Re: [agi] hello
On 8/13/08, Jim Bromer [EMAIL PROTECTED] wrote: On Wed, Aug 13, 2008 at 4:14 AM, rick the ponderer [EMAIL PROTECTED] wrote: Thanks for replying YKY Is the logic learning you are talking about inductive logic programming. If so, isn't ilp basically a search through the space of logic programs (i may be way off the mark here!), wouldn't it be too large of a search space to explore if you're trying reach agi. And if you're determined to learn a symbolic representation, wouldn't genetic programming be a better choice, since it won't get stuck in local minima. There is no reason why symbolic reasoning could not incorporate some kind of random combinatoric search methods like those used in GA searches. Categorical imagination can be used to examine the possible creation of new categories; the method does not have to be limited to the examination of new combinations of previously derived categories. And it does not have to be limited to incremental methods either. For example, the method might be used to combine fragments of surface features observed in the IO data environment. Combinatoric search can be also used with the creation and consideration of conjectures about possible explanations of observed data events. One of the most important aspects of these kinds of searches is that they can be used in serendipitous methods to detect combinations or conjectures that might be useful in some other problem even when they don't solve the current search goal that they were created for. While discussions about these subjects must utilize some traditional frames of reference, the conventions of their use in conversation should not be considered as absolute limitations on their possible modifications. They can be used as starting points of further conversation. YKY's and Ben Goetzel's recent comments sound as if they are referring to strictly predefined categories when they talk about symbolic methods, but I would be amazed if that represents their ultimate goals in AI research. Similarly, other unconventional methods can be considered when thinking about ANN's and GA's, but I think that novel approaches to symbolic methods offers the best bet for some of the same reasons that YKY mentioned. Jim Bromer --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com For example, the method might be used to combine fragments of surface features observed in the IO data environment. Combinatoric search can be also used with the creation and consideration of conjectures about possible explanations of observed data events. One of the most important aspects of these kinds of searches is that they can be used in serendipitous methods to detect combinations or conjectures that might be useful in some other problem even when they don't solve the current search goal that they were created for. Is that any different to clustering? --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] hello
On 8/15/08, rick the ponderer [EMAIL PROTECTED] wrote: On 8/13/08, Jim Bromer [EMAIL PROTECTED] wrote: On Wed, Aug 13, 2008 at 4:14 AM, rick the ponderer [EMAIL PROTECTED] wrote: Thanks for replying YKY Is the logic learning you are talking about inductive logic programming. If so, isn't ilp basically a search through the space of logic programs (i may be way off the mark here!), wouldn't it be too large of a search space to explore if you're trying reach agi. And if you're determined to learn a symbolic representation, wouldn't genetic programming be a better choice, since it won't get stuck in local minima. There is no reason why symbolic reasoning could not incorporate some kind of random combinatoric search methods like those used in GA searches. Categorical imagination can be used to examine the possible creation of new categories; the method does not have to be limited to the examination of new combinations of previously derived categories. And it does not have to be limited to incremental methods either. For example, the method might be used to combine fragments of surface features observed in the IO data environment. Combinatoric search can be also used with the creation and consideration of conjectures about possible explanations of observed data events. One of the most important aspects of these kinds of searches is that they can be used in serendipitous methods to detect combinations or conjectures that might be useful in some other problem even when they don't solve the current search goal that they were created for. While discussions about these subjects must utilize some traditional frames of reference, the conventions of their use in conversation should not be considered as absolute limitations on their possible modifications. They can be used as starting points of further conversation. YKY's and Ben Goetzel's recent comments sound as if they are referring to strictly predefined categories when they talk about symbolic methods, but I would be amazed if that represents their ultimate goals in AI research. Similarly, other unconventional methods can be considered when thinking about ANN's and GA's, but I think that novel approaches to symbolic methods offers the best bet for some of the same reasons that YKY mentioned. Jim Bromer --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com For example, the method might be used to combine fragments of surface features observed in the IO data environment. Combinatoric search can be also used with the creation and consideration of conjectures about possible explanations of observed data events. One of the most important aspects of these kinds of searches is that they can be used in serendipitous methods to detect combinations or conjectures that might be useful in some other problem even when they don't solve the current search goal that they were created for. Is that any different to clustering? where you talk about discovering new categories from IO data. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] hello
