Re: [agi] Parsing theories
This is based purely on reading the wikipedia entry on Operator grammar, which I find very interesting. I'm hoping someone out there knows enough about this to answer some questions :^) Wikipedia says that various quantities are learnable because they can in principle be determined by data. What is known about whether they are efficiently learnable, e.g. (a) whether a child would acquire enough data to learn the language and (b) whether given the data, learning the language would be computationally feasible? (e.g. polynomial time.) Keep in mind that, you have to learn the language well enough to deal with the fact that you can generate and understand (and thus pretty much have to be able to calculate the likelihood of) a virtually infinite number of sentences never before seen. I presume the answer to these two questions (how much data you need and how easy it is to learn from it) will depend on how you parametrize the various knowledge you learn. So, for example, take a word that takes two arguments. One way to parametrize the likelihood of various arguments would be with a table over all two word combinations, the i,j entry gives the likelihood that the ith word and the jth word are the two arguments. But most likely, in reality, the likelihood of the jth word will be much pinned down conditional on the ith. So one might imagine parametrizing these learned coherent selection tables in some powerful way that exposes underlying structure. If you just use lookup tables, I'm guessing learning is computationally trivial, but data requirements are prohibitive. On the other hand, if you posit underlying structure, you can no doubt lower the amount of data required to be able to deal with novel sentences, but I would expect you'd run into the standard problems that finding the optimal structure becomes NP-hard. At this point, a heuristic might or might not suffice, it would be an empirical question. Is there empirical work with this model? Also, I don't see how you can call a model semantic when it makes no reference to the world. The model as described by Wikipedia could have the capability of telling me whether a sentence is natural or highly unlikely, but unless I misunderstand something, there is no possibility it could tell me whether a sentence describes a scene. Matt --- Chuck Esterbrook [EMAIL PROTECTED] wrote: Any opinions on Operator Grammar vs. Link Grammar? http://en.wikipedia.org/wiki/Operator_Grammar http://en.wikipedia.org/wiki/Link_grammar Link Grammar seems to have spawned practical software, but Operator Grammar has some compelling ideas including coherent selection, information content and more. Maybe these ideas are too hard or too ill-defined to implement? Or, in other words, why does Link Grammar win the GoogleFight? Matt http://www.googlefight.com/index.php?lang=en_GBword1=%22link+grammar%22word2=%22operator+grammar%22 (http://tinyurl.com/yvu9xr) Matt Link grammar has a website and online demo at Matt http://www.link.cs.cmu.edu/link/submit-sentence-4.html Matt But as I posted earlier, it gives the same parse for: Matt - I ate pizza with pepperoni. - I ate pizza with a friend. - I Matt ate pizza with a fork. Matt which shows that you can't separate syntax and semantics. Many Matt grammars have this problem. Matt Operator grammar seems to me to be a lot closer to the way Matt natural language actually works. It includes semantics. The Matt basic constraints (dependency, likelihood, and reduction) are Matt all learnable. It might have gotten less attention because its Matt main proponent, Zellig Harris, died in 1992, just before it Matt became feasible to test the grammar in computational models Matt (e.g. perplexity or text compression). Also, none of his Matt publications are online, but you can find reviews of his books Matt at http://www.dmi.columbia.edu/zellig/ Matt -- Matt Mahoney, [EMAIL PROTECTED] Matt - This list is sponsored by AGIRI: Matt http://www.agiri.org/email To unsubscribe or change your Matt options, please go to: Matt 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=231415user_secret=e9e40a7e
Re: [agi] NARS: definition of intelligence
On 5/22/07, Derek Zahn [EMAIL PROTECTED] wrote: Pei, As part of my ongoing AGI education, I am beginning to study NARS in some detail. Thanks for the interest. I'll do my best to help, though since I'm on vacation in China, I may not be able to process my emails as usual. As has been discussed recently here, you define intelligence as: Intelligence is the capability of an information system to adapt to its environment while operating with insufficient knowledge and resources. In later discussion about an adaptive system, you introduce the phrase it attempts to improve its performance in carrying out the tasks. This would seem to be an important further specification. Would it be accurate for my own understanding to rephrase your definition to be: Intelligence is the capability of a task-performing information system to adapt to its environment while operating with insufficient knowledge and resources where task-performing means that the system's purpose is the performance of one or more simultaneously active tasks where a task is defined in terms of a goal state and a (perhaps approximate) method for measuring whether the goal state has been achieved? If goal state is not a good way to describe tasks in the sense you intend, could you explain a little bit about your definitions of carrying out the tasks and improve its performance? Sorry if this seems like a trivial issue, I'm just trying to understand as clearly as possible how you define the goals for the NARS project. It is not a trivial point at all, though I haven't had the pressure (until now) to explain this aspect of my definition publicly. I mostly agree with your description, though rather not to modify the definition in that way, because to me task performing and goal achieving have been mostly implied by the notion of information system, so your description sounds redundent to me. I touched this issue in my first book, though plan to reserve it for my other book, which will be less technical and more philosophical. A few people on this list who was associated with Webmind Inc. should had browsed my extended abstract years ago. The relevent part of that book is to put intelligent system into a larger picture, within a hierachy roughly like the following: 1. system: things/events that should be analyzed as interalating parts, with internal structure and external function 1.1 information system: systems whose structure and function can be analyzed abstractly as goal-achieving (or task-performing), without depending too much on the lower level description (using the terms of physics, chemistry, biology, ...) 