[agi] Interpreting Brain damage experiments
Richard, Did you know, for example, that certain kinds of brain damage can leave a person with the ability to name a visually presented object, but then be unable to pick the object up and move it through space in a way that is consistent with the object's normal use . and that another type of brain damage can result in a person have exactly the opposite problem: they can look at an object and say I have no idea what that is, and yet when you ask them to pick the thing up and do what they would typically do with the object, they pick it up and show every sign that they know exactly what it is for (e.g. object is a key: they say they don't know what it is, but then they pick it up and put it straight into a nearby lock). Now, interpreting that result is not easy, but it does seem to tell us that there are two almost independent systems in the brain that handle vision-for-identification and vision-for-action. That's not exact explanation. In both cases vision module works good. Vision-to-identification works fine in both cases. In this case identified object cannot produce proper actions, because connection with action module was damaged. In another case identified object cannot be resolved into language concept, because connection with language module was damaged. Agree? - 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=73488174-e8e4c8
Re: Human Irrationality [WAS Re: [agi] None of you seem to be able ...]
Mike Tintner wrote: Well, I'm not sure if not doing logic necessarily means a system is irrational, i.e if rationality equates to logic. Any system consistently followed can classify as rational. If for example, a program consistently does Freudian free association and produces nothing but a chain of associations with some connection: bird - - feathers - four..tops or on the contrary, a 'nonsense' chain where there is NO connection.. logic.. sex... ralph .. essence... pi... Loosemore... then it is rational - it consistently follows a system with a set of rules. And the rules could, for argument's sake, specify that every step is illogical - as in breaking established rules of logic - or that steps are alternately logical and illogical. That too would be rational. Neural nets from the little I know are also rational inasmuch as they follow rules. Ditto Hofstadter Johnson-Laird from again the little I know also seem rational - Johnson-Laird's jazz improvisation program from my cursory reading seemed rational and not truly creative. Sorry to be brief, but: This raises all sorts of deep issues about what exactly you would mean by rational. If a bunch of things (computational processes) come together and each contribute something to a decision that results in an output, and the exact output choice depends on so many factors coming together that it would not necessarily be the same output if roughly the same situation occurred another time, and if none of these things looked like a rule of any kind, then would you still call it rational? If the answer is yes then whatever would count as not rational? Richard Loosemore I do not know enough to pass judgment on your system, but you do strike me as a rational kind of guy (although probably philosophically much closer to me than most here as you seem to indicate). Your attitude to emotions seems to me rational, and your belief that you can produce an AGI that will almost definitely be cooperative , also bespeaks rationality. In the final analysis, irrationality = creativity (although I'm using the word with a small c, rather than the social kind, where someone produces a new idea that no one in society has had or published before). If a system can change its approach and rules of reasoning at literally any step of problem-solving, then it is truly crazy/ irrational (think of a crazy path). And it will be capable of producing all the human irrationalities that I listed previously - like not even defining or answering the problem. It will by the same token have the capacity to be truly creative, because it will ipso facto be capable of lateral thinking at any step of problem-solving. Is your system capable of that? Or anything close? Somehow I doubt it, or you'd already be claiming the solution to both AGI and computational creativity. But yes, please do send me your paper. P.S. I hope you won't - I actually don't think - that you will get all pedantic on me like so many AI-ers say ah but we already have programs that can modify their rules. Yes, but they do that according to metarules - they are still basically rulebound. A crazy/ creative program is rulebreaking (and rulecreating) - can break ALL the rules, incl. metarules. Rulebound/rulebreaking is one of the most crucial differences between narrow AI/AGI. Richard: In the same way computer programs are completely neutral and can be used to build systems that are either rational or irrational. My system is not rational in that sense at all. Richard, Out of interest, rather than pursuing the original argument: 1) Who are these programmers/ systembuilders who try to create programs (and what are the programs/ systems) that are either irrational or non-rational (and described as such)? I'm a little partied out right now, so all I have time for is to suggest: Hofstadter's group builds all kinds of programs that do things without logic. Phil Johnson-Laird (and students) used to try to model reasoning ability using systems that did not do logic. All kinds of language processing people use various kinds of neural nets: see my earlier research papers with Gordon Brown et al, as well as folks like Mark Seidenberg, Kim Plunkett etc. Marslen-Wilson and Tyler used something called a Cohort Model to describe some aspects of language. I am just dragging up the name of anyone who has ever done any kind of computer modelling of some aspect of cognition: all of these people do not use systems that do any kind of logical processing. I could go on indefinitely. There are probably hundreds of them. They do not try to build complete systems, of course, just local models. When I have proposed (in different threads) that the mind is not rationally, algorithmically programmed I have been met with uniform and often fierce resistance both on this and another AI forum. Hey, join the club! You have read my little brouhaha with
Re: [agi] Do we need massive computational capabilities?
