|posted 1/4/2007 16:53|
I am seeking input on some work I've been doing. Since my day-job is in an unrealted field, I have no opportunity for the type of natural daily professional interaction regarding programming ideas.
I need to find other work that is related; also, as is the case with a person's social sanity, isolation is hardly optimal, and the input of others with similar interests would be appreciated.
It's learning-algorithm/natural-language/knowledge-representation. You may read about it at www2.potsdam.edu/lanzcc/artificialintelligence.htm.
Please respond to my personal email if possible: I don't see these posts every day.
Best wishes, and thanks!
"Ambition is a poor excuse for not having the sense to be lazy."
| the geometry of meaning|
|posted 3/9/2007 20:37|
|Interesting. Good luck to your work, abstraction is key to an AI.|
|posted 4/27/2007 03:59|
|Potsdam, NY? I grew up there!|
Most people's filters require some foreplay, or whatnot, before reading a 25 page paper.
What are the highlights of the thesis?
|posted 5/28/2007 00:44|
|Replying to Andrew (finally)|
Sorry for the delay - no content to it - after I put this message up initially, seemed like forever passed with no replies, so I quit coming to this forum.
I like your wording - filters requiring foreplay - so here we go:
1) I am disinterested in theorizing - I only like empirically demonstrable knowledge. The paper is "planning", rather.
2) Word definitions and program instructions/behaviors all exist in a large feature space of perhaps 500 dimensions. This way of quantifying linguistic definitions has a number of advantages
- morphemes, graphemes, words, phrases, sentences, subjects, are all represented to the program in the same ways, and there are straightforward ways to get from one to the other
- there's no barrier between conversation & program operation
- the space in which the definitions and behaviors exist consists of axes that are themselves defined within the space
- the topology of the space is transformed by conversation
- getting from one sentence to another (translation: generating a reply) is a transform between two entities in the space
- class-words, generalization, and parts-of-speech have straightforward meaning in the space - in fact they are generated
3) the paper describes the 23 years of experiments & programs that lead up to the current ptoject
4) SOme useful AI is done by limiting the subject area and trying to figure out how a) it works in humans or b)how it can be made to work in a computer. Like speech synthesis, or vision.
I am only interested in algorithms that are general - I can speak, sing, and navigate. Any program I write must be equally capable in each area, without modification to anything but the input pathway.
One must crawl before one can walk. Clearly one could write programs that speak incompetently, sing in a limited way, and navigate in a limited environment, rather easily ( I have done so ). Now the question is, how to construct a "single" core operating procedure for all three areas, that is at least as "single" as is the structure of the cortex, and that actually reasons, composes, and, what, I suppose, dances? (The cortex itself is pretty homogeneous across regions of radically differing function. I mean this only to a certain extent - structurally homogeneous - I used to make silver-stain slides of it - not functionally so, after learning has taken place.)
Let's see what else
I don't know, it depends what area you're coming from. A computational-linguist would be amused at how this data-structure makes phrase-isolation possible, and how it generates its own parts-of-speech. A polyglot would be amused at the program's complete equality for working in any language. Don Knuth was amused at the broad & complicated uses of content-addressable memory (although not at much else - 'course that was 20 years ago).
Fans of production-systems and associative learning would be amused at the mechanisms for tabula-rasa startup.
And my mentor, an experimental physicist, would say "shut up about it until all these fantasies have come to fruition" but I have decided to talk about it anyway, for the reasons hinted at in my initial post.