1.1.1 intelligent system: information systems that are adaptive and work with insufficient knowledge and reources 1.1.2 instinctive system: information systems that work with sufficient knowledge and resources 1.2 non-information system: systems whose structure and function cannot be analyzed abstractly, and have to be explained in terms of physics, chemistry, biology, ... I know the above description is brief and controversal --- the working definition of information is no less complicated than that of intelligence. Since you asked, I give the above position statement, though I won't argue for it, since it is not that crucial for AGI at the current time. Another topic is goal --- as you noticed, I don't follow the common practice of specifying goal as goal state, because to me this is a big mistake of traditional AI. In the usual sense, a state is indicated by a COMPLETE description of the relevant part of the domain/environment, which cannot be obtained if insufficient knowledge and resources is assumed. Roughly speaking, in NARS a goal is a description, which is a PARTIAL description of the environment. Furthermore, a goal is usualy achieved/satisfied to a degree, which is not a matter of yes/no. Since each goal in NARS is satisfied by a statement, the degree of satisfaction is related to (though not completely reduced to) the truth value of the statement. A more detailed and formal description is in my book, and I'm also working on a paper focusing on this aspect of the system. I'll post a draft when it is finished. Pei - 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=231415user_secret=e9e40a7e
Re: [agi] Pure reason is a disease.
Mark Waser wrote: AGIs (at least those that could run on current computers) cannot really get excited about anything. It's like when you represent the pain intensity with a number. No matter how high the number goes, it doesn't really hurt. Real feelings - that's the key difference between us and them and the reason why they cannot figure out on their own that they would rather do something else than what they were asked to do. So what's the difference in your hardware that makes you have real pain and real feelings? Are you *absolutely positive* that real pain and real feelings aren't an emergent phenomenon of sufficiently complicated and complex feedback loops? Are you *really sure* that a sufficiently sophisticated AGI won't experience pain? I think that I can guarantee (as in, I'd be willing to bet a pretty large sum of money) that a sufficiently sophisticated AGI will act as if it experiences pain . . . . and if it acts that way, maybe we should just assume that it is true. Jiri, I agree with Mark's comments here, but would add that I think we can do more than just take a hands-off Turing attitude to such things as pain: I believe that we can understand why a system built in the right kind of way *must* experience feelings of exactly the sort we experience. I won't give the whole argument here (I presented it at the Consciousness conference in Tucson last year, but have not yet had time to write it up as a full paper). I think it is a serious mistake for anyone to say that the difference between machines cannot in principle experience real feelings. Sure, if they are too simple they will not, but all of our discussions, on this list, are not about those kinds of too-simple systems. Having said that: there are some conventional approaches to AI that are so crippled that I don't think they will ever become AGI, let alone have feelings. If you were criticizing those specifically, rather than just AGI in general, I'm on your side! :-; Richard Loosemore - 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=231415user_secret=e9e40a7e
Re: [agi] Pure reason is a disease.
Hi, On 5/23/07, Mark Waser [EMAIL PROTECTED] wrote: - Original Message - From: Jiri Jelinek [EMAIL PROTECTED] On 5/20/07, Mark Waser [EMAIL PROTECTED] wrote: - Original Message - From: Jiri Jelinek [EMAIL PROTECTED] On 5/16/07, Mark Waser [EMAIL PROTECTED] wrote: - Original Message - From: Jiri Jelinek [EMAIL PROTECTED] Mark and Jiri, I beg you, could you PLEASE stop top-posting? I guess it is just a second for you to cut it, or even better, to change the settings of your mail program to cut it, and it takes a second for every message you send for everyone who reads it to scroll through it, not to mention looking inside for content just in case it was not entirely top-posted. Please, cut it! - lk - 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=231415user_secret=e9e40a7e
Re: [agi] Parsing theories
Check bigrams (or, more interestingly, trigrams) in computational linguistics. Department of Information Systems Email: [EMAIL PROTECTED] Phone: (+27)-(0)21-6504256 Fax: (+27)-(0)21-6502280 Office: Leslie Commerce 4.21 Eric Baum [EMAIL PROTECTED] 2007/05/23 15:36:20 One way to parametrize the likelihood of various arguments would be with a table over all two word combinations, the i,j entry gives the likelihood that the ith word and the jth word are the two arguments. But most likely, in reality, the likelihood of the jth word will be much pinned down conditional on the ith. Is there empirical work with this model? - 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=231415user_secret=e9e40a7e
Re: [agi] Write a doctoral dissertation, trigger a Singularity
Universal compassion and tolerance are the ultimate consequences of enlightenment which one Matt on the list equated IMHO erroneously to high-orbit intelligence methinx subtle humour is a much better proxy for intelligence Jean-Paul member of the 'let Murray stay' advocacy group aka 'the write 2 doctorates, trigger 2 singularities movement' just back from 2 weeks enlightenment-seeking in Indian ashram ;-) Department of Information Systems Email: [EMAIL PROTECTED] Phone: (+27)-(0)21-6504256 Fax: (+27)-(0)21-6502280 Office: Leslie Commerce 4.21 Benjamin Goertzel [EMAIL PROTECTED] 2007/05/20 20:38:35 Personally, I find many of his posts highly entertaining... If your sense of humor differs, you can always use the DEL key ;-) -- Ben G On 5/20/07, Eliezer S. Yudkowsky [EMAIL PROTECTED] wrote: Why is Murray allowed to remain on this mailing list, anyway? As a warning to others? The others don't appear to be taking the hint. - 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=231415user_secret=e9e40a7e