Sounds like the worst case scenario: computations that need between say 20 and
100 PCs. Too big to run on a very souped up server (4-way Quad processor with
128GB RAM) but to scale up to a 100 Beowulf PC cluster typically means a factor
10 slow-down due to communications (unless it's a
local-data/computation-intensive algorithm) so you actually haven't gained much
in the process. {Except your AGI is now ready for a distributed computing
environment, which I believe luckily Novamenta was explicitely designed for.}
:)
=Jean-Paul
Research Associate: CITANDA
Post-Graduate Section Head
Department of Information Systems
Phone: (+27)-(0)21-6504256
Fax: (+27)-(0)21-6502280
Office: Leslie Commerce 4.21
Benjamin Goertzel [EMAIL PROTECTED] 2007/12/07 15:06
I don't think we need more than hundreds of PCs to deal with these things,
but we need more than a current PC, according to the behavior of our
current algorithms.
-
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=73568490-365c88
Re: Re[2]: [agi] Solution to Grounding problem
Your bot is having a conversation - in words. Words are in fact continually made sense of - grounded - by the human brain - converted into sensory images - and have to be. I've given simple examples of snatches of conversation, which are in fact obviously thus grounded and have to be. The only way a human - or a machine - can make sense of sentence 1 is by referring to a mental image/movie of Bush walking. Merely referring to more words won't cut it. Ditto for sentence 2 - it is essential to refer to an image of Dennis to establish whether he is handsome. If s.o. asks me right now if you are handsome, I can half understand the words, but I can't understand if they are true, because I have never seen you (although I'm sure you're incredibly butch). Ditto with sentence 3, a human or a machine can only really tell whether a person's dialogue is getting emotional, by forming a sensory/sound image of the dialogue on the page and thus of the tone - which you do all the time whether you're aware of it or not. Words and all symbols are totally abstract - if you don't have a sensory image of what they refer to, you can't understand or ground them - that's the grounding problem. Get me a grundchen, Dennis. Meaningless. Ungrounded. But if I show you a picture of a grundchen, you will have no problem knowing what it is, and getting one. Oh, just to make your day, if you don't have a body, you can't understand the images either - because all images have a POV - and are at a distance from an observer - which will take a little more time to explain. That's the extended grounding problem. Is all that clear? If it is, it's grounded. Mike, Was it your explanation of what Grounding Problem is? If it was - you missed the explanation and gave only examples ... Dennis: 1) Grounding Problem (the *real* one, not the cheap substitute that everyone usually thinks of as the symbol grounding problem). Say, we are trying to build AGI for the purpose of running intelligent chat-bot. What would be the grounding problem in this case? Example: understanding: Bush walks like a cowboy, doesn't he? Dennis Gorelik is v. handsome, no? You're getting v. emotional about 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/?; -- No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.503 / Virus Database: 269.16.17/1176 - Release Date: 12/6/2007 11:15 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=73559499-725a29
Re: [agi] Do we need massive computational capabilities?
On Dec 7, 2007 7:09 AM, Mike Tintner [EMAIL PROTECTED] wrote: Matt,:AGI research needs special hardware with massive computational capabilities. Could you give an example or two of the kind of problems that your AGI system(s) will need such massive capabilities to solve? It's so good - in fact, I would argue, essential - to ground these discussions. Problems that would likely go beyond the capability of a current PC to solve in a realistic amount of time, in the current NM architecture, would include for instance: -- Learning a new type of linguistic relationship (in the context of link grammar, this would mean e.g. learning a new grammatical link type) -- Learning a new truth value formula for a probabilistic inference rule -- Recognizing objects in a complex, rapidly-changing visual scene (Not that we have written the code to let the system solve these particular problems yet ... but the architecture should allow it...) I don't think we need more than hundreds of PCs to deal with these things, but we need more than a current PC, according to the behavior of our current algorithms. -- 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=73544012-c56a06
Re: [agi] Do we need massive computational capabilities?
