Path: csiph.com!v102.xanadu-bbs.net!xanadu-bbs.net!news.glorb.com!npeer01.iad.highwinds-media.com!news.highwinds-media.com!feed-me.highwinds-media.com!post01.iad.highwinds-media.com!newsfe07.iad.POSTED!00000000!not-for-mail From: seeWebInstead@rem.intarweb.org (Robert Maas, http://tinyurl.com/uh3t) Errors-To: ErrorsToHere@YahooGroups.Com X-Spam-This: SpamCopies@YahooGroups.Com X-Twitter: CalRobert Newsgroups: comp.ai.philosophy,comp.programming Subject: Towards true A.I. (was: John McCarthy R.I.P.) References: <738sp4gb4v79.soj02a849hyl$.dlg@40tude.net> <1d202j8607dz7.dgimh8xoxlwo$.dlg@40tude.net> Message-ID: Lines: 223 X-Complaints-To: abuse@rawbandwidth.com NNTP-Posting-Date: Tue, 27 Mar 2012 19:26:58 UTC Date: Tue, 27 Mar 2012 12:25:47 -0700 X-Received-Bytes: 13094 Xref: csiph.com comp.ai.philosophy:3615 comp.programming:1383 DAK> We know how car functions, that is why there exist objective DAK> features which characterize a car. These features are used as DAK> criteria for comparison (for the properties of interest). All DAK> this does not apply to intelligence. REM>... in recent years it's become apparent that the brain is a REM>... loose collection of special-purpose processors, to perform REM>... routine data-processing functions such as visual-feature REM>... extraction and muscle-servo, and specialized but complex REM>... problem-solving functions such as building a model of REM>... what's in the visual field, fitted into a longer-term model REM>... of the entire local geography, and figuring out how to REM>... perform navigation and hand-arm manipulation actions upon REM>... external objects. Accordingly I believe it would be REM>... appropriate to define classes of tasks performed by the REM>... various processing centers of the brain, and then to try to REM>... devise computer systems to perform each of these tasks. > From: "Dmitry A. Kazakov" > Yes. A "bottom up" approach is more productive. At least it > produces solutions for some problems. However, we still don't > know the architecture of intelligence. The tasks we had > identified reflect our understanding of intelligence, incomplete > and likely wrong. We don't know in which relation these tasks are > to general intelligence. I'm leaning toward the view that there's no such thing as "general intelligence", that what we appreciate in humans (and to a lesser degree in other great apes, cetaceans, cephalopod, and some birds), and what we thus try to measure in IQ tests, isn't a single "general intelligence", but rather a hodge podge of specialized types of cognitive skills. We are blind to skills that we humans don't have, thus blind to whatever other capabilities a true "general intelligence" would have if it existed, thus unable to distinguish between our own hodge podge and a hypothetical "general intelligence", because all we see of either is the intersection between what we can see and what's actually there, namely our hodge podge in both cases. It's analagous between humans not being able to distinguish a 3-color photo of a flower and a true all-spectrai view (including UV) of a flower, because we can't see what's different, namely the presence or absence of the UV. (Insects however *can* see the UV, so they would not confuse the two.) > Furthermore, considering these tasks in the context of > intelligence, it seems not so important to have a solution, > rather the method used. For a long time it was thought, for > example, that playing chess requires intelligence. Then the task > was solved using a stupid machine. Arguably, solutions of this > kind combined will never emerge into intelligence. Humans who play a high level of Chess use a more general mental capacity that *is* one of the components of the hodge podge of natural human cognitive skills. But skill at playing Chess is such a narrow specialized skill that no such general method is needed, a "stupid machine" is sufficient. If we knew the full capabilities of that particular skill, we might devise a test that is much more general than just Chess, perhaps the ability to learn any new kind of board game, and then to self-train at such a game and get better, and also the ability to recognize game-theoretical ideas within the natural world and human society, such as the "war between the sexes" and diplomacy between nations, respectively. If we could then build a computer system that did well at this full range of game-theoretic rules-learning and strategy-optimization, that might turn out to be a component of A.I. rather than just a "stupid machine" algorithm. > This is the key issue of AI: if all subtasks were solved by > whatever means, would that result in intelligence? If the full range of subtasks were solved by tools that were specialized only enough to match the capabilities of the various processing centers of the human brain, maybe yes. But if we break down the skillset to overly-specialized skills, we'd need billions of different programs, one per too-specific skill, and every time we discover a new overly-specialized skill that humans can solve, we'd have to start from scratch writing yet another program to solve that one new skill, and it'd be a "mad queen race" between people inventing new skills that humans can solve but the computer can't yet solve, and people inventing special programs for the computer to solve those new skills. Thus I agree with you if we devise algorithms to solve too-specific skills such as Chess, but if we instead define wide-span skill-types that match parts of the human brain that might indeed "result in intelligence" to match human capabilities. > Or is intelligence rather a method than any concrete task at hand. It's a combination of different methods, each of which can solve one rather general kind of task, if my leaning is correct. > ... since it is unknown what intelligence means, The word "intelligence" means anything we define it to mean. In some context we might define it simply to mean the ability to operate an android with sufficient skill to pass as a human through all the ordinary natural and social situations of a normal human life, including moving about and socializing with humans, engaging in intimate relationships, learning useful job skills from a human-oriented train program and consequently performing useful work to "earn a living", etc. In another context, when robotics aren't yet