On Wed, Aug 13, 2008 at 6:31 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: To use Thorton's example, he demontrated that a checkerboard pattern can be learned using logic easily, but it will drive a NN learner crazy. Note that neural networks are a broad subject and don't only include perceptrons, but also self-organising maps and other connectionist set ups. In particular, Hopfield networks are an associative memory system that would have no problem learning/memorising a checkerboard pattern (or any other pattern, the only problem occurs when memorized patterns begin to overlap). A logic system system would be a lot more efficient though. J --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] hello
On 8/13/08, rick the ponderer [EMAIL PROTECTED] wrote: Reading this, I get the view of ai as basically neural networks, where each individual perceptron could be any of a number of algorithms (decision tree, random forest, svm etc). I also get the view that academics such as Hinton are trying to find ways of automatically learning the network, whereas there could also be a parallel track of engineering the network, manually creating it perceptron by percetron, in the way Rodney Brooks advocates bottom up subsumption architecture. How does opencog relate to the above viewpoint. Is there something fundamentally flawed in the above as an approach to achieving agi. NN *may* be inadequate for AGI, because logic-based learning seems to be, at least for some datasets, more efficient than NN learning (that includes variants such as SVMs). This has been my intuition for some time, and recently I've found a book that explores this issue in more detail. See Chris Thorton 2000, Truth from Trash -- how learning makes sense, MIT press, or some of his papers on his web site. To use Thorton's example, he demontrated that a checkerboard pattern can be learned using logic easily, but it will drive a NN learner crazy. It doesn't mean that the NN approach is hopeless, but it faces some challenges. Or, maybe this intuition is wrong (ie, do such heavily logical datasets occur in real life?) YKY --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] hello
On 8/13/08, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: On 8/13/08, rick the ponderer [EMAIL PROTECTED] wrote: Reading this, I get the view of ai as basically neural networks, where each individual perceptron could be any of a number of algorithms (decision tree, random forest, svm etc). I also get the view that academics such as Hinton are trying to find ways of automatically learning the network, whereas there could also be a parallel track of engineering the network, manually creating it perceptron by percetron, in the way Rodney Brooks advocates bottom up subsumption architecture. How does opencog relate to the above viewpoint. Is there something fundamentally flawed in the above as an approach to achieving agi. NN *may* be inadequate for AGI, because logic-based learning seems to be, at least for some datasets, more efficient than NN learning (that includes variants such as SVMs). This has been my intuition for some time, and recently I've found a book that explores this issue in more detail. See Chris Thorton 2000, Truth from Trash -- how learning makes sense, MIT press, or some of his papers on his web site. To use Thorton's example, he demontrated that a checkerboard pattern can be learned using logic easily, but it will drive a NN learner crazy. It doesn't mean that the NN approach is hopeless, but it faces some challenges. Or, maybe this intuition is wrong (ie, do such heavily logical datasets occur in real life?) YKY -- *agi* | Archives https://www.listbox.com/member/archive/303/=now https://www.listbox.com/member/archive/rss/303/ | Modifyhttps://www.listbox.com/member/?;Your Subscriptionhttp://www.listbox.com Thanks for replying YKY Is the logic learning you are talking about inductive logic programming. If so, isn't ilp basically a search through the space of logic programs (i may be way off the mark here!), wouldn't it be too large of a search space to explore if you're trying reach agi. And if you're determined to learn a symbolic representation, wouldn't genetic programming be a better choice, since it won't get stuck in local minima. Would neural networks be better in that case because they have the mechanisms as in Geoff Hinton's paper that improve on random searching. Also, if you did manage to learn a giant logic program that represented ai, could it be easily parallelized the way a neural network can be (so that it can run in real time). --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] hello
On 8/13/08, rick the ponderer [EMAIL PROTECTED] wrote: Thanks for replying YKY Is the logic learning you are talking about inductive logic programming. If so, isn't ilp basically a search through the space of logic programs (i may be way off the mark here!), wouldn't it be too large of a search space to explore if you're trying reach agi. ** Yes, and I guess the search space would be huge no matter what kind of learning substrate we use. At least one redeeming trick (for symbolic AI) is that we can limit the depth of the search of programs, and my intuition is that commonsense reasoning is mostly shallow (ie, involving few inference steps). And if you're determined to learn a symbolic representation, wouldn't genetic programming be a better choice, since it won't get stuck in local minima. * It is possible to use GA to search the ILP space; there is research in that area. I may use that too. One interesting question is to compare ILP search in the space of logic programs vs genetic programming (ie search in program spaces such as Lisp or combinator logic or lambda calculus). Unfortunately I'm unfamiliar with the latter, so I need some time to study that. Would neural networks be better in that case because they have the mechanisms as in Geoff Hinton's paper that improve on random searching. ** This is just the age-old debate of symbolic AI vs connectionism, given a new twist in the context of machine learning. Note that that first debate was never really settled. So, my bet is that we need NN-style learning at the low levels, and symbolic-style learning at the high levels. I tend to focus on the symbolic side. I'm very skeptical whether NN learning can solve high-level symbolic problems. Also, if you did manage to learn a giant logic program that represented ai, could it be easily parallelized the way a neural network can be (so that it can run in real time). Yes, logical inference can be parallelized. I have a book about it, but I haven't bothered to study that -- design first, optimize later. YKY --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] hello
