On Tuesday 02 October 2007 05:50:57 pm, Edward W. Porter wrote:
> The below is a good post:

Thank you!
 
> I have one major question for Josh.  You said
> 
> “PRESENT-DAY TECHNIQUES CAN DO MOST OF THE THINGS THAT AN AI NEEDS 
> TO DO,  WITH THE EXCEPTION OF COMING UP WITH NEW REPRESENTATIONS AND
> TECHNIQUES. THAT'S THE SELF-REFERENTIAL KERNEL, THE TAIL-BITING,
> GÖDEL-INVOKING COMPLEX CORE OF THE WHOLE PROBLEM.”
> 
> Could you please elaborate on exactly what the “complex core of the whole
> problem” is that you still think is currently missing.

No, but I will try to elaborate inexactly.   :^)

Let me quote Tom Mitchell, Head of Machine Learning Dept. at CMU:

"It seems the real problem with current AI is that NOBODY to my knowledge is 
seriously trying to design a 'never ending learning' machine."
(Private communication)

By which he meant what we tend to call "RSI" here. I think the "coming up with 
new representations and techniques" part is pretty straightforward, the 
question is how to do it. Search works, a la a GA, if you have billions of 
years and trillions of organisms to work with. I personally am too impatient, 
so I'd like to understand how the human brain does it in billions of seconds 
and 3 pounds of mush.

Another way to understand the problem is to say that all AI learning systems 
to date have been "wind-up toys" -- they could learn stuff in some small 
space of possibilities, and then they ran out of steam. That's what happened 
famously with AM and Eurisko.

I conjecture that this will happen with ANY fixed learning process. That means 
that for RSI, the learning process must not only improve the world model and 
whatnot, but must improve (=> modify) *itself*. Kind of the way civilization 
has (more or less) moved from religion to philosophy to science as the 
methodology of choice for its sages.

That, of course, is self-modifying code -- the dark place in a computer 
scientist's soul where only the Kwisatz Haderach can look.   :^)

> Why for example would a Novamente-type system’s representations and
> techniques not be capable of being self-referential in the manner you seem
> to be implying is both needed and currently missing?

It might -- I think it's close enough to be worth the experiment. BOA/Moses 
does have a self-referential element in the Bayesian analysis of the GA 
population. Will it be enough to invent elliptic function theory and 
zero-knowledge proofs and discover the Krebs cycle and gamma-ray bursts and 
write Finnegan's Wake and Snow Crash? We'll see...
 
Josh

-----
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=49291559-b3bbfd

Reply via email to