developed well enough to make a functional android, we might use a lesser definition whereby the "A.I." system can browse newsgroups and compose replies that make more sense than Xah Lee. REM> I disagree. "Jeopardy!" is already a feature that can be measured, REM> and in fact has been measured, where the computer did quite well. > But is it a *relevant* feature? IMO it's a broader-span skill than expertise at Chess. IMO it doesn't closely match one of the built-in skills of the human brain, but I don't know whether it's a broad-enough skill to be a component in an A.I. Maybe the underlying technology of Watson (with minor generalization to avoid the trick answer-before-question facade of Jeopardy, to answer several other formats of free-language fact-lookup questions) is broad enough to serve as a component of an A.I. that is rather non-human, or maybe it's too narrow at all and needs a new research breakthrough to widen it enough. > I think that sweeping floor is a much harder problem, I think that sweeping floor is a totally different type of skill, not comparable to playing Jeopardy, like apples and not oranges and not even earthworms but thermophilic sulfur bacteria. Maybe the underlying Watson technology can be generalized to one component of the hodge podge, and sweeping floor can be generalized to another, and twenty more equally generalized components will suffice to match human cognitive ability. By the way, I expect Japanese robotics to be able to build a floor-sweeping robot within the next ten years. But given that automated vacuum cleaners are a better way to clean tiny debris off the floor, and in fact automated vacuum cleaners are already nearly consumer ready, there'll be no economic incentive to actually produce a floor-sweeping robot, except if some major company such as Google offers an "X prize" for it. (Aside: I learned just within the past week that Google has offered an "X prize" for any non-government whatever to send a robot rover to Luna.) > much closer to general intelligence than indexing massives of texts. Which is closer to having a complete automobile, one wheel, or one cylinder of an engine? In fact you need to generalize the cylinder to a complete engine, and generalize the wheel to a complete wheel+tire+axel+driveshaft system, and also add the chassis+frame and several other systems before you have a working car. Indexing massives of texts may turn out to be one key part of just one component of human intelligence. REM> Google and Bing have been working on another test set, users' REM> queries in their search engines, to try to present what the user REM> really is asking about before the other keyword matches. Last I saw REM> they haven't done a good job. > Actually they become worse each year. I don't have an objective test to determine if your suspicion is correct, but I have a vague feeling you might be correct, except the effective spelling checker built into Google's search engine hasn't gotten noticeably worse lately, and might be getting slightly better. The one area I've noticed is utter crap and not getting any better is disambiguating search terms such as people's names. I've been designing a cybernetic (mix of human and computer components) system that will organize the various Google search results according to the meaning of the term. For example, if you search for "Heather Thompson" it would disambiguate that into the several hundred people by that name, perhaps organized into a hierarchy or a sub-search engine, and then once you have picked *which* individual person you are looking for (the one who was beaten by her husband, or the math professor, or the one who drove wrecklessly on a country road, or the one who works at a bank, etc.), it'd show you the Google search results *only* for that one person. Watch TinyURL.Com/RLlink which will include disambiguation and identification of individual people as the first component. > As with an AI for sweeping floor they show unintelligence by > total inability to classify content between "valuable" and "junk." That's yet another dimension, after relevance to what you really wanted to know about. But whereas relevance (which "Heather Thompson" are you asking about) is a matter of fact, and truth-value is also a matter of fact (for which TinyURL.Com/TruFut will be a useful cybernetic aid), value contributing toward your current task-goal is a matter of opinion, *your* opinion ("you" being the person requesting the search), because nobody else knows what you are aiming for. You are in a bit of a dilemma. If you publicize your entire research programme, so that Google can in principle "read your mind" as to what relevant to your research needs, somebody else can steal your programme and publish before you do and thus deprive you of all your work. But if you keep the purpose of your query confidential, Google can't even in principle "read your mind" to know what will really help you and what will be useless to your current need. I suppose you can play a game where you give Google just enough clues to be able to find what's relevant to your needs, but not enough information to be able to steal your research programme. There several types of back-and-forth interactive query-disambiguation systems that might be used to guide Google towards increased relevance: - Salton's original idea of the user scoring each search result per relevance and then the ISR system using a sort of Bayesian fitting of clues with relevance to add weight to some clues and discount others. - Wikipedia's current system of disambiguation pages. - User manually adding additional search terms. I suspect an A.I. system might operate both ends of the Google/User relationship, playing role of user to suggest queries that would be of use, playing role of Google to retrieve some information, playing role of user to evaluate those results and thus do one of these: - formally rate the search results, per Salton's system; - select a sub-category of search results, per Wikipedia's method; - modify the query for the next round. One thing IMDB.Com does is to assign unique terms to each person or each movie/TVprogram/episode per disambiguation of the original free-form search terms. Such a technique would allow an intermix of disambiguation of terms and additional terms to refine the search. Google-groups-search-key: imtrgfdi