On Thu, Oct 25, 2012 at 5:07 AM, Mike Tintner <[email protected]>wrote:
> Aaron, you haven’t described anything to me. I asked you/all to analyse > graphically how you would recognize a chair – you gave me zip, just general > airy stuff. > > Aaron, This is a perfect example of how your examples and explanations weren't interpreted as examples and explanations. The reason is partly psychological but it is also because you cannot talk about the key issue in a way that is subtle enough to get him to think about what you are saying. For instance, you noted that the combinatorial explosion was a problem and yet his response was, "...you gave me zip, just general airy stuff... Language is an unbelievably complicated thing – and nets don’t even scrape the surface." The mention of combinatorial explosion (or complexity) does not affect his thinking. It never has and until you have a sign that it does, you won't be able to use it. The more intuitive idea that computer programming is only able to use natural language in a simple way might be the basis of some future conversation but again, you won't know until you try it. Jim Bromer On Thu, Oct 25, 2012 at 5:07 AM, Mike Tintner <[email protected]>wrote: > Aaron, you haven’t described anything to me. I asked you/all to analyse > graphically how you would recognize a chair – you gave me zip, just general > airy stuff. > As for uniform blocks, yes your nets are totally general and uniform. For > example, how would your system understand: > I WOULDN’T CALL THAT A TREE, WOULD YOU? > IT’S A CROSS BETWEEN A TREE AND A POLE. > THE SCIENCES AND ARTS FORM A TREE. > NO, IT’S NOT TREE, IT’S FOUR. > Your semantic nets, which seem fairly standard & similar to what others > are doing, are toy stuff with toy results. They are not for the real world > of language and language users. Language is an unbelievably complicated > thing – and nets don’t even scrape the surface. > *From:* Aaron Hosford <[email protected]> > *Sent:* Thursday, October 25, 2012 12:46 AM > *To:* AGI <[email protected]> > *Subject:* Re: [agi] The Fundamental Misunderstanding in AGI [was > Superficiality] > Why do you insist that the "blocks" have to be uniform for semantic nets > to work? I already described to you how non-uniform building blocks > (outlines, connected blobs of color, etc.) could be built up through a > hierarchical semantic net to represent the contents of an arbitrary image, > including parts of objects, objects, and entire scenes at different levels > in the hierarchy. There was no assumption of uniformity there, other than > that the bits & pieces being put together are at most 2D, since all that a > 2D image can provide. > The "3D" that we think we perceive is really just 2 1/2 D, where a flat > image has been labeled with depth information, but we can't see the > backsides of objects. Including that sort of information doesn't > fundamentally change anything, although the extra info should make it even > easier to carve the image up into outlines & blobs & subsequently determine > how to connect them together appropriately. > Besides, the process can generalize from 2D images to full 3D (or higher) > just as easily as it generalizes from 1D sentences, where it's already > working to parse sentences, up to 2D images. All that's necessary for it to > work on an input class of any topology or dimensionality is simply to have > the input carved into small, connectable chunks ("blocks") of *arbitrary* > shape and to provide a "grammar" of how those chunks can be connected to > each other recursively. The only caveat lies not in the ability of semantic > nets to represent the structure of arbitrary data, but in the combinatoric > explosion of ways to put the pieces back together when assembling the > semantic net, and that can be gotten around using shrewd design and/or a > faster processor. > > > On Wed, Oct 24, 2012 at 4:03 AM, Mike Tintner <[email protected]>wrote: > >> Aaron, >> The problem is the same whether you start with parts or the whole of the >> object/chair, such as a silhouette. You’re always dealing with sets of >> multiform [odds and ends] blocks and configurations - and *not* uniform >> blocks, on wh. semantic nets depend – wh. is why they have never begun to >> work for AGI. >> *From:* Aaron Hosford <[email protected]> >> *Sent:* Wednesday, October 24, 2012 2:16 AM >> *To:* AGI <[email protected]> >> *Subject:* Re: [agi] The Fundamental Misunderstanding in AGI [was >> Superficiality] >> Maybe parts of a chair are a bad place to start. Maybe you have to look >> at the whole thing, Mike. Contextual information tells you what a part of >> an object is, not the part by itself. Is that vertical piece of wood the >> leg of a chair or a baseball bat leaned against something? You'll never >> know until you look at the scene it appears in. Is "go" a noun or a verb? >> You'll never know until you put it into a sentence. Have a "go" at it. It's >> the relationships between the elements which is key. That's where the >> pattern lies. Not in any one element looking a certain way. >> You keep taking things out of their natural contexts, and then demanding >> we show how our approaches would recognize them. My parser recognizes "go" >> as both a noun and a verb, depending on use. It builds up a set of trees >> that show how the elements, the words, must be connected into bigger >> elements, the phrases, and finally on to the entirety, the sentence. At >> each step, it weeds out the combinations that make no sense. What we need >> is not the ability to recognize the tops of chairs taken out of context. (I >> wouldn't recognize those blobs if I didn't know what they are already, so >> how is software that's meant to copy my abilities going to do it?) What we >> need is a "parser" for images. Something that identifies the outlines and >> fields that make up an image as the base-level elements, and builds up >> higher level elements out of them, excluding combinations that make no >> sense along the way, until the whole image has been explained in terms of >> the *relationships* between those elements. Only then we can look at how >> the parts fit together to determine an object's shape, and from its shape, >> its characteristics (being self supporting, being artifactual, having >> "sittability", etc.), and from its shape and/or characteristics, its >> identity. >> > ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
