Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Mike Tintner wrote: Sounds a little confusing. Sounds like you plan to evolve a system through testing thousands of candidate mechanisms. So one way or another you too are taking a view - even if it's an evolutionary, I'm not taking a view view - on, and making a lot of asssumptions about -how systems evolve -the known architecture of human cognition. No, I think because of the paucity of information I gave you have misunderstood slightly. Everything I mentioned was in the context of an extremely detailed framework that tries to include all of the knowledge we have so far gleaned by studying human cognition using the methods of cognitive science. So I am not making assumptions about the architecture of human cognition I am using every scrap of experimental data I can. You can say that this is still assuming that the framework is correct, but that is nothing compared to the usual assumptions made in AI, where the programmer just picks up a grab bag of assorted ideas that are floating around in the literature (none of them part of a coherent theory of cognition) and starts hacking. And just because I talk of thousands of candidate mechanisms, that does not mean that there is evolution involved: it just means that even with a complete framework for human cognition to start from there are still so many questions about the low-level to high-level linkage that a vast number of mechanisms have to be explored. about which science has extremely patchy and confused knowledge. I don't see how any system-builder can avoid taking a view of some kind on such matters, yet you seem to be criticising Ben for so doing. Ben does not start from a complete framework for human cognition, nor does he feel compelled to stick close to the human model, and my criticisms (at least in this instance) are not really about whether or not he has such a framework, but about a problem that I can see on his horizon. I was hoping that you also had some view on how a system 's symbols should be grounded, especially since you mention Harnad, who does make vague gestures towards the brain's levels of grounding. But you don't indicate any such view. On the contrary, I explained exactly how they would be grounded: if the system is allowed to build its own symbols *without* me also inserting ungrounded (i.e. interpreted, programmer-constructed) symbols and messing the system up by forcing it to use both sorts of symbols, then ipso fact it is grounded. It is easy to build a grounded system. The trick is to make it both grounded and intelligent at the same time. I have one strategy for ensuring that it turns out intelligent, and Ben has another my problem with Ben's strategy is that I believe his attempt to ensure that the system is intelligent ends up compromising the groundedness of the system. Sounds like you too, pace MW, are hoping for a number of miracles - IOW creative ideas - to emerge, and make your system work. I don't understand where I implied this. You have to remember that I am doing this within a particular strategy (outlined in my CSP paper). When you see me exploring 'thousands' of candidate mechanisms to see how one parameter plays a role, this is not waiting for a miracle, it is a vital part of the strategy. A strategy that, I claim, is the only viable one. Anyway, you have to give Ben credit for putting a lot of his stuff principles out there on the line. I think anyone who wants to mount a full-scale assault on him ( why not?) should be prepared to reciprocate. Nice try, but there are limits to what I can do to expose the details. I have not yet worked out how much I should release and how much to withhold (I confess, I nearly decided to go completely public a month or so ago, but then changed my mind after seeing the dismally poor response that even one of the ideas provoked). Maybe in the near future I will write a summary account. In the mean time, yes, it is a little unfair of me to criticise other projects. But not that unfair. When a scientist sees a big problem with a theory, do you suppose they wait until they have a completely worked out alternative before discussing the fact that there is a problem with the theory that other people may be praising? That is not the way of science. 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=8660244id_secret=65349870-56ef76
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Linas Vepstas wrote: On Tue, Nov 13, 2007 at 12:34:51PM -0500, Richard Loosemore wrote: Suppose that in some significant part of Novamente there is a representation system that uses probability or likelihood numbers to encode the strength of facts, as in [I like cats](p=0.75). The (p=0.75) is supposed to express the idea that the statement [I like cats] is in some sense 75% true. Either way, we have a problem: a fact like [I like cats](p=0.75) is ungrounded because we have to interpret it. Does it mean that I like cats 75% of the time? That I like 75% of all cats? 75% of each cat? Are the cats that I like always the same ones, or is the chance of an individual cat being liked by me something that changes? Does it mean that I like all cats, but only 75% as much as I like my human family, which I like(p=1.0)? And so on and so on. Eh? You are standing at the proverbial office water coooler, and Aneesh says Wen likes cats. On your drive home, you mind races .. does this mean that Wen is a cat fancier? You were planning on taking Wen out on a date, and this tidbit of information could be useful ... when you try to build the entire grounding mechanism(s) you are forced to become explicit about what these numbers mean, during the process of building a grounding system that you can trust to be doing its job: you cannot create a mechanism that you *know* is constructing sensible p numbers and facts during all of its development *unless* you finally bite the bullet and say what the p numbers really mean, in fully cashed out terms. But has a human, asking Wen out on a date, I don't really know what Wen likes cats ever really meant. It neither prevents me from talking to Wen, or from telling my best buddy that ...well, I know, for instance, that she likes cats... Lack of grounding is what makes humour funny, you can do a whole Pygmalion / Seinfeld episode on she likes cats. No: the real concept of lack of grounding is nothing so simple as the way you are using the word grounding. Lack of grounding makes an AGI fall flat on its face and not work. I can't summarize the grounding literature in one post. (Though, heck, I have actually tried to do that in the past: didn't do any good). 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=8660244id_secret=64980585-67cbc9
