On Jun 28, 2004, at 7:08 PM, Brad Wyble wrote:
The more one studies the specifics of the solutions the brain uses, the more one realizes the incredible varieties of strategies that could have been used, yes economically, and yes in our universe.


But what is so extraordinary about this? This is just another expression of the Church-Turing thesis.

No it isn't. The CT thesis talks about implementing many different solutions on one hardware.


I think you're misunderstanding me. I'm not saying that the brain has come up with 100 clever ways to solve the same problem.

I'm saying the brain has had to cherry-pick (ie face/speech recognition) what problems it can afford to
spend the resources to hardwire, and then tack on some general purpose machinery to handle everything else. But even that general purpose machinery is highly specialized in many ways (ie using emotion as a
shortcut to decision making in a cognitively intractable space).


Armed with this insight, one could imagine brains from different evolutionary timelines that solve a completely different set of problems, and end up thinking in completely different ways. Even within the relatively narrow track of the brains created by our genome, there is incredible diversity in the types of thinking that can be done: intuitive, logical, artistic, etc. Twiddle a few base pairs and Mozart becomes Turing or Newton. These people may use the same underlying NAP representation of their knowledge, but by manipulating parameters in the various problem solving apparati that give us our cognitive faculties, they can solve radically different kinds of problem.

In essence:
They each have access to areas of knowledge space that are inaccessible to the other.



Now if your claim is that our first generation of real AGI's will be able to enter all domains of knowledge with equal fluidity, I'm going to have to agree to disagree. We will beat evolution someday, but not on our first try. And I don't think such a thing as a truly general AGI exists. There will always be some degree of specialization when resources are not infinite.




To recycle an analogy, the design space of airplanes is both rich and extremely bounded. The Wright brothers' first flyer and the very advanced F-22 occupy the same narrow design space even though the technology to express that design space has improved dramatically.


Flight is not an open-ended problem.  Knowledge generation is.



Hrmmm.... let me rephrase.

For any theoretical approach that starts from the ground and works up in mathematics, there is very little disagreement. A model that isn't grounded even loosely in basic theoretical principles isn't much of a model at all. Most AI research, hard or weak, has revolved around making adjustments to the epicycles of the intelligence solar system.

In my not so humble opinion, THE historical problem with AI research and research on intelligence in general is that it has been a random walk through the phase space. There is no good mathematical reasons to buy into most of the models proposed because almost all of them amount to throwing things at the wall to see what sticks in implementation. You'll have a hard time finding the solution if you haven't specified what the solution will look like in refined and rigorous terms.

You are preaching to the choir. This paragraph here is exactly why I jumped ship into understanding the brain as a means to understanding how to do AI. Cognitive science is messy, but at least there's a rubrick to follow. I'd rather be a blind man trying to figure out that he's touching an elephant, than a blind man walking through the desert.


But there are a lot of smart people working on AI, and while I do not favor their methodology personally, I can see the value in approaching the AI problem from a completely unguided perspective. After all, evolution came up with *us* by throwing things at a wall.

Now yes, approaches that start from the same underlying set of mathematics may tend to agree on an approach and presumed result, but that's not because there's only one result. It's because they started from the same point, think the same way, and are therefore restricted to the same knowledge space.



No, knowledge is derivative patterns. 1+2=3 is a simple low-order pattern. That pattern has no meaning. When that pattern mixes and interacts with other patterns, you may discover other useful patterns. Knowledge is the derived relationships and associations between patterns (all of which are themselves patterns).


I think you are drawing a false dichotomy when you claim that 1+2=3 is fundamentally different from more complex forms of knowledge. How complex does a mathetical statement have to be before you claim that it is real knowledge?



Nonsense. A chemistry textbook is a collection symbols and patterns. One has to have a not insignificant set of patterns that one already knows to be able to extract useful derivative patterns from the chemistry text, as a chemistry textbook requires that one has already learned a rather vast amount of prior patterns that serve as context.

One *always* has to have a pre-existing set of knowledge to extract value from other knowledge.




Data is the raw information stream. Knowledge is the result of the inductive process on that stream.

You've done a great job of dispelling my concrete examples of knowledge. Can you provide me with some concrete examples of what knowledge is, by your definition? I'm struggling to see the distinction as anything more than a vast gray area.



No, it is dependent on my view that there is only one *efficient* solution to the general problem. Or more accurately, it is a peak in the phase space and that you have to be pretty near that peak to realize AI in practice.


Again, agree to disagree. The space is the Rocky Mountains, and while one of them may be higher than all the others, there are plenty of runner-ups which go high enough. Further, it is in an intractable and fruitless pursuit to find the highest of them.



Heh. This is a narrow AI perspective rather than a General Intelligence perspective.


For simple and extremely well-specified domains, narrow domain implementations will outperform GI, in this you are correct. But for everything else, the ability of a GI to adaptively tailor itself to the patterns of complex domains will generally outperform narrow intelligence.

Any GI worth a damn should be able to thoroughly extract all the patterns in an economic reality from the low-order to the high-order, rather than selectively like you suggest. If it can't do this, it isn't "general".

Here we debate semantics of what GI means. I think your definition of GI is impractical. I think that a more specialized version of a GI (let's call it SSGI: Somewhat Specialized GI) will scoop you years before you approach a functional product, and hence gain all of the bootstrapping benefits you speak of (or 10 companies will divide the pie, either way, you get none of it)


There is much to be gained in specialization. After all, it is the ubiquitous hallmark of the designs explored by evolution, a tireless process which has learned over the past many millions of years one unassailable truth: specialization is supremely efficient.

-Brad

-------
To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Reply via email to