Josh, Again a good reply. So it appears the problem is they don't have good automatic learning of semantics.
But, of course, that's vertually impossible to do in small systems except, perhaps, about trivial domains. It becomes much easier in tera-machines. So if my interpretation of what you are saying is true, it bodes well for the ease of overcoming this problem in the coming years with the coming hardware. I look forward to reading Pei's article on this subject. It may shed some new light on my understanding of the subject. But it may take me some time. I read and understand symbolic logic slowly. Edward W. Porter Porter & Associates 24 String Bridge S12 Exeter, NH 03833 (617) 494-1722 Fax (617) 494-1822 [EMAIL PROTECTED] -----Original Message----- From: J Storrs Hall, PhD [mailto:[EMAIL PROTECTED] Sent: Thursday, October 04, 2007 4:30 PM To: agi@v2.listbox.com Subject: Re: [agi] breaking the small hardware mindset Let me answer with an anecdote. I was just in the shop playing with some small robot motors and I needed a punch to remove a pin holding a gearbox onto one of them. I didn't have a purpose-made punch, so I cast around in the toolbox until Aha! an object close enough to use. (It was a small rattail file) Now the file and a true punch have many things in common and many other things different. Among the common things that were critical are the fact that the hardened steel of the file wouldn't bend and wedge beside the pin, and I could hammer on the other end of it. These semantic aspects of the file had to match the same ones of the punch before I could see it as one. Where did these semantic aspects come from? Somehow I've learned enough about punches and files to know what a punch needs (i.e. which of its properties are necessary for it to work) and what a file gives. In Copycat, the idea is to build up an interpretation of an object (analogy as perception) under pressures from what it has to match. So far, well and good -- that's what I was doing. But in Copycat (and tabletop and ...) the semantics is built in and ad hoc. And there isn't really all that much of an analogy net matching algorithm without the semantics (codelets). In my case, I have lots of experience misusing tools, so I have built up an internal theory of which properties are likely to matter and which aren't. I think this most closely matches your even-numbered points below :-) Perhaps more succinctly, they have a general purpose representation but it's snippets of hand-written lisp code, and no way to automatically generate more like it. Josh On Thursday 04 October 2007 02:59:38 pm, Edward W. Porter wrote: > Josh, > > (Talking of breaking the small hardware mindset, thank god for the > company with the largest hardware mindset -- or at least the largest > physical embodiment of one-- Google. Without them I wouldnt have > known what FARG meant, and would have had to either (1) read your > valuable response with less than the understanding it deserves or (2) > embarrassed myself by admitting ignorance and asking for a > clarification.) > > With regard to your answer, copied below, I thought the answer would > be something like that. > > So which of the below types of representational problems are the > reasons why their basic approach is not automatically extendable? > > 1. They have no general purpose representation that can represent > almost anything in a sufficiently uniform representational scheme to > let their analogy net matching algorithm be universally applied > without requiring custom patches for each new type of thing to be > represented. > > 2. They have no general purpose mechanism for determining what are > relevant similarities and generalities across which to allow slippage > for purposes of analogy. > > 3. They have no general purpose mechanism for > automatically finding which compositional patterns map to which lower > level representations, and which of those compositional patterns are > similar to each other in a way appropriate for slippages. > > 4. They have no general purpose mechanism for > automatically determining what would be appropriately coordinated > slippages in semantic hyperspace. > > 5. Some reason not listed above. > > I dont know the answer. There is no reason why you should. But if > you > -- or any other interested reader do, or if you have any good thoughts > on the subject, please tell me. > > I may be naïve. I may be overly big-hardware optimistic. But based > on the architecture I have in mind, I think a Novamente-type system, > if it is not already architected to do so, could be modified to handle > all of these problems (except perhaps 5, if there is a 5) and, thus, > provide powerful analogy drawing across virtually all domains. > > Edward W. Porter > Porter & Associates > 24 String Bridge S12 > Exeter, NH 03833 > (617) 494-1722 > Fax (617) 494-1822 > [EMAIL PROTECTED] > ----- 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=50129778-7b4bfa