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Hi, No: the real concept of lack of grounding is nothing so simple as the way you are using the word grounding. Lack of grounding makes an AGI fall flat on its face and not work. I can't summarize the grounding literature in one post. (Though, heck, I have actually tried to do that in the past: didn't do any good). FYI, I have read the symbol-grounding literature (or a lot of it), and generally found it disappointingly lacking in useful content... though I do agree with the basic point that non-linguistic grounding is extremely helpful for effective manipulation of linguistic entities... -- Ben G - 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=8660244id_secret=64981284-09925d
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Benjamin Goertzel wrote: On Nov 13, 2007 2:37 PM, Richard Loosemore [EMAIL PROTECTED] mailto:[EMAIL PROTECTED] wrote: Ben, Unfortunately what you say below is tangential to my point, which is what happens when you reach the stage where you cannot allow any more vagueness or subjective interpretation of the qualifiers, because you have to force the system to do its own grounding, and hence its own interpretation. I don't see why you talk about forcing the system to do its own grounding -- the probabilities in the system are grounded in the first place, as they are calculated based on experience. The system observes, records what it sees, abstracts from it, and chooses actions that it guess will fulfill its goals. Its goals are ultimately grounded in in-built feeling-evaluation routines, measuring stuff like amount of novelty observed, amount of food in system etc. So, the system sees and then acts ... and the concepts it forms and uses are created/used based on their utility in deriving appropriate actions. There is no symbol-grounding problem except in the minds of people who are trying to interpret what the system does, and get confused. Any symbol used within the system, and any probability calculated by the system, are directly grounded in the system's experience. There is nothing vague about an observation like Bob_Yifu was observed at time-stamp 599933322, or a fact Command 'wiggle ear' was sent at time-stamp 54. These perceptions and actions are the root of the probabilities the system calculated, and need no further grounding. What you gave below was a sketch of some more elaborate 'qualifier' mechanisms. But I described the process of generating more and more elaborate qualifier mechanisms in the body of the essay, and said why this process was of no help in resolving the issue. So, if a system can achieve its goals based on choosing procedures that it thinks are likely to achieve its goals, based on the knowledge it gathered via its perceived experience -- why do you think it has a problem? I don't really understand your point, I guess. I thought I did -- I thought your point was that precisely specifying the nature of a conditional probability is a rats-nest of complexity. And my response was basically that in Novamente we don't need to do that, because we define conditional probabilities based on the system's own knowledge-base, i.e. Inheritance A B .8 means If A and B were reasoned about a lot, then A would (as measred by an weighted average) have 80% of the relationships that B does But apparently you were making some other point, which I did not grok, sorry... Anyway, though, Novamente does NOT require logical relations of escalating precision and complexity to carry out reasoning, which is one thing you seemed to be assuming in your post. You are, in essence, using one of the trivial versions of what symbol grounding is all about. The complaint is not your symbols are not connected to experience. Everyone and their mother has an AI system that could be connected to real world input. The simple act of connecting to the real world is NOT the core problem. If you have an AGI system in which the system itself is allowed to do all the work of building AND interpreting all of its symbols, I don't have any issues with it. Where I do have an issue is with a system which is supposed to be doing the above experiential pickup, and where the symbols are ALSO supposed to be interpretable by human programmers who are looking at things like probability values attached to facts. When a programmer looks at a situation like ContextLink .7,.8 home InheritanceLink Bob_Yifu friend ... and then follows this with a comment like: which suggests that Bob is less friendly at home than in general. ... they have interpreted the meaning of that statement using their human knowledge. So here I am, looking at this situation, and I see: AGI system intepretation (implicit in system use of it) Human programmer intepretation and I ask myself which one of these is the real interpretation? It matters, because they do not necessarily match up. The human programmer's intepretation has a massive impact on the system because all the inference and other mechanisms are built around the assumption that the probabilities mean a certain set of things. You manipulate those p values, and your manipulations are based on assumptions about what they mean. But if the system is allowed to pick up its own knowledge from the environment, the implicit meaning of those p values will not necessarily match the human interpretation. As I say, the meaning is then implicit in the way the system *uses* those p values (and other stuff). It is a nontrivial question to ask whether the implicit system interpretation does indeed match the human intepretation built into the inference