On Wed, Aug 13, 2008 at 4:14 AM, rick the ponderer [EMAIL PROTECTED] wrote: Thanks for replying YKY Is the logic learning you are talking about inductive logic programming. If so, isn't ilp basically a search through the space of logic programs (i may be way off the mark here!), wouldn't it be too large of a search space to explore if you're trying reach agi. And if you're determined to learn a symbolic representation, wouldn't genetic programming be a better choice, since it won't get stuck in local minima. There is no reason why symbolic reasoning could not incorporate some kind of random combinatoric search methods like those used in GA searches. Categorical imagination can be used to examine the possible creation of new categories; the method does not have to be limited to the examination of new combinations of previously derived categories. And it does not have to be limited to incremental methods either. For example, the method might be used to combine fragments of surface features observed in the IO data environment. Combinatoric search can be also used with the creation and consideration of conjectures about possible explanations of observed data events. One of the most important aspects of these kinds of searches is that they can be used in serendipitous methods to detect combinations or conjectures that might be useful in some other problem even when they don't solve the current search goal that they were created for. While discussions about these subjects must utilize some traditional frames of reference, the conventions of their use in conversation should not be considered as absolute limitations on their possible modifications. They can be used as starting points of further conversation. YKY's and Ben Goetzel's recent comments sound as if they are referring to strictly predefined categories when they talk about symbolic methods, but I would be amazed if that represents their ultimate goals in AI research. Similarly, other unconventional methods can be considered when thinking about ANN's and GA's, but I think that novel approaches to symbolic methods offers the best bet for some of the same reasons that YKY mentioned. Jim Bromer --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
RE: [agi] Hello from Kevin Copple
Gary Miller wrote: I also agree that the AGI approach of modeling and creating a self learning system is a valid bottom up approach to AGI. But it is much harder for me with my limited mathematical and conceptual knowledge of the research to grasp how and when these systems will be able jumpstart themselves and evolve to the point of communicating in English. Sure. In my view, the path involves teaching an AGI to carry out simple tasks in an environment (physical or digital) and then teaching it to communicate about these tasks and related entities in its environment... While it is true that most bots today generate a reflexive response based only on the user's input, it is possible to extend bot technology by generating the response based upon the following additional internal stimuli not provided in the current input they are responding to. These stimuli provide at least a portion of the grounding I think you are referring to. Hm... Actually, I think you're getting at a deep point here. Potentially, *conversational pragmatics* and *inferred psychology* can be used to ground *semantics*, for a chat bot... For example, suppose there's a pattern of word usage, sentence length, etc., which correlates with humans being angry. The bot can learn to correlate this pattern with the word angry. It is thus grounding the word angry with a nonlinguistic pattern... It may then learn different patterns corresponding to very angry versus slightly angry .. Suppose there's also a pattern of word usage, sentence length, punctuation use, etc., that corresponds to the emotion of happy ... and very happy vs. slightly happy If it also understands very long sentence vs. slightly long sentence vs. not long sentence [via grounding these in sentence lengths], then it may be able to extrapolate from these examples to form an abstract model of very-ness in general... Based on this line of thinking, I have to modify and partially retract my previous statement. If a chat bot is given the ability to study patterns in language usage, such as the ones mentioned above, then it may use these patterns as a nonlinguistic domain in which to ground its linguistic knowledge... So, I think that truly intelligent language usage COULD potentially be learned by a chat bot I still think this is trickier than learning it via a more physical-world-ish grounding domain, but it's far from impossible Very interesting point, Gary, thanks!! -- Ben --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: [agi] Hello from Kevin Copple