If I had 100 of the highest specification PCs on my desktop today (and it would be a big desk!) linked via a high speed network this wouldn't help me all that much. Provided that I had the right knowledge I think I could produce a proof of concept type AGI on a single PC today, even if it ran like a tortoise. It's the knowledge which is mainly lacking I think. Although I do a lot of stuff with computer vision I find myself not being all that restricted by computational limitations. This certainly wasn't the case a few years ago. Generally even the lowest end hardware these days has enough compute power to do some pretty sophisticated stuff, especially if you include the GPU. - 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=73586802-476a69
Re: [agi] Evidence complexity can be controlled by guiding hands
I have a doubt about role of stochastic variance in this parallel terraced scan as it proceeds in humans (or could proceed with the same functional behavior in AIs). Could it be that low-level mechanisms are not that stochastic and just compute a 'closure' of given context? Closure brings up a specific collection of answer-candidates, and if they are unsatisfactory or if there is time to contemplate some more, deliberation level slightly changes a context by introducing particular bias in it, so that 'closure' gives a different set of answers. Effectively, this process is separated on two levels, where low-level process doesn't work stochastically, and high-level process messes with initial conditions on low-level process, using some kind of ad-hoc pseudorandom generation of biases (for example, based on collection of simple procedures that iterate on available concepts). It certainly feels this way introspectively, and I'm not sure how it can be determined experimentally, probably by delays between phases of this process. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73604704-0ab273
Re: [agi] Interpreting Brain damage experiments
Dennis Gorelik wrote: Richard, Did you know, for example, that certain kinds of brain damage can leave a person with the ability to name a visually presented object, but then be unable to pick the object up and move it through space in a way that is consistent with the object's normal use . and that another type of brain damage can result in a person have exactly the opposite problem: they can look at an object and say I have no idea what that is, and yet when you ask them to pick the thing up and do what they would typically do with the object, they pick it up and show every sign that they know exactly what it is for (e.g. object is a key: they say they don't know what it is, but then they pick it up and put it straight into a nearby lock). Now, interpreting that result is not easy, but it does seem to tell us that there are two almost independent systems in the brain that handle vision-for-identification and vision-for-action. That's not exact explanation. In both cases vision module works good. Vision-to-identification works fine in both cases. In this case identified object cannot produce proper actions, because connection with action module was damaged. In another case identified object cannot be resolved into language concept, because connection with language module was damaged. Agree? I don't think this works, unfortunately, because that was the first simple explanation that people came up with, and it did not match up with the data at all. I confess I do not have time to look this up right now. You wouldn't be able to read one of the latest cognitive neuropsychology books (not cognitive neuroscience, note) and let me know would you? ;-) 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=73608091-c6ef93
Re: [agi] Evidence complexity can be controlled by guiding hands
Vladimir Nesov wrote: I have a doubt about role of stochastic variance in this parallel terraced scan as it proceeds in humans (or could proceed with the same functional behavior in AIs). Could it be that low-level mechanisms are not that stochastic and just compute a 'closure' of given context? Closure brings up a specific collection of answer-candidates, and if they are unsatisfactory or if there is time to contemplate some more, deliberation level slightly changes a context by introducing particular bias in it, so that 'closure' gives a different set of answers. Effectively, this process is separated on two levels, where low-level process doesn't work stochastically, and high-level process messes with initial conditions on low-level process, using some kind of ad-hoc pseudorandom generation of biases (for example, based on collection of simple procedures that iterate on available concepts). It certainly feels this way introspectively, and I'm not sure how it can be determined experimentally, probably by delays between phases of this process. You are asking good questions about the mechanisms, which I am trying to explore emprically. No good answers to this yet, although I have many candidate solutions, some of which (I think) look like your above model. I certainly agree with the sentiment that not *all* of the process can be as fluid as the higher level parts (if that is what you are meaning). 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=73609008-799dfa
Re: [agi] Do we need massive computational capabilities?