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Benjamin Goertzel wrote: Hi, No: the real concept of lack of grounding is nothing so simple as the way you are using the word grounding. Lack of grounding makes an AGI fall flat on its face and not work. I can't summarize the grounding literature in one post. (Though, heck, I have actually tried to do that in the past: didn't do any good). FYI, I have read the symbol-grounding literature (or a lot of it), and generally found it disappointingly lacking in useful content... though I do agree with the basic point that non-linguistic grounding is extremely helpful for effective manipulation of linguistic entities... Ben, As you will recall, Harnad himself got frustrated with the many people who took the term symbol grounding and trivialized or distorted it in various ways. One of the reasons the grounding literature is such a waste of time (and you are right: it is) is that so many people talked so much nonsense about it. As far as I am concerned, your use of it is one of those trivial senses that Harnad complained of. (Essentially, if the system uses world input IN ANY WAY during the building of its symbols, then the system is grounded). The effort I put into that essay yesterday will have been completely wasted if your plan is to stick to that interpretation and not discuss the deeper issue that I raised. I really have no energy for pursuing yet another discussion about symbol grounding. Sorry: don't mean to blow you off, but you and I both have better things to do, and I foresee a big waste of time ahead if we pursue it. So let's just drop it? 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=8660244id_secret=64998305-6bdb18
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Richard, So here I am, looking at this situation, and I see: AGI system intepretation (implicit in system use of it) Human programmer intepretation and I ask myself which one of these is the real interpretation? It matters, because they do not necessarily match up. That is true, but in some cases they may approximate each other well.. In others, not... This happens to be a pretty simple case, so the odds of a good approximation seem high. The human programmer's intepretation has a massive impact on the system because all the inference and other mechanisms are built around the assumption that the probabilities mean a certain set of things. You manipulate those p values, and your manipulations are based on assumptions about what they mean. Well, the PLN inference engine's treatment of ContextLink home InheritanceLink Bob_Yifu friend is in no way tied to whether the system's implicit interpretation of the ideas of home or friend are humanly natural, or humanly comprehensible. The same inference rules will be applied to cases like ContextLink Node_66655 InheritanceLink Bob_Yifu Node_544 where the concepts involved have no humanly-comprehensible label. It is true that the interpretation of ContextLink and InheritanceLink are fixed by the wiring of the system, in a general way (but what kinds of properties are referred to by them may vary in a way dynamically determined by the system). In order to completely ground the system, you need to let the system build its own symbols, yes, but that is only half the story: if you still have a large component of the system that follows a programmer-imposed interpretation of things like probability values attached to facts, you have TWO sets of symbol-using mechanisms going on, and the system is not properly grounded (it is using both grounded and ungrounded symbols within one mechanism). I don't think the system needs to learn its own probabilistic reasoning rules in order to be an AGI. This, to me, is too much like requiring that a brain needs to learn its own methods for modulating the conductances of the bundles of synapses linking between the neurons in cell assembly A and cell assembly B. I don't see a problem with the AGI system having hard-wired probabilistic inference rules, and hard-wired interpretations of probabilistic link types. But the interpretation of any **particular** probabilistic relationship inside the system, is relative to the concepts and the empirical and conceptual relationships that the system has learned. You may think that the brain learns its own uncertain inference rules based on a lower-level infrastructure that operates in terms entirely unconnected from ideas like uncertainty and inference. I think this is wrong. I think the brain's uncertain inference rules are the result, on the cell assembly level, of Hebbian learning and related effects on the neuron/synapse level. So I think the brain's basic uncertain inference rules are wired-in, just as Novamente's are, though of course using a radically different infrastructure. Ultimately an AGI system needs to learn its own reasoning rules and radically modify and improve itself, if it's going to become strongly superhuman! But that is not where we need to start... -- Ben - 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=8660244id_secret=64998317-8c4281
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
RL:In order to completely ground the system, you need to let the system build its own symbols V. much agree with your whole argument. But - I may well have missed some vital posts - I have yet to get the slightest inkling of how you yourself propose to do this. - 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=8660244id_secret=65013351-96e8f0
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
On Nov 14, 2007 1:36 PM, Mike Tintner [EMAIL PROTECTED] wrote: RL:In order to completely ground the system, you need to let the system build its own symbols Correct. Novamente is designed to be able to build its own symbols. what is built-in, are mechanisms for building symbols, and for probabilistically interrelating symbols once created... ben g V. much agree with your whole argument. But - I may well have missed some vital posts - I have yet to get the slightest inkling of how you yourself propose to do this. - 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=8660244id_secret=65100803-21ddd3
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
On Wednesday 14 November 2007 11:28, Richard Loosemore wrote: The complaint is not your symbols are not connected to experience. Everyone and their mother has an AI system that could be connected to real world input. The simple act of connecting to the real world is NOT the core problem. Are we sure? How much of the real world are we able to get into our AGI models anyway? Bandwidth is limited, much more limited than in humans and other animals. In fact, it might be the equivalent to worm tech. To do the calculations would I just have to check out how many neurons are in a worm, how many sensory neurons, and rough information theoretic estimations as to the minimum and maximums as to amounts of information processing that the worm's sensorium could be doing? - Bryan - 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=8660244id_secret=65191610-b12544