Re: [agi] Parsing theories
A google search on operator grammar + trigram yields nada. A google search on operator grammar + bigram yields nothing interesting. I've seen papers on statistical language parsing before, including trigrams etc. Not so clear to me the extent to which they've been merged with Harris's work. Jean-Paul Check bigrams (or, more interestingly, trigrams) in Jean-Paul computational linguistics. Jean-Paul Department of Information Systems Email: Jean-Paul [EMAIL PROTECTED] Phone: (+27)-(0)21-6504256 Jean-Paul Fax: (+27)-(0)21-6502280 Office: Leslie Commerce 4.21 Eric Baum [EMAIL PROTECTED] 2007/05/23 15:36:20 Jean-Paul One way to parametrize the likelihood of various arguments Jean-Paul would be with a table over all two word combinations, the Jean-Paul i,j entry gives the likelihood that the ith word and the Jean-Paul jth word are the two arguments. But most likely, in Jean-Paul reality, the likelihood of the jth word will be much pinned Jean-Paul down conditional on the ith. Jean-Paul Is there empirical work with this model? - 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=231415user_secret=e9e40a7e
Re: [agi] Parsing theories
I'll take a shot at answering some of your questions as someone who has done some work and research but is certainly not claiming to be an expert . . . . Wikipedia says that various quantities are learnable because they can in principle be determined by data. What is known about whether they are efficiently learnable, e.g. (a) whether a child would acquire enough data to learn the language and (b) whether given the data, learning the language would be computationally feasible? (e.g. polynomial time.) Operator grammar in many respects reminds me of conceptual classification systems in that there has been success in processing huge amounts (corpuses, corpi? :-) of data and producing results -- but it's *clearly* not the way in which humans (i.e. human children) do it. My belief is that if you had the proper structure-building learning algorithms that your operator grammar system would simply (re-)discover the basic parts of speech and would then successfully proceed from there. I suspect that doing so is probably even computationally feasible (particularly if you accidentally bias it -- which would be *really* tough to avoid). All human languages fundamentally have the same basic parts of speech. I believe that operator grammar is reinventing the wheel in terms of it's unnecessary generalization of dependency. Is there empirical work with this model? It depends upon what you mean. My current project is positing an underlying structure of the basic parts of speech. Does it count -- or would I need to (IMO foolishly ;-) discard that for it to count? Also, I don't see how you can call a model semantic when it makes no reference to the world. Ah, but this is where it gets tricky. While the model makes no reference to the world, it is certainly influenced by the fact that 100% of it's data comes from the world -- which then forces the model to build itself based upon the world (i.e. effectively, it is building a world model) -- and I would certainly call that semantics. natural or highly unlikely, but unless I misunderstand something, there is no possibility it could tell me whether a sentence describes a scene. Do you mean that it couldn't perform sensory fusion or that it can't recognize meaning? I would agree with the former but (as an opinion -- because I can't definitively prove it) disagree with the latter. Mark - Original Message - From: Eric Baum [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, May 23, 2007 9:36 AM Subject: Re: [agi] Parsing theories This is based purely on reading the wikipedia entry on Operator grammar, which I find very interesting. I'm hoping someone out there knows enough about this to answer some questions :^) Wikipedia says that various quantities are learnable because they can in principle be determined by data. What is known about whether they are efficiently learnable, e.g. (a) whether a child would acquire enough data to learn the language and (b) whether given the data, learning the language would be computationally feasible? (e.g. polynomial time.) Keep in mind that, you have to learn the language well enough to deal with the fact that you can generate and understand (and thus pretty much have to be able to calculate the likelihood of) a virtually infinite number of sentences never before seen. I presume the answer to these two questions (how much data you need and how easy it is to learn from it) will depend on how you parametrize the various knowledge you learn. So, for example, take a word that takes two arguments. One way to parametrize the likelihood of various arguments would be with a table over all two word combinations, the i,j entry gives the likelihood that the ith word and the jth word are the two arguments. But most likely, in reality, the likelihood of the jth word will be much pinned down conditional on the ith. So one might imagine parametrizing these learned coherent selection tables in some powerful way that exposes underlying structure. If you just use lookup tables, I'm guessing learning is computationally trivial, but data requirements are prohibitive. On the other hand, if you posit underlying structure, you can no doubt lower the amount of data required to be able to deal with novel sentences, but I would expect you'd run into the standard problems that finding the optimal structure becomes NP-hard. At this point, a heuristic might or might not suffice, it would be an empirical question. Is there empirical work with this model? Also, I don't see how you can call a model semantic when it makes no reference to the world. The model as described by Wikipedia could have the capability of telling me whether a sentence is natural or highly unlikely, but unless I misunderstand something, there is no possibility it could tell me whether a sentence describes a scene. Matt --- Chuck Esterbrook [EMAIL PROTECTED] wrote: Any opinions on Operator Grammar vs. Link Grammar?