I wonder if I've made it into the net-kook files yet?
|posted 5/29/2007 21:59|
|I thought I had replied to this a while back... ah,well. Anyway, did you write that link yourself?!?!? The content is extremely original, well thought-out, and advanced. Really, there is only one suggestion I could make - perhaps you could formalize the connection you noted between this linguistic model and quantum mechanics. There already seems, to me, to be an elusive sort of wavefunction notion occuring in the space, and tying the connection between the two fields could make your work a lot easier. All the theorems and programs that work for QM would work for your model also. Furthermore, it would be philosophically interesting to find out just what the linguistic manifestation of an atom, force field, and so forth, is. By the way, have you considered using the sort of program you described on a set of mathematical input, perhaps Principia Mathematica or some other formalized system? The connections between mathematics and language is very crucial in all the modern hype about Godel, Turing, and Chaitan, so maybe giving it as input itself could be insightful. However, whatever other connections your concept may have with other fields, your ideas are 100% top-notch. So far, you seem to have done unbelievingly well without the help of an expert, although I certainly understand what it is like to not having anyone to test ideas with. Either way, best of luck, and keep on researching!|
|posted 5/30/2007 19:37|
Nevertheless, this seems as sheer theorizing.
lanzcc wrote @ 5/28/2007 12:44:00 AM:
1) I am disinterested in theorizing - I only like empirically demonstrable knowledge...
Any chances of you BUILDING something that can be DEMONSTRATED?
|posted 6/3/2007 01:43|
|Thank you for your kind words.|
Plato Demosthenes wrote @ 5/29/2007 9:59:00 PM:
>>did you write that link yourself?!?!?
Look, Ma, no hands!
>>perhaps you could formalize the connection you noted between this linguistic model and quantum mechanics.
I am very attracted by this idea, and I have been powerfully aided by the general ideas of QM. I am just beginning to encode the interactions among objects.
Stealing mathematics from un-connected branches of applied science is no problem, but it gets a little kitchy, or cute, to go around saying "oooh Chaos theory! I wonder how I could capitalize on everyone's amusement with strange attractors! Maybe I can look cool if I drop names! So anyway I'm a bit cautious about that.
>>All the theorems and programs that work for QM would work for your model also.
Well, some of it. Not down too far in level, though. I'm not sure what the significance of, oh, I don't know, the fine-structure-constant, would be, to computational linguistics.
>>By the way, have you considered using the sort of program you described on a set of mathematical input, perhaps Principia Mathematica or some other formalized system?
Well, actually, I'm working for a guy, writing a sort of generalized version of this learning algorithm for him to use on other environments. The idea of "general" is so central to my thinking that it's often reasonable to branch out.
|posted 6/3/2007 01:47|
Well, yeah, I thought I had addressed that. I'm building every day. I have done maybe 10% of what I expect to be a 2-year project.
yaki wrote @ 5/30/2007 7:37:00 PM:
Nevertheless, this seems as sheer theorizing.
Any chances of you BUILDING something that can be DEMONSTRATED?
If one just keeps one's head down, buried in the sand of isolated and intense work, one is in danger of becoming ensnared by one's own desires, and one can be victimized by errors that others might see early on, saving one long fruitless treks down dead-end paths.
Since I teach full time in another field, this forum is one of only a few ways for me to communicate in this important way, during the early stages of the project.
|posted 6/6/2007 03:38|
|lanzcc, Ive got a funny question for you. Whats the most general method of understanding what a program does without understanding the programming language being used, and without actually initializing the program itself|
|posted 6/7/2007 19:07|
|I'm sorry, that's the sort of question I can't help you with. I believe a mathematician would just reply "formal logic" or if you agree with Zadeh "formal fuzzy logic".|
When you say something extreme like "the most efficient" you suddenly propel yourself into a realm of engineering and mathematical extremes. It's much wiser just to look for "a good method".
The guy I'm working for always says, of program documentation, "just tell me the inputs & outputs, just tell me the inputs & outputs, I don't care about the rest" so I suppose you might consider that. As one centrally concerned with how things work, I find that to be, well, shallow.
The implication of your post is that you don't know what my work is aimed at, so here's another informal sound-bite:
input: observed human linguistic, musical, and motile behavior
output: computer imitations of a human generating original linguistic, musical, and navigational behavior.
And an inspirational sound-bite:
It's Krazy, the Kat that walks by himself.