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Bryan Bishop wrote: On Wednesday 14 November 2007 11:28, Richard Loosemore wrote: The complaint is not your symbols are not connected to experience. Everyone and their mother has an AI system that could be connected to real world input. The simple act of connecting to the real world is NOT the core problem. Are we sure? How much of the real world are we able to get into our AGI models anyway? Bandwidth is limited, much more limited than in humans and other animals. In fact, it might be the equivalent to worm tech. To do the calculations would I just have to check out how many neurons are in a worm, how many sensory neurons, and rough information theoretic estimations as to the minimum and maximums as to amounts of information processing that the worm's sensorium could be doing? I'm not quite sure where this is at . but the context of this particular discussion is the notion of 'symbol grounding' raised by Steven Harnad. I am essentially talking about how to solve the problem he described, and what exactly the problem was. Hence a lot of background behind this one, which if you don't know it might make it confusing. 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=8660244id_secret=65202116-6cf6d0
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
On Nov 14, 2007 11:58 PM, Bryan Bishop [EMAIL PROTECTED] wrote: Are we sure? How much of the real world are we able to get into our AGI models anyway? Bandwidth is limited, much more limited than in humans and other animals. In fact, it might be the equivalent to worm tech. To do the calculations would I just have to check out how many neurons are in a worm, how many sensory neurons, and rough information theoretic estimations as to the minimum and maximums as to amounts of information processing that the worm's sensorium could be doing? Pretty much. Let's take as our reference computer system a bog standard video camera connected to a high-end PC, which can do something (video compression, object recognition or whatever) with the input in real time. On the worm side, consider the model organism Caenorhabditis elegans, which has a few hundred neurons. It turns out that the computer has much more bandwidth. Then again, while intelligence unlike bandwidth isn't a scalar quantity even to a first approximation, to the extent they are comparable our best computer systems do seem to be considerably smarter than C. elegans. If we move up to something like a mouse, then the mouse has intelligence we can't replicate, and also has much more bandwidth than the computer system. Insects are somewhere in between, enough so that the comparison (both bandwidth and intelligence) doesn't produce an obvious answer; it's therefore considered not unreasonable to say present-day computers are in the ballpark of insect-smart. Of course that doesn't mean if we took today's software and connected it to mouse-bandwidth hardware it would become mouse-smart, but hopefully it means when we have that hardware we'll be able to use it to develop software that matches some of the things mice can do. (And it's still my opinion that by accepting - embracing - slowness on existing hardware we can work on the software at the same time as the hardware guys are working on their end, parallel rather than serial development.) - 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=8660244id_secret=65207531-031731
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Sounds a little confusing. Sounds like you plan to evolve a system through testing thousands of candidate mechanisms. So one way or another you too are taking a view - even if it's an evolutionary, I'm not taking a view view - on, and making a lot of asssumptions about -how systems evolve -the known architecture of human cognition. about which science has extremely patchy and confused knowledge. I don't see how any system-builder can avoid taking a view of some kind on such matters, yet you seem to be criticising Ben for so doing. I was hoping that you also had some view on how a system 's symbols should be grounded, especially since you mention Harnad, who does make vague gestures towards the brain's levels of grounding. But you don't indicate any such view. Sounds like you too, pace MW, are hoping for a number of miracles - IOW creative ideas - to emerge, and make your system work. Anyway, you have to give Ben credit for putting a lot of his stuff principles out there on the line. I think anyone who wants to mount a full-scale assault on him ( why not?) should be prepared to reciprocate. - RL: Mike Tintner wrote: RL:In order to completely ground the system, you need to let the system build its own symbols V. much agree with your whole argument. But - I may well have missed some vital posts - I have yet to get the slightest inkling of how you yourself propose to do this. Well, for the purposes of the present discussion I do not need to say how, only to say that there is a difference between two different research strategies for finding out what the mechanism is that does this. One strategy (the one that I claim has serious problems) is where you try to have your cake and eat it too: let the system build its own symbols, with attached parameters that 'mean' whatever they end up meaning after the symbols have been built, BUT then at the same time insist that some of the parameters really do 'mean' things like probabilities or likelihood or confidence values. If the programmer does anything at all to include mechanisms that rely on these meanings (these interpretations of what the parameters signify) then the programmer has second-guessed what the system itself was going to use those things for, and you have a conflict between the two. My strategy is to keep my hands off, not do anything to strictly interpret those parameters, and experimentally observe the properties of systems that seem loosely consistent with the known architecture of human cognition. I have a parameter, for instance, that seems to be a happiness or consistency parameter attached to a knowledge-atom. But beyond roughly characterising it as such, I do not insert any mechanisms that (implicitly or explicitly) lock the system into such an intepretation. Instead, I have a wide variety of different candidate mechanisms that use that parameter, and I look at the overall properties of systems that use these different candidate mechanisms. I let the system use the parameter according to the dictates of whatever mechanism is in place, but then I just explore the consequences (the high level behavior of the system). In this way I do not get a conflict between what I think the parameter 'ought' to mean and what the system is implicitly taking it to 'mean' by its use of the parameter. I could start talking about all the different candidate mechanisms, but there are thousands of them (at least thousands of candidates that I go so far as to test: they are generated in a semi-automatic way, so there are an unlimited number of potential candidates). 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/?; -- No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.503 / Virus Database: 269.15.30/1125 - Release Date: 11/11/2007 9:50 PM - 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=8660244id_secret=65232546-91c089