On Dec 7, 2007 10:21 AM, Bob Mottram [EMAIL PROTECTED] wrote: If I had 100 of the highest specification PCs on my desktop today (and it would be a big desk!) linked via a high speed network this wouldn't help me all that much. Provided that I had the right knowledge I think I could produce a proof of concept type AGI on a single PC today, even if it ran like a tortoise. It's the knowledge which is mainly lacking I think. I agree that at the moment hardware is NOT the bottleneck. This is why, while we've instrumented the Novamente system to be straightforwardly extensible to a distributed implementation, we haven't done much actual distributed processing implementation yet. We have build commercial systems incorporating the NCE in simple distributed architectures, but haven't gone the distributed-AGI direction yet in practice -- because, as you say, it seems likely that the key AGI problems can be worked out on a single machine, and you can then scale up afterwards. -- 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=73609156-15fdf3
Re: [agi] None of you seem to be able ...
On Dec 6, 2007 8:06 PM, Ed Porter [EMAIL PROTECTED] wrote: Ben, To the extent it is not proprietary, could you please list some of the types of parameters that have to be tuned, and the types, if any, of Loosemore-type complexity problems you envision in Novamente or have experienced with WebMind, in such tuning and elsewhere? Ed Porter A specific list of parameters would have no meaning without a huge explanation which I don't have time to give... Instead I'll list a few random areas where choices need to be made, that appear localized at first but wind up affecting the whole -- attention allocation is handled by an artificial economy mechanism, which has the same sorts of parameters as any economic system (analogues of interest rates, rent rates, etc.) -- program trees representing internal procedures are normalized via a set of normalization rules, which collectively cast procedures into a certain normal form. There are many ways to do this. -- the pruning of (backward and forward chaining) inference trees uses a statistical bandit problem methodology, which requires a priori probabilities to be ascribed to various inference steps Fortunately though in each of the above three examples there is theory that can guide parameter tning (different theories in the three cases -- dynamic systems theory for the artificial economy; formal computer science and language theory for program tree reduction; and Bayesian stats for the pruning issue) Webmind AI Engine had too many parameters and too much coupling between subsystems. We cast parameter optimization as an AI learning problem but it was a hard one, though we did make headway on it. Novamente Engine has much coupling btw subsystems, but no unnecessary coupling; and many fewer parameters on which system behavior can sensitively depend. Definitely, minimization of the number of needful-of-adjustment parameters is a very key aspect of AGI system design. -- 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=73598324-4bf78b
Re: Human Irrationality [WAS Re: [agi] None of you seem to be able ...]
Mike, I think you are going to have to be specific about what you mean by irrational because you mostly just say that all the processes that could possibly exist in computers are rational, and I am wondering what else is there that irrational could possibly mean. I have named many processes that seem to me to fit the irrational definition, but without being too clear about it you have declared them all to be just rational, so now I have no idea what you can be meaning by the word. Richard Loosemore Mike Tintner wrote: Richard:This raises all sorts of deep issues about what exactly you would mean by rational. If a bunch of things (computational processes) come together and each contribute something to a decision that results in an output, and the exact output choice depends on so many factors coming together that it would not necessarily be the same output if roughly the same situation occurred another time, and if none of these things looked like a rule of any kind, then would you still call it rational?If the answer is yes then whatever would count as not rational? I'm not sure what you mean - but this seems consistent with other impressions I've been getting of your thinking. Let me try and cut through this: if science were to change from its prevailing conception of the human mind as a rational, computational machine to what I am suggesting - i.e. a creative, compositional, irrational machine - we would be talking of a major revolution that would impact right through the sciences - and radically extend the scope of scientific investigation into human thought. It would be the end of the deterministic conception of humans and animals and ultimately be a revolution of Darwinian proportions. Hofstadter co are absolutely not revolutionaries. Johnson-Laird conceives of the human mind as an automaton. None of them are fundamentally changing the prevailing conceptions of cognitive science. No one has reacted to them with shock or horror or delight. I suspect that what you are talking about is loosely akin to the ideas of some that quantum mechanics has changed scientific determinism. It hasn't - the fact that we can't measure certain quantum phenomena with precision does not mean that they are not fundamentally deterministic. And science remains deterministic. Similarly, if you make a computer system very complex, keep changing the factors involved in computations, add random factors whatever, you are not necessarily making it non-rational. You make it v. difficult to understand the computer's rationality, (and possibly extend our conception of rationality), but the system may still be basically rational, just as quantum particles are still in all probability basically deterministic. As a side-issue, I don't believe that human reasoning, conscious and unconscious, is remotely, even infinitesimally as complex as that of the AI systems you guys all seem to be building. The human brain surely never seizes up with the kind of complex, runaway calculations that y'all have been conjuring up in your arguments. That only happens when you have a rational system that obeys basically rigid (even if complex) rules. The human brain is cleverer than that - it doesn't have any definite rules for any activities. In fact, you should be so lucky as to have a nice, convenient set of rules, even complex ones, to guide you when you sit down to write your computer programs. - 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=73610112-93352e
Re: [agi] None of you seem to be able ...