If you can think of more specific questions, maybe I can be of more help.
|posted 6/8/2007 19:54|
|Oh I did look through your work, and I think it's very intelligent! As for my question, I apoligize, I just supposed you would have an idea on it, as my own approach is at this time is a bit thrown into the engineering aspect as a starting point. You did give an interesting reply though. :)|
Foods for thought. If your input/output was a taste, then your reply to andrew was a snack, and your paper is a full course meal! The level of detail in it is incredible, and hardly something to be swallowed whole. In an attempt to swallow it quickly, I did choke a bit. Definitely something I would have to grab a pen and paper in order to analyze, cause the attempt to do it all in my head was quite a spin, as I'm hardly familiar with all the concepts.
Yes, the AI Forum is definitely a buffet of foods for thought.
The "context" and the stuff relating to it.....Definitely need a pen and paper for that. I'll try and give an intelligent question on it in a day or two. :p
|posted 6/9/2007 18:03|
I began to understand with Andrew's post above, and LordJakian's reinforces the same idea.
Therefore, I am in the process of re-writing the whole paper as a hypertext document so that you can read all the way through in about 1/3 the time and just click for more details.
You guys, you have to understand, ;) all the detail is included because I wished to be exhaustive and complete. I just had a meeting with a close friend who is a prof of Math here, and we laughed heartily because he observed that I am a "kook". We discussed in detail what that means, how to avoid the negative aspects of it, and the fact that, for instance Fourier was a kook too.
But anyway I am seriously paranoid about the "kook" thing, and one way to lessen kookiness is completeness.
The paper that's online now is 3 months old anyway, and I've written a whole s--tload of code since then, and refined, altered, or enhanced enough that it's worthwhile going at it again. The new version will not be up for another couple of weeks, but I'll come here and "announce" it
( ooooooo ) when it's done.
|posted 6/21/2007 23:49|
|OK. Almost finished, and so anyone who goes to the site now will have an easier time - hopefully interest will be tickled without too great a time investment.|
www2.potsdam.edu/lanzcc/ and don't go to version 1.0
hope this helps
Thanks everybody for the feedback
|posted 6/22/2007 16:34|
|Lancc, look at it this way :)|
When Archimedes ran down the street naked crying Eureka! I solved it!, people probaly thought he was a bit kooky, but I'm sure the king was glad to see him anyway.
My one question so far is this. Is english the only language you have contcentrated on as of yet? I ask simply because I would think it would be interesting comparing different geometries of meaning together and showing where they differ and where they are nearly similar.
With your idea that could be possible, or not?
|posted 6/22/2007 20:36|
|Yes, other languages are plausible. The software is not quite completely immune to language change, however. The parsing tricks for plurals, possessives, verb-forms, and homonyms are English specific, and the extensive compound-word situation in German would require more concentration on the disassemby of words.|
I'm pretty sure, though, that the axes would all be the same, including the axis-defining ones. This means the meaning space would be the same too. Largely differrent labels at the points in the space.
I considered using Latin, with all its inflections that tell you what the function/part-of-speech of all words are, and with its general lack-of-caring about word order as compared to English.
The problem is, I don't know Latin as well, and it's REALLY a lot easier to write one of these program if you're REALLY familiar with the space yourself.
And this is such a hard project, I have pretty much taken every chance to simplify initial attempts.
|posted 1/8/2008 00:36|
|How goes the project?|
|posted 1/12/2008 02:10|
|12,000 lines of code done. 2nd version of on-line paper done.|
What's done is so banal. You know, beaucoup data types & the files, display & management of each one. 2 weeks spent coding the display for short-term-memory trees. Yuk.
It'll be eons before I know - empirically - what this idea can do. Nonetheless it is very beautiful.
I did re-read Minsky, after 20 years. It appears I'm a plagiarist. There's a new appendix in my paper that describes how his ideas are related to my programming structure.
|posted 3/14/2009 21:46|
|How goes the project? |
Lol, I know I'm awakening a long not disturbed thread, but I thought it was interesting.
So, any new thoughts as of yet?
|Last edited by lordjakian @ 3/14/2009 9:47:00 PM|
|posted 10/10/2009 18:34|
|See the new pdfLaTeX version of the paper|
A Geometric Approach to Meaning
now at 26,000 lines of code, and learning CUDA !