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Mike Tintner wrote: RL:In order to completely ground the system, you need to let the system build its own symbols V. much agree with your whole argument. But - I may well have missed some vital posts - I have yet to get the slightest inkling of how you yourself propose to do this. Well, for the purposes of the present discussion I do not need to say how, only to say that there is a difference between two different research strategies for finding out what the mechanism is that does this. One strategy (the one that I claim has serious problems) is where you try to have your cake and eat it too: let the system build its own symbols, with attached parameters that 'mean' whatever they end up meaning after the symbols have been built, BUT then at the same time insist that some of the parameters really do 'mean' things like probabilities or likelihood or confidence values. If the programmer does anything at all to include mechanisms that rely on these meanings (these interpretations of what the parameters signify) then the programmer has second-guessed what the system itself was going to use those things for, and you have a conflict between the two. My strategy is to keep my hands off, not do anything to strictly interpret those parameters, and experimentally observe the properties of systems that seem loosely consistent with the known architecture of human cognition. I have a parameter, for instance, that seems to be a happiness or consistency parameter attached to a knowledge-atom. But beyond roughly characterising it as such, I do not insert any mechanisms that (implicitly or explicitly) lock the system into such an intepretation. Instead, I have a wide variety of different candidate mechanisms that use that parameter, and I look at the overall properties of systems that use these different candidate mechanisms. I let the system use the parameter according to the dictates of whatever mechanism is in place, but then I just explore the consequences (the high level behavior of the system). In this way I do not get a conflict between what I think the parameter 'ought' to mean and what the system is implicitly taking it to 'mean' by its use of the parameter. I could start talking about all the different candidate mechanisms, but there are thousands of them (at least thousands of candidates that I go so far as to test: they are generated in a semi-automatic way, so there are an unlimited number of potential candidates). 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=8660244id_secret=65198894-3ece99
Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Mark Waser wrote: I'm going to try to put some words into Richard's mouth here since I'm curious to see how close I am . . . . (while radically changing the words). I think that Richard is not arguing about the possibility of Novamente-type solutions as much as he is arguing about the predictability of *very* flexible Novamente-type solutions as they grow larger and more complex (and the difficulty in getting it to not instantaneously crash-and-burn). Indeed, I have heard a very faint shadow of Richard's concerns in your statements about the tuning problems that you had with BioMind. This is true, but not precise enough to capture the true nature of my worry. Let me focus on one aspect of the problem. My goal here is to describe in a little detail how the Complex Systems Problem actually bites in a particular case. Suppose that in some significant part of Novamente there is a representation system that uses probability or likelihood numbers to encode the strength of facts, as in [I like cats](p=0.75). The (p=0.75) is supposed to express the idea that the statement [I like cats] is in some sense 75% true. [Quick qualifier: I know that this oversimplifies the real situation in Novamente, but I need to do this simplification in order to get my point across, and I am pretty sure this will not affect my argument, so bear with me]. We all know that this p value is not quite a probability or likelihood or confidence factor. It plays a very ambigous role in the system, because on the one hand we want it to be very much like a probability in the sense that we want to do calculations with it: we NEED a calculus of such values in order to combine facts in the system to make inferences. But we also do not want to lock ourselves into a particular interpretation of what it means, because we know full well that we do not really have a clear semantics for these numbers. Either way, we have a problem: a fact like [I like cats](p=0.75) is ungrounded because we have to interpret it. Does it mean that I like cats 75% of the time? That I like 75% of all cats? 75% of each cat? Are the cats that I like always the same ones, or is the chance of an individual cat being liked by me something that changes? Does it mean that I like all cats, but only 75% as much as I like my human family, which I like(p=1.0)? And so on and so on. Digging down to the root of this problem (and this is the point where I am skipping from baby stuff to hard core AI) we want these numbers to be semantically compositional and interpretable, but in order to make sure they are grounded, the system itself is going to have to build them interpret them without our help ... and it is not clear that this grounding can be completely implemented. Why is it not clear? Because when you try to build the entire grounding mechanism(s) you are forced to become explicit about what these numbers mean, during the process of building a grounding system that you can trust to be doing its job: you cannot create a mechanism that you *know* is constructing sensible p numbers and facts during all of its development *unless* you finally bite the bullet and say what the p numbers really mean, in fully cashed out terms. [Suppose you did not do this. Suppose you built the grounding mechanism