Re: [agi] Pure reason is a disease.
A meta-question here with some prefatory information . . . . The reason why I top-post (and when I do so, I *never* put content inside) is because I frequently find it *really* convenient to have the entire text of the previous message or two (no more) immediately available for reference. On the other hand, I, too, find top-posting annoying whenever I'm reading a list as a digest but feel that it is offset by it's usefulness. That being said, I am more than willing to stop top-posting if even a sizeable minority find it frustrating (I've seen this meta-discussion on several other lists and seen it go about 50/50 with a very slight edge for allowing top-posting with a skew towards low-volume lists liking it and high-volume lists not). Mark - 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=231415user_secret=e9e40a7e
RE: [agi] NARS: definition of intelligence
Pei Wang writes: Thanks for the interest. I'll do my best to help, though since I'm on vacation in China, I may not be able to process my emails as usual. Thank you for your response. I'm planning over the course of the rest of the year to look in-depth at all of the AGI projects that include a significant implementation component (that is, those that are not just books musing about the nature of intelligence -- I am also reading those in parallel but there are so many that I don't know if anybody could have a solid understanding of all of them). NARS is very well described so it's a good one to start with. I am working from your book Rigid Flexibility which I assume is the best source. I'm sorry that I wasn't able to justify the high cost of buying it new; I got it used from a vendor affiliated with amazon.com. Unless I hit some fundamental roadblock I can easily wait to ask any questions (I don't want to pick nits or ask dumb things anyway) until you're back from your vacation. One thing I'm curious about: peeking ahead, the book sketches a rather long string of increasingly-ambitious implementation stages (if I remember correctly, up to NAL-8). What stage is the current implementation? Thanks again! As has been discussed recently here, you define intelligence as: Intelligence is the capability of an information system to adapt to its environment while operating with insufficient knowledge and resources. In later discussion about an adaptive system, you introduce the phrase it attempts to improve its performance in carrying out the tasks. This would seem to be an important further specification. Would it be accurate for my own understanding to rephrase your definition to be: Intelligence is the capability of a task-performing information system to adapt to its environment while operating with insufficient knowledge and resources where task-performing means that the system's purpose is the performance of one or more simultaneously active tasks where a task is defined in terms of a goal state and a (perhaps approximate) method for measuring whether the goal state has been achieved? If goal state is not a good way to describe tasks in the sense you intend, could you explain a little bit about your definitions of carrying out the tasks and improve its performance? Sorry if this seems like a trivial issue, I'm just trying to understand as clearly as possible how you define the goals for the NARS project. It is not a trivial point at all, though I haven't had the pressure (until now) to explain this aspect of my definition publicly. I mostly agree with your description, though rather not to modify the definition in that way, because to me task performing and goal achieving have been mostly implied by the notion of information system, so your description sounds redundent to me. I touched this issue in my first book, though plan to reserve it for my other book, which will be less technical and more philosophical. A few people on this list who was associated with Webmind Inc. should had browsed my extended abstract years ago. The relevent part of that book is to put intelligent system into a larger picture, within a hierachy roughly like the following: 1. system: things/events that should be analyzed as interalating parts, with internal structure and external function 1.1 information system: systems whose structure and function can be analyzed abstractly as goal-achieving (or task-performing), without depending too much on the lower level description (using the terms of physics, chemistry, biology, ...) 1.1.1 intelligent system: information systems that are adaptive and work with insufficient knowledge and reources 1.1.2 instinctive system: information systems that work with sufficient knowledge and resources 1.2 non-information system: systems whose structure and function cannot be analyzed abstractly, and have to be explained in terms of physics, chemistry, biology, ... I know the above description is brief and controversal --- the working definition of information is no less complicated than that of intelligence. Since you asked, I give the above position statement, though I won't argue for it, since it is not that crucial for AGI at the current time. Another topic is goal --- as you noticed, I don't follow the common practice of specifying goal as goal state, because to me this is a big mistake of traditional AI. In the usual sense, a state is indicated by a COMPLETE description of the relevant part of the domain/environment, which cannot be obtained if insufficient knowledge and resources is assumed. Roughly speaking, in NARS a goal is a description, which is a PARTIAL description of the environment. Furthermore, a goal is usualy
