to move
things "over" a table/bridge/ground/hill/ mountain/ wave/logs etc etc.
2b) Movement concepts can of course be used in purely general ways - "He
likes to kick things". But a great deal of the time, they have to be
instantiated in specific forms, whether we are reading sentences about
movements, or taking briefs/instructions about how to move from others.
And they may have to be instantiated in any of the vast range of possible
ways.
"He reached for 1. the light bulb 2. showerhead 3. her shoulder 4. her
breast 5. the box on the table 6. book on the chair 7.hem of her dress 8.
his shoelace...etc" may all require trajectories at different angles to be
understood.
"He handled the 1. pin 2. orange 3. pen 4. knife 5 hand 6. soil 7. mud
...etc" may require any of a virtually infinite range of hand positions to
be understood.
And these different instantiations *have* to be fairly precise, if we are to
understand a text, or effect an instruction, successfully. The next sentence
in the text may demand that we know the rough angle of reaching - and that,
say, it was impossible because there was a particular kind of object in the
way.
3) The ideal and simplest way to work out which specific movement is
required is by imaginative simulation - here some visual, and/or some other
sensorimotor, schemas. It is typically straightforward to mentally draw a
line/trajectory from the relevant limb to the relevant target, from the arm
to the lightbulb, say, or the shoelace. We also find it simple when working
out how to handle different-shape objects, to kinaesthetically and sometimes
visually form our hands to corresponding shapes in advance.
It would be absurd and almost certainly impossible to try working out
movements by symbolic means - by, say, listing every possible angle at which
an arm can reach out, and listing the normal heights of different objects
that can be reached for - or trying to apply some set of mathematical,
formulaic approach to the problem.
Clearly, the human mind doesn't do this. We rely on intuitive, sensorimotor
models of how we and others move, and how objects move, both to plan our
movements, and, when reading texts, to work out others' movements. These are
for the most part purely imaginative models, without symbolic labels. This
is evident in how we have extreme difficulty when we try to
symbolically/verbally analyse most of our physical skills and habits, like
how we ride a bike, reach for a knife, handle a shopping bag etc etc. Our
imaginative knowledge generally of how we and the rest of the world move, is
vastly greater than our culture's vocabulary.
In addition, working out specific movements depends typically on often
complex imaginative models of environments - of rooms, for example, and
where tables, showers, paintings, chairs etc are normally placed - models
which would also be absurd to try and hold in symbolic as opposed to
imaginative form.
And the neuroscientific evidence keeps piling up that we do indeed plan
movements by maps/ imaginative simulations.
4)It shouldn't be too hard to see that the necessity of testing symbolic
language by imaginative simulation applies, by extension, to many other
areas of the world, as well as that of the movements of objects and
creatures. Descriptions of the forms of all objects and things. All
physical activities - hunting, sex, eating. All interactions between
creatures. Conversations. Emotions... Statements about all these also
typically depend on physical, imaginative knowledge of things' forms,
movements and behaviour.
In fact, there is, as Lakoff argues, no area that can be understood without
imaginative simulation, But I accept the need to demonstrate this further
with respect to more abstract areas. By all means challenge me, and I'll
think about it.
In the meantime, I believe I have made a convincing case that you cannot
understand how the world moves - and the core movement vocabulary of
language - without imaginative simulation. And if you can't do that, you
can't have a viable worldview.
-------------------------------------------
agi
Archives: http://www.listbox.com/member/archive/303/=now
RSS Feed: http://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4
Powered by Listbox: http://www.listbox.com