but remained ambiguous about the meaning of the p numbers. What would the resulting system be computing? From end to end it would be building facts with p numbers, but you the human observer would still be imposing an interpretation on the facts. And if you are still doing anything to interpret, it cannot be grounded]. Now, as far as I understand it, the standard approach to this condundrum is that researchers (in Novamente and elsewhere) do indeed make an attempt to disambiguate the p numbers, but they do it by developing more sophisticated logical systems. First, perhaps, error-value bands of p values instead of sharp values. And temporal logic mechanisms to deal with time. Perhaps clusters of p and q and r and s values, each with some slightly different zones of applicability. More generally, people try to give structure to the qualifiers that are appended to the facts: [I like cats](qualfier=value) instead of [I like cats](p=0.75). The question is, does this process of refinement have an end? Does it really lead to a situation where the qualifier is disambiguated and the semantics is clear enough to build a trustworthy grounding system? Is there a closed-form solution to the problem of building a logic that disambiguates the qualifiers? Here is what I think will happen if this process is continued. In order to make the semantics unambiguous enough to let the system ground its own knowledge without the interpretation of p values, researchers will develop more and more sophisticated logics (with more and more structured replacements for that simple p value), until they are
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Richard, The idea of the PLN semantics underlying Novamente's probabilistic truth values is that we can have **both** -- simple probabilistic truth values without highly specific interpretation -- more complex, logically refined truth values, when this level of precision is necessary To make the discussion more concrete, I'll use a specfic example to do with virtual animals in Second Life. Our first version of the virtual pets won't use PLN in this sort of way, it'll be focused on MOSES evolutionary learning; but, this is planned for the second version and is within the scope of what Novamente can feasibly be expected to do with modest effort. Consider an avatar identified as Bob_Yifu And, consider the concept of friend, which is a ConceptNode -- associated to the WordNode friend via a learned ReferenceLink -- defined operationally via a number of links such as ImplicationLink AND InheritanceLink X friend EvaluationLink near (I, X) Pleasure (this one just says that being near a friend confers pleasure. Other links about friendship may contain knowledge such as that friends often give one food, friends help one find things, etc.) The concept of friend may be learned, via mining of the animal's experience-base -- basically, this is a matter of learning that there are certain predicates whose SatisfyingSets (the set of Atoms that fulfill the predicate) have significant intersection, and creating a ConceptNode to denote that intersection. Then, once the concept of friend has been formed, more links pertaining to it may be learned via mining the experience base and via inference rules. Then, we can may find that InheritanceLink Bob_Yifu friend .9,1 (where the .9,1 is an interval probability, interpreted according to the indefinite probabilities framework) and this link mixes intensional and extensional inheritance, and thus is only useful for heuristic reasoning (which however is a very important kind). What this link means is basically that Bob_Yifu's node in the memory has a lot of the same links as the friend node -- or rather, that it **would**, if all its links were allowed to exist rather than being pruned to save memory. So, note that the semantics are actually tied to the mind itself. Or we can make more specialized logical constructs if we really want to, denoting stuff like -- at certain times Bob_Yifu is a friend -- Bob displays some characteristics of friendship very strongly, and others not at all -- etc. We can also do crude, heuristic contextualization like ContextLink .7,.8 home InheritanceLink Bob_Yifu friend which suggests that Bob is less friendly at home than in general. Again this doesn't capture all the subtleties of Bob's friendship in relation to being at home -- and one could do so if one wanted to, but it would require introducing a larger complex of nodes and links, which is not always the most appropriate thing to do. The PLN inference rules are designed to give heuristically correct conclusions based on heuristically interpreted links; or more precise conclusions based on more precisely interpreted links. Finally, the semantics of PLN relationships is explicitly an **experiential** semantics. (One of the early chapters in the PLN book, to appear via Springer next year, is titled Experiential Semantics.) So, all node and link truth values in PLN are intended to be settable and adjustable via experience, rather than via programming or importation from databases or something like that. Now, the above example is of course a quite simple one. Discussing a more complex example would go beyond the scope of what I'm willing to do in an email conversation, but the mechanisms I've described are not limited to such simple examples. I am aware that identifying Bob_Yifu as a coherent, distinct entity is a problem faced by humans and robots, and eliminated via the simplicity of the SL environment. However, there is detailed discussion in the (proprietary) NM book of how these same mechanisms may be used to do object recognition and classification, as well. You may of course argue that these mechanisms won't scale up to large knowledge bases and rich experience streams. I believe that they will, and have arguments but not rigorous proofs that they will. -- Ben G On Nov 13, 2007 12:34 PM, Richard Loosemore [EMAIL PROTECTED] wrote: Mark Waser wrote: I'm going to try to put some words into Richard's mouth here since I'm curious to see how close I am . . . . (while radically changing the words). I think that Richard is not arguing about the possibility of Novamente-type solutions as much as he is arguing about the predictability of *very* flexible Novamente-type solutions as they grow larger and more complex (and the difficulty in getting it to not instantaneously crash-and-burn). Indeed, I have heard a very faint shadow of Richard's concerns in your statements about the tuning problems that you had