Re: [agi] Write a doctoral dissertation, trigger a Singularity
The scholar and gentleman Jean-Paul Van Belle wrote: Universal compassion and tolerance are the ultimate consequences of enlightenment which one Matt on the list equated IMHO erroneously to high-orbit intelligence methinx subtle humour is a much better proxy for intelligence Jean-Paul member of the 'let Murray stay' advocacy group aka 'the write 2 doctorates, trigger 2 singularities movement' just back from 2 weeks enlightenment-seeking in Indian ashram ;-) Satyan eva jayate -- Sanskrit for Truth alone prevails -- quoted by Mahatma Mohandas Karamchand Ghandi, who also said, First they laugh at you, then they fear you, then they fight you, then you win. By way of explanation... The original message of this thread is also at http://mentifex.virtualentity.com/edcohelp.html as a kind of staging area for AGI Help Wanted appeals from the SourceForge AI Mind project. Now v.t.y. Mentifex here is preparing to ask for Russian and German translations of AI docs. Members of this liberal, all-ideas-welcome list may enjoy some of the Everything2 links below. http://www.everything2.com/index.pl?node_id=1013306 AI should be our top priority http://www.everything2.com/index.pl?node_id=1043865 AI virus http://www.everything2.com/index.pl?node_id=563003 aspects of American society that may be new to you http://www.everything2.com/index.pl?node_id=1228930 the birth of artificial intelligence http://www.everything2.com/index.pl?node_id=11298 But who codes the coders? http://www.everything2.com/index.pl?node_id=452676 butterfly effect http://www.everything2.com/index.pl?node_id=51480 coding standards http://www.everything2.com/index.pl?node_id=134452 Cogito ergo sum http://www.everything2.com/index.pl?node_id=12718 Dark Side of the Moon http://www.everything2.com/index.pl?node_id=32693 Dr. Strangelove, or How I Learned to Stop Worrying and Love the Bomb http://www.everything2.com/index.pl?node_id=774320 +* Excuse me, may I blow your mind? [the failure of Mentifex is not] http://www.everything2.com/index.pl?node_id=1521490 the failure of artificial intelligence http://www.everything2.com/index.pl?node_id=938762 From now on, any ordinary knowledge is no longer going to satisfy you, I'm afraid http://www.everything2.com/index.pl?node_id=76962 futurism http://www.everything2.com/index.pl?node_id=55865 futurist http://www.everything2.com/index.pl?node_id=40987 FWIW http://www.everything2.com/index.pl?node_id=624119 * Geeks of the world unite http://www.everything2.com/index.pl?node_id=472395 hack reality http://www.everything2.com/index.pl?node_id=445357 + A Heartbreaking Work of Staggering Genius http://www.everything2.com/index.pl?node_id=965284 A Heartbreaking Work of Staggering Hubris http://www.everything2.com/index.pl?node_id=426116 * I am not a hacker http://www.everything2.com/index.pl?node_id=675507 I Am Not a Lawyer http://www.everything2.com/index.pl?node_id=113825 I am not making this up http://www.everything2.com/index.pl?node_id=494930 I can't decide whether to change the world or just become a bitter recluse http://www.everything2.com/index.pl?node_id=559881 ** I just bought real estate in your mind http://www.everything2.com/index.pl?node_id=670519 ** I refute him thus! http://www.everything2.com/index.pl?node_id=584208 * I speak for the Borg http://www.everything2.com/index.pl?node_id=870562 I'm at a programming roadblock http://www.everything2.com/index.pl?node_id=1336607 I'm sorry Dave, I'm afraid I can't do that http://www.everything2.com/index.pl?node_id=1188429 in defense of robot domination http://www.everything2.com/index.pl?node_id=19005 Information wants to be free http://www.everything2.com/index.pl?node_id=606019 Information War is coming: whose side are you on? http://www.everything2.com/index.pl?node_id=745413 Is development in AI bad? http://www.everything2.com/index.pl?node_id=73157 Let's Play Global Thermonuclear War One of the http://www.everything2.com/index.pl?node_id=1522443 limitations on artificial intelligence is that True AI needs to be translated into more languages. http://www.everything2.com/index.pl?node_id=61306 The Matrix http://www.everything2.com/index.pl?node_id=525319 The Matrix is going down for a reboot in 5 minutes: all users, please save your data and log out http://www.everything2.com/index.pl?node_id=111373 meatspace http://www.everything2.com/index.pl?node_id=12366 meme http://www.everything2.com/index.pl?node_id=1401073 meme hijack http://www.everything2.com/index.pl?node_id=48303 mission statement http://www.everything2.com/index.pl?node_id=121864 MIT Artificial Intelligence Lab http://www.everything2.com/index.pl?node_id=36338 noosphere http://www.everything2.com/index.pl?node_id=177121 Omega http://www.everything2.com/index.pl?node_id=523623 Omega Point http://www.everything2.com/index.pl?node_id=877088 only in America http://www.everything2.com/index.pl?node_id=45103 otaku