From: Kevin Copple [EMAIL PROTECTED] It seems to me that rout memorization is an aspect of human learning, so why not include a variety of jokes, poems, trivia, images, and so on as part of an AI knowledge base? In the EllaZ system we refer to these chunks of data as Convuns (conversational units). This is an important issue. One extreme approach in AI and CogSci is to reduce the meaning of linguistic chunks (phrases, sentenses, paragraphs, and texts) into basic components (example: Schank's CD Theory, Wierzbicka's primes, and Frege's principle of compositionality). I think such an approach is fine for certain formal languages, but definitely not OK for some others, and especially, won't work well for any natural language. However, I feel many statistical NLP approaches are going to the other extreme, that is, to take a linguistic chunk as a whole, without analysing its semantic relation with its components. I don't think we can go very far in this direction, neither. I hope Ella dosen't fall into this category. Ella was lucky enough to win the 2002 Loebner Prize Contest, which can be somewhat arbitrary with the limited number of judges and limited length of conversations. She has a number of functional features that I suspect the engineering students selected as judges were more likely to test and appreciate. I don't find any document about the system on the website. Is there any that you can share with us? Though I'm also interested in I Ching, your claim The I Ching (Yi Jing), dating as far back as 2000 B.C., can be considered to be the first computational AI, and the first binary computer. is still way too strong for me to agree. ;-) I am currently living Tianjin, China, having sold my import/export chemicals business to a competitor. My wife, Zhang Ying, is a local girl who doesn't care for the food in the US and doesn't like being away from her friends and family. So, I am between jobs and working on www.EllaZ.com for the next year or so. Tianjin is my hometown, and I'm back there every summer in the recent years. I hope you enjoy your life there. Pei We are always on the outlook for collaborators and ideas we can borrow :-) Cheers . . . Kevin Copple --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
RE: [agi] Hello from Kevin Copple
Hi Kevin, I know something about the Loebner prize from following the AI career of Jason Hutchens, whose chatbot won the contest in the late 90's, and who I knew slightly when I was living in Perth (Western Australia) in the mid-90's Your approach, on the surface, seems fairly similar to Jason's. I'm guessing you're familiar with his work, but others may not be so I'll post some links here. An old but good paper from his on chatbots and the Loebner contest: http://ciips.ee.uwa.edu.au/Papers/Technical_Reports/1997/05/ His company a-i.com, which essentially shut down about a year ago http://www.a-i.com/ My own approach is quite opposite. While I do aim to make my Novamente AI system (www.realai.net) chat eventually, I've detailed a complex design for integrative cognition, and I think we need a lot of that implemented before we can have the system do chat in a meaningful way [i.e. chat while understanding what it's talking about]. I have a lot of skepticism about any approach to AGI that is primarily or entirely language-focused. I doubt it's going to be possible to get a system to have any significant general intelligence unless it has access, not only to language, but also to a nonlinguistic domain in which some of its linguistic experience can be grounded or anchored [to use two related terms from the cog sci literature]. I agree that a significant part of human conversation consists of rote memory, and reflexive responses according to habitual communication patterns. To me, however, these are the least interesting parts of human conversation And I'm not sure how far mimicking these parts of human conversation gets you, in terms of emulating the other parts of human conversation, which involve deeper thought. One thing is sure though: The database of conversational units you're compiling could be *very useful* to a Novamente system [or other AGI system] that was trying to learn to chat [though we're not ready for that quite yet]. It will be a great source of information on conversational pragmatics Please, continue maintaining that DB, and maintain it carefully!!! One of these days I'll be wanting to talk to you about an arrangement for sharing it ;) -- Ben Goertzel -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]]On Behalf Of Kevin Copple Sent: Sunday, December 08, 2002 4:11 AM To: [EMAIL PROTECTED] Subject: [agi] Hello from Kevin Copple I just recently joined this e-mail list after following some links posted by Tony Lofthouse in the Generation5 forum. I am working on a natural language project that can be seen at www.EllaZ.com, and am interested in what you all are up to. The e-mails I have received from this list in the last day or so have been interesting and informative. Thanks! My approach to doing something in the AI field is to start with basic interface, knowledge, and functional features that can be implemented and demonstrated. Now that a basic framework is in place, the system can be expanded and built upon as various techniques are identified as useful and incorporated. It seems to me that rout memorization is an aspect of human learning, so why not include a variety of jokes, poems, trivia, images, and so on as part of an AI knowledge base? In the EllaZ system we refer to these chunks of data as Convuns (conversational units). One plan is for the system to log interactions with users and identify patterns of interest. The system would then be able to predict which Convuns a user would most likely be interested in, and also be able to evaluate the interest in a particular Convun. Ella was lucky enough to win the 2002 Loebner Prize Contest, which can be somewhat arbitrary with the limited number of judges and limited length of conversations. She has a number of functional features that I suspect the engineering students selected as judges were more likely to test and appreciate. I am currently living Tianjin, China, having sold my import/export chemicals business to a competitor. My wife, Zhang Ying, is a local girl who doesn't care for the food in the US and doesn't like being away from her friends and family. So, I am between jobs and working on www.EllaZ.com for the next year or so. We are always on the outlook for collaborators and ideas we can borrow :-) Cheers . . . Kevin Copple --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]