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Mike Tintner wrote: RL:Suppose that in some significant part of Novamente there is a representation system that uses probability or likelihood numbers to encode the strength of facts, as in [I like cats](p=0.75). The (p=0.75) is supposed to express the idea that the statement [I like cats] is in some sense 75% true. This essay seems to be a v.g. demonstration of why the human system almost certainly does not use numbers or anything like, as stores of value - but raw, crude emotions. How much do you like cats [or marshmallow ice cream]? Miaow//[or yummy] [those being an expression of internal nervous and muscular impulses] And black cats [or strawberry marshmallow] ? Miaow-miaoww![or yummy yummy] . It's crude but it's practical. It is all a question of what role the numbers play. Conventional AI wants them at the surface, and transparently interpretable. I am not saying that there are no numbers, but only that they are below the surface, and not directly interpretable. that might or might not gibe with what you are saying ... although I would not go so far as to put it in the way you do. 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=8660244id_secret=64636829-14d428
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
Ben, Unfortunately what you say below is tangential to my point, which is what happens when you reach the stage where you cannot allow any more vagueness or subjective interpretation of the qualifiers, because you have to force the system to do its own grounding, and hence its own interpretation. What you gave below was a sketch of some more elaborate 'qualifier' mechanisms. But I described the process of generating more and more elaborate qualifier mechanisms in the body of the essay, and said why this process was of no help in resolving the issue. Richard Loosemore Benjamin Goertzel wrote: Richard, The idea of the PLN semantics underlying Novamente's probabilistic truth values is that we can have **both** -- simple probabilistic truth values without highly specific interpretation -- more complex, logically refined truth values, when this level of precision is necessary To make the discussion more concrete, I'll use a specfic example to do with virtual animals in Second Life. Our first version of the virtual pets won't use PLN in this sort of way, it'll be focused on MOSES evolutionary learning; but, this is planned for the second version and is within the scope of what Novamente can feasibly be expected to do with modest effort. Consider an avatar identified as Bob_Yifu And, consider the concept of friend, which is a ConceptNode -- associated to the WordNode friend via a learned ReferenceLink -- defined operationally via a number of links such as ImplicationLink AND InheritanceLink X friend EvaluationLink near (I, X) Pleasure (this one just says that being near a friend confers pleasure. Other links about friendship may contain knowledge such as that friends often give one food, friends help one find things, etc.) The concept of friend may be learned, via mining of the animal's experience-base -- basically, this is a matter of learning that there are certain predicates whose SatisfyingSets (the set of Atoms that fulfill the predicate) have significant intersection, and creating a ConceptNode to denote that intersection. Then, once the concept of friend has been formed, more links pertaining to it may be learned via mining the experience base and via inference rules. Then, we can may find that InheritanceLink Bob_Yifu friend .9,1 (where the .9,1 is an interval probability, interpreted according to the indefinite probabilities framework) and this link mixes intensional and extensional inheritance, and thus is only useful for heuristic reasoning (which however is a very important kind). What this link means is basically that Bob_Yifu's node in the memory has a lot of the same links as the friend node -- or rather, that it **would**, if all its links were allowed to exist rather than being pruned to save memory. So, note that the semantics are actually tied to the mind itself. Or we can make more specialized logical constructs if we really want to, denoting stuff like -- at certain times Bob_Yifu is a friend -- Bob displays some characteristics of friendship very strongly, and others not at all -- etc. We can also do crude, heuristic contextualization like ContextLink .7,.8 home InheritanceLink Bob_Yifu friend which suggests that Bob is less friendly at home than in general. Again this doesn't capture all the subtleties of Bob's friendship in relation to being at home -- and one could do so if one wanted to, but it would require introducing a larger complex of nodes and links, which is not always the most appropriate thing to do. The PLN inference rules are designed to give heuristically correct conclusions based on heuristically interpreted links; or more precise conclusions based on more precisely interpreted links. Finally, the semantics of PLN relationships is explicitly an **experiential** semantics. (One of the early chapters in the PLN book, to appear via Springer next year, is titled Experiential Semantics.) So, all node and link truth values in PLN are intended to be settable and adjustable via experience, rather than via programming or importation from databases or something like that. Now, the above example is of course a quite simple one. Discussing a more complex example would go beyond the scope of what I'm willing to do in an email conversation, but the mechanisms I've described are not limited to such simple examples. I am aware that identifying Bob_Yifu as a coherent, distinct entity is a problem faced by humans and robots, and eliminated via the simplicity of the SL environment. However, there is detailed discussion in the (proprietary) NM book of how these same mechanisms may be used to do object recognition and classification, as well. You may of course argue that these mechanisms won't scale up to large knowledge bases and rich experience streams. I believe that they will, and have arguments but not rigorous proofs that they will. -- Ben G On Nov 13, 2007 12:34 PM, Richard Loosemore
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