Re: [agi] Parsing theories
As I think about it, one problem is, depending on how its parametrized, its not going to build much of a world model. Say for example it uses trigrams. The average hs grad knows something like 50,000 words. So there are something like 10^17 trigrams. It will never see enough data to build a model capturing much semantics, unless it builds an incredibly compact model, in which case-- what is the underlying structure and how (computationally) are you going to learn it? Absolutely correct. That's why I said My belief is that if you had the proper structure-building learning algorithms that your operator grammar system would simply (re-)discover the basic parts of speech and would then successfully proceed from there. and why I slammed it for reinventing the wheel in terms of it's unnecessary generalization of dependency In unsupervised learning, you can learn a lot, say you can cluster the world into two clusters. But until you get supervision, you can't learn the final few bits to distinguish good from bad, or whatever. I'm afraid that I disagree completely with the latter sentence. Operator grammar might be very useful for getting a structure that could then be rapidly trained to produce meaning, but I don't think you can finish the job until you interact with sensation. It seems as if you're now talking sensory fusion (which is a whole 'nother can o' worms). Mark - 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=231415user_secret=e9e40a7e
Re: [agi] Parsing theories
On 5/23/07, Mark Waser [EMAIL PROTECTED] wrote: systems in that there has been success in processing huge amounts (corpuses, corpi? :-) of data and producing results -- but it's *clearly* not the way corpora - 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=231415user_secret=e9e40a7e
Re: [agi] Pure reason is a disease.
Richard Mark Waser wrote: AGIs (at least those that could run on current computers) cannot really get excited about anything. It's like when you Richard represent the pain intensity with a number. No matter how high the number Richard goes, it doesn't really hurt. Real feelings - that's the key difference between us and them and the reason why they cannot figure out on Richard their own that they would rather do something else than what they were Richard asked to do. So what's the difference in your hardware that makes you have real pain and real feelings? Are you *absolutely positive* that real pain and real feelings aren't an emergent phenomenon of sufficiently complicated and complex feedback loops? Are you *really sure* that a sufficiently sophisticated AGI won't experience pain? I think that I can guarantee (as in, I'd be willing to bet a pretty large sum of money) that a sufficiently sophisticated AGI will act as if it experiences pain . . . . and if it acts that way, maybe we should just assume that it is true. Richard Jiri, Richard I agree with Mark's comments here, but would add that I think Richard we can do more than just take a hands-off Turing attitude to Richard such things as pain: I believe that we can understand why a Richard system built in the right kind of way *must* experience Richard feelings of exactly the sort we experience. Richard I won't give the whole argument here (I presented it at the Richard Consciousness conference in Tucson last year, but have not Richard yet had time to write it up as a full paper). What is Thought? argues the same thing (Chapter 14). I'd be curious to see if your argument is different. Richard I think it is a serious mistake for anyone to say that the Richard difference between machines cannot in principle experience Richard real feelings. Sure, if they are too simple they will not, Richard but all of our discussions, on this list, are not about those Richard kinds of too-simple systems. Richard Having said that: there are some conventional approaches to Richard AI that are so crippled that I don't think they will ever Richard become AGI, let alone have feelings. If you were criticizing Richard those specifically, rather than just AGI in general, I'm on Richard your side! :-; Richard Richard Loosemore Richard - This list is sponsored by AGIRI: Richard http://www.agiri.org/email To unsubscribe or change your Richard options, please go to: Richard 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=231415user_secret=e9e40a7e
Re: [agi] Pure reason is a disease.
