I think YKY is right on this one. There was a Dave Barry column about going to 
the movies with kids in which a 40-foot image of a handgun appears on the 
screen, at which point every mother in the theater turns to her kid and says, 
"Oh look, he's got a GUN!"

Communication in natural language is extremely compressed. It's a code that 
expresses the *difference* between the speaker's and the hearer's states of 
knowledge, not a full readout of the meaning. (this is why misunderstanding 
is so common, as witness the "intelligence" discussion here)  

Even a theoretical Solomonoff/Hutter AI would flounder if given a completely 
compressed bit-stream: it would be completely random, incompressible and 
unpredictable like Chaitin's Omega number. Language is a lot closer to this 
than is the sensory input stream of a kid.

There's a quote widely attributed to a "William Martin" (anybody know who he 
is?): "You can't learn anything unless you almost know it already." In 
general, the hearer needs a world model almost the same as the speaker's. 

Let's call this "Winograd's Theory of Understanding": that having a model 
capable of simulating the domain of discourse is necessary and sufficient for 
understanding discourse about it. (NB: (a) there are different levels of 
completeness and accuracy for simulations and also for understanding; (b) 
"symbol grounding" in the sense of associations to physical sensory/motor 
signals is *not necessary*.)

I find SHRDLU and its intellectual descencents a convincing demonstration of 
WTU. This implies that understanding an NL sentence consists not only in 
parsing it into an internal representation and stashing it somewhere, but, if 
it's something you didn't already know, modifying and augmenting the 
mechanism of your world model to reflect the new knowledge in future 
simulations. In other words, building a working mechanism and integrating it 
into an existing vast, complex machine.

Josh

On Saturday 28 April 2007 03:29, YKY (Yan King Yin) wrote:
> "Layered learning" is not just better, it's actually the only
> computationally feasible approach.
>
> We may talk to a baby like:
> "MILK?"
> "You want to play BALL?"
> "Oh you POO-POO again" etc.
> And these things are said simultaneously as some *physical* events (eg
> milk, ball, poo) are happening, which allows the baby to correctly *bind*
> the words to concepts, ie achieve grounding.
>
> Contrast this with something from Wall Street Journal:
> Headline: "Employees of a new plan to get Dell back on the road to growth,
> including streamlining management and looking at new methods of
> distribution beyond the computer company's direct-selling model."
> Can a baby really learn from THIS ^^^ ?

-----
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=231415&user_secret=fabd7936

Reply via email to