On Nov 13, 2007 2:37 PM, Richard Loosemore [EMAIL PROTECTED] wrote: Ben, Unfortunately what you say below is tangential to my point, which is what happens when you reach the stage where you cannot allow any more vagueness or subjective interpretation of the qualifiers, because you have to force the system to do its own grounding, and hence its own interpretation. I don't see why you talk about forcing the system to do its own grounding -- the probabilities in the system are grounded in the first place, as they are calculated based on experience. The system observes, records what it sees, abstracts from it, and chooses actions that it guess will fulfill its goals. Its goals are ultimately grounded in in-built feeling-evaluation routines, measuring stuff like amount of novelty observed, amount of food in system etc. So, the system sees and then acts ... and the concepts it forms and uses are created/used based on their utility in deriving appropriate actions. There is no symbol-grounding problem except in the minds of people who are trying to interpret what the system does, and get confused. Any symbol used within the system, and any probability calculated by the system, are directly grounded in the system's experience. There is nothing vague about an observation like Bob_Yifu was observed at time-stamp 599933322, or a fact Command 'wiggle ear' was sent at time-stamp 54. These perceptions and actions are the root of the probabilities the system calculated, and need no further grounding. What you gave below was a sketch of some more elaborate 'qualifier' mechanisms. But I described the process of generating more and more elaborate qualifier mechanisms in the body of the essay, and said why this process was of no help in resolving the issue. So, if a system can achieve its goals based on choosing procedures that it thinks are likely to achieve its goals, based on the knowledge it gathered via its perceived experience -- why do you think it has a problem? I don't really understand your point, I guess. I thought I did -- I thought your point was that precisely specifying the nature of a conditional probability is a rats-nest of complexity. And my response was basically that in Novamente we don't need to do that, because we define conditional probabilities based on the system's own knowledge-base, i.e. Inheritance A B .8 means If A and B were reasoned about a lot, then A would (as measred by an weighted average) have 80% of the relationships that B does But apparently you were making some other point, which I did not grok, sorry... Anyway, though, Novamente does NOT require logical relations of escalating precision and complexity to carry out reasoning, which is one thing you seemed to be assuming in your post. Ben - 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=8660244id_secret=64644318-8bbdee
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
On Tue, Nov 13, 2007 at 12:34:51PM -0500, Richard Loosemore wrote: Suppose that in some significant part of Novamente there is a representation system that uses probability or likelihood numbers to encode the strength of facts, as in [I like cats](p=0.75). The (p=0.75) is supposed to express the idea that the statement [I like cats] is in some sense 75% true. Either way, we have a problem: a fact like [I like cats](p=0.75) is ungrounded because we have to interpret it. Does it mean that I like cats 75% of the time? That I like 75% of all cats? 75% of each cat? Are the cats that I like always the same ones, or is the chance of an individual cat being liked by me something that changes? Does it mean that I like all cats, but only 75% as much as I like my human family, which I like(p=1.0)? And so on and so on. Eh? You are standing at the proverbial office water coooler, and Aneesh says Wen likes cats. On your drive home, you mind races .. does this mean that Wen is a cat fancier? You were planning on taking Wen out on a date, and this tidbit of information could be useful ... when you try to build the entire grounding mechanism(s) you are forced to become explicit about what these numbers mean, during the process of building a grounding system that you can trust to be doing its job: you cannot create a mechanism that you *know* is constructing sensible p numbers and facts during all of its development *unless* you finally bite the bullet and say what the p numbers really mean, in fully cashed out terms. But has a human, asking Wen out on a date, I don't really know what Wen likes cats ever really meant. It neither prevents me from talking to Wen, or from telling my best buddy that ...well, I know, for instance, that she likes cats... Lack of grounding is what makes humour funny, you can do a whole Pygmalion / Seinfeld episode on she likes cats. --linas - 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=8660244id_secret=64672202-2af80e
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
But has a human, asking Wen out on a date, I don't really know what Wen likes cats ever really meant. It neither prevents me from talking to Wen, or from telling my best buddy that ...well, I know, for instance, that she likes cats... yes, exactly... The NLP statement Wen likes cats is vague in the same way as the Novamente or NARS relationship EvaluationLink likes ListLink Wen cats is vague The vagueness passes straight from NLP into the internal KR, which is how it should be. And that same vagueness may be there if the relationship is learned via inference based on experience, rather than acquired by natural language. I.e., if the above relationship is inferred, it may just mean that {the relationship between Wen and cats} shares many relationships with other person/object relationships that have been categorized as 'liking' before In this case, the system can figure out that Wen likes cats without ever actually making explicit what this means. All it knows is that, whatever it means, it's the same thing that was meant in other circumstances where liking was used as a label. So, vagueness can not only be important into an AI system from natural language, but also propagated around the AI system via inference. This is NOT one of the trickier things about building probabilistic AGI, it's really kind of elementary... -- Ben G - 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=8660244id_secret=64674694-3ada83
Re: Essay - example of how the CSP bites [WAS Re: [agi] What best evidence for fast AI?]
So, vagueness can not only be important imported, I meant into an AI system from natural language, but also propagated around the AI system via inference. This is NOT one of the trickier things about building probabilistic AGI, it's really kind of elementary... -- Ben G - 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=8660244id_secret=64674943-4b25e0