AGIs (at least those that could run on current computers) cannot really get excited about anything. It's like when you represent the pain intensity with a number. No matter how high the number goes, it doesn't really hurt. Real feelings - that's the key difference between us and them and the reason why they cannot figure out on their own that they would rather do something else than what they were asked to do. Mark So what's the difference in your hardware that makes you have Mark real pain and real feelings? Are you *absolutely positive* that Mark real pain and real feelings aren't an emergent phenomenon of Mark sufficiently complicated and complex feedback loops? Are you Mark *really sure* that a sufficiently sophisticated AGI won't Mark experience pain? Mark I think that I can guarantee (as in, I'd be willing to bet a Mark pretty large sum of money) that a sufficiently sophisticated AGI Mark will act as if it experiences pain . . . . and if it acts that Mark way, maybe we should just assume that it is true. If you accept the proposition (for which Turing gave compelling arguments) that a computer with the right program could simulate the workings of your brain in detail, then it follows that your feelings are identifiable with some aspect or portion of the computation. I claim that if feelings are identified with the decision making computations of a top level module, (which might reasonably be called a homunculus) everything is concisely explained. What you are then *unaware* of is all the many and varied computations done in subroutines that the decision making module is isolated from by abstraction boundary (this is by far most of the computation) as well as most internal computations of the decision making module itself (which it will no more be programmed to be able to report than my laptop can report its internal transistor voltages). What you feel and can report and the qualitative nature of your sensations is then determined by the code being run as it makes decisions. I claim that the subjective nature of every feeling is very naturally explained in this context. Pain, for example, is the weighing of programmed-in negative reinforcement. (How could you possibly modify the sensation of pain to make it any clearer it is negative reinforcement?) What is Thought? ch 14 goes through about 10 sensations that a philosopher had claimed were not plausibly explainable by a computational model, and argues that each has exactly the nature you'd expect evolution to program in. You then can't have a zombie that behaves the way you do but doesn't have sensations, since to behave like you do it has to make decisions, and it is in fact the decision making computation that is identified with sensation. (Computations that are better preprogrammed because they don't require decision, such as pulling away from a hot stove or driving the usual route home for the thousandth time, are dispatched to subroutines and are unconscious.) This picture is subject to empirical test, through psychophysics (and also as we increasingly understand the genetic programming that builds much of this code.) A good example is Ramanchandran's amputee experiment. Amputees frequently feel pain in their phantom (missing) limb. They can feel themselves clenching their phantom hand so hard, that their phantom finger nails gouge their phantom hands, causing intense real pain. Ramanchandran predicted that this was caused by the mind sending a signal to the phantom hand saying: relax, but getting no feedback assuming that the hand had not relaxed, and inferring that pain should be felt (including computing details of its nature). He predicted that if he provided a feedback telling the mind that relaxation had occurred the pain would go away, which he then provided through a mirror device in which patients could place both real and phantom limbs, relax both simultaneously, and get visual feedback that the phantom limb had relaxed (in the mirror). Instantly the pain vanished, confirming the prediction that the pain was purely computational. - 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=231415user_secret=e9e40a7e
Re: [agi] Pure reason is a disease.
Mike Eric Baum: What is Thought [claims that] feelings.are Mike explainable by a computational model. Mike Feelings/ emotions are generated by the brain's computations, Mike certainly. But they are physical/ body events. Does your Turing Mike machine have a body other than that of some kind of computer Mike box? And does it want to dance when it hears emotionally Mike stimulating music? Mike And does your Turing Machine also find it hard to feel - get in Mike touch with - feelings/ emotions? Will it like humans massively Mike overconsume every substance in order to get rid of unpleasant Mike emotions? If its running the right code. If you find that hard to understand, its because your understanding mechanism has certain properties, and one of them is that it has having trouble with this concept. I claim its not surprising either that evolution programmed in an understanding mechanism like that, but I suggest it is possible to overcome in the same way that physicists were capable of coming to understand quantum mechanics. - 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=231415user_secret=e9e40a7e
Re: [agi] NARS: definition of intelligence
On 5/23/07, Derek Zahn [EMAIL PROTECTED] wrote: I'm planning over the course of the rest of the year to look in-depth at all of the AGI projects that include a significant implementation component (that is, those that are not just books musing about the nature of intelligence -- I am also reading those in parallel but there are so many that I don't know if anybody could have a solid understanding of all of them). I'm working on an introduction of AGI projects, which may help people like you. I'll post it when I'm back in late June. NARS is very well described so it's a good one to start with. I am working from your book Rigid Flexibility which I assume is the best source. I'm sorry that I wasn't able to justify the high cost of buying it new; I got it used from a vendor affiliated with amazon.com. Yes, the book is the best source for most of the topics. Sorry for the absurd price, which I have no way to influence. One thing I'm curious about: peeking ahead, the book sketches a rather long string of increasingly-ambitious implementation stages (if I remember correctly, up to NAL-8). What stage is the current implementation? The book corresponds to NARS 4.3.0, which implements the basic of NAL-8 (the final layer of the NAL family). The current on-line demo is 4.3.1, and I'm coding 4.3.2. I consider NAL-1 to NAL-6 to be relatively mature, while I'm still adding details into NAL-7 and NAL-8. For the overall engineering plan, see http://nars.wang.googlepages.com/wang.roadmap.pdf , which is more clear and up-to-date than the book on this topic. Pei - 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=231415user_secret=e9e40a7e
[agi] Computer explains your error by showing how it should have been done
For those of you interested in type-driven program synthesis: http://www.cs.washington.edu/homes/blerner/seminal.html (quick link: http://www.cs.washington.edu/homes/blerner/files/seminal-visitdays.ppt) - 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=231415user_secret=e9e40a7e
Re: [agi] Pure reason is a disease.
P.S. Eric, I haven't forgotten your question to me, will try to address it in time - the answer is complex. - 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=231415user_secret=e9e40a7e
Re: [agi] Pure reason is a disease.
Eric, The point is simply that you can only fully simulate emotions with a body as well as a brain. And emotions while identified by the conscious brain are felt with the body I don't find it at all hard to understand - I fully agree - that emotions are generated as a result of computations in the brain. I agree with cog. sci. that they are highly functional in helping us achieve goals. My underlying argument, though, is that your (or any) computational model of emotions, if it does not also include a body, will be fundamentally flawed both physically AND computationally. Mike Eric Baum: What is Thought [claims that] feelings.are Mike explainable by a computational model. Mike Feelings/ emotions are generated by the brain's computations, Mike certainly. But they are physical/ body events. Does your Turing Mike machine have a body other than that of some kind of computer Mike box? And does it want to dance when it hears emotionally Mike stimulating music? Mike And does your Turing Machine also find it hard to feel - get in Mike touch with - feelings/ emotions? Will it like humans massively Mike overconsume every substance in order to get rid of unpleasant Mike emotions? If its running the right code. If you find that hard to understand, its because your understanding mechanism has certain properties, and one of them is that it has having trouble with this concept. I claim its not surprising either that evolution programmed in an understanding mechanism like that, but I suggest it is possible to overcome in the same way that physicists were capable of coming to understand quantum mechanics. - 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/?; -- No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.467 / Virus Database: 269.7.6/815 - Release Date: 22/05/2007 15:49 - 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=231415user_secret=e9e40a7e
Re: [agi] NARS: definition of intelligence
Pei, Yes, the book is the best source for most of the topics. Sorry for the absurd price, which I have no way to influence. It's $190. Somebody is making a lot of money on each copy and I'm sure it's not you. To get a 400 page hard cover published at lulu.com is more like $25. Shane - 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=231415user_secret=e9e40a7e
Re: [agi] NARS: definition of intelligence
Shane, Well, I actually considered Lulu and similar publishers, though as the last option. It is much easier to publish with them, but given the nature of NARS, such a publisher will make the book even more likely to be classified as by a crackpot. :( I continued to look for a publisher with tough peer-review procedure, even after the manuscript had been rejected by more than a dozen of them. Though the price excludes most of individual buyers, it may be more likely for a research library to buy a $190 book from Springer than a $25 book from Lulu, given the topic. Pei On 5/24/07, Shane Legg [EMAIL PROTECTED] wrote: Pei, Yes, the book is the best source for most of the topics. Sorry for the absurd price, which I have no way to influence. It's $190. Somebody is making a lot of money on each copy and I'm sure it's not you. To get a 400 page hard cover published at lulu.com is more like $25. Shane 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=231415user_secret=e9e40a7e
Re: [agi] Pure reason is a disease.
On Wednesday 23 May 2007 06:34:29 pm Mike Tintner wrote: My underlying argument, though, is that your (or any) computational model of emotions, if it does not also include a body, will be fundamentally flawed both physically AND computationally. Does everyone here know what an ICE is in the EE sense? (In-Circuit Emulator -- it's a gadget that plugs into a circuit and simulates a given chip, but has all sorts of debugging readouts on the back end that allow the engineer to figure out why it's screwing up.) Now pretend that there is a body and a brain and we have removed the brain and plugged in a BrainICE instead. There's this fat cable running from the body to the ICE (just as there is in electronic debugging) that carries all the signals that the brain would be getting from the body. Most of the cable's bandwidth is external sensation (and indeed most of that is vision). Motor control is most of the outgoing bandwidth. There is some extra portion of the bandwidth that can be counted as internal affective signals. (These are very real -- the body takes part in quite a few feedback loops with such mechanisms as hormone release and its attendant physiological effects.) Let us call these internal feedback loop closure mechanisms the affect effect. Now here is * Hall's Conjecture: The computational resources necessary to simulate the affect effect are less than 1% of that necessary to implement the computational mechanism of the brain. * I think that people have this notion that because emotions are so unignorable and compelling subjectively, that they must be complex. In fact the body's contribution, in an information theoretic sense, is tiny -- I'm sure I way overestimate it with the 1%. 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=231415user_secret=e9e40a7e
Re: [agi] NARS: definition of intelligence
On May 23, 2007, at 4:17 PM, Pei Wang wrote: I continued to look for a publisher with tough peer-review procedure, even after the manuscript had been rejected by more than a dozen of them. Though the price excludes most of individual buyers, it may be more likely for a research library to buy a $190 book from Springer than a $25 book from Lulu, given the topic. Most books of this type are priced so that it will turn a profit on library sales alone. There are hundreds of libraries that will buy a copy of every single book published by a major publisher in a given publishing program regardless of either price or specific content. Because this is the business model, sales to individuals are not even a relevant consideration -- individual sales are pure gravy to the publisher. Given that, the economics of the pricing becomes obvious: price = (production cost * 1.2) / library buyers or something like that. What the market for individual sales will bear does not even enter the picture. Note that the library buyers will buy or not buy a program based on the quality/strength of an individual program at a publisher, they do not simply buy all the books from every STM publisher for a given topic area. Program quality rather than price is the deciding factor, as is in evidence here. Cheers, J. Andrew Rogers - 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=231415user_secret=e9e40a7e