Jean-Paul Van Belle wrote: Interesting - after drafting three replies I have come to realize that it is possible to hold two contradictory views and live or even run with it. Looking at their writings, both Ben Richard know damn well what complexity means and entails for AGI. Intuitively, I side with Richard's stance that, if the current state of 'the new kind of science' cannot even understand simple chaotic systems - the toy-problems of three-variable differential quadratic equations and 2-D Alife, then what hope is there to find a theoretical solution for a really complex system. The way forward is by experimental exploration of part of the solution space. I don't think we'll find general complexity theories any time soon. On the other hand, practically I think that it *is* (or may be) possible to build an AGI system up carefully and systematically from the ground up i.e. inspired by a sound (or at least plausible) theoretical framework or by modelling it on real-world complex systems that seem to work (because that's the way I proceed too), finetuning the system parameters and managing emerging complexity as we go along and move up the complexity scale. (Just like engineers can build pretty much anything without having a GUT.) Both paradagmatic approaches have their merits and are in fact complementary: explore, simulate, genetically evolve etc. from the top down to get a bird's eye view of the problem space versus incrementally build up from the bottom up following a carefully chartered path/ridge inbetween the chasms of the unknown based on a strong conceptual theoretical founding. It is done all the time in other sciences - even maths! Interestingly, I started out wanting to use a simulation tool to check the behaviour (read: fine-tune the parameters) of my architectural designs but then realised that the simulation of a complex system is actually a complex system itself and it'd be easier and more efficient to prototype than to simulate. But that's just because of the nature of my architecture. Assuming Ben's theories hold, he is adopting the right approach. Given Richard's assumption or intuitions, he is following the right path too. I doubt that they will converge on a common solution but the space of conceivably possible AGI architectures is IMHO extremely large. In fact, my architectural approach is a bit of a poor cousin/hybrid: having neither Richard's engineering skills nor Ben's mathematical understanding I am hoping to do a scruffy alternative path :) Interesting thoughts: remind me, if I forget, that when I get my website functioning and can put longer papers into a permanent repository, that we all need to have a forward-looking discussion about some of the detailed issues that might arise here. That is, going beyond merely arguing about whether or not there is a problem. I have many thoughts about what you say, but no time right now, so I will come back to this. The short version of my thoughts is that we need to look into some of the details of what I propose to do, and try to evaluate the possible dangers of not taking the path I suggest. 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=73591687-f58813
Re: [agi] Solution to Grounding problem
Dennis Gorelik wrote: Richard, It seems that under Real Grounding Problem you mean Communication Problem. Basically your goal is to make sure that when two systems communicate with each other -- they understand each other correctly. Right? If that's the problem -- I'm ready to give you my solution. BTW, I had to read your explanation 3 times to get it [if I got it]. :-) Don't feel bad: my explanation was horribly compressed, and not necessarily very well articulated, and the actual claim is extremely abstract and susceptible to misinterpretation (about 95% of the literature on the SGP is a complete misinterpretation!). I don't think it is quite a communication problem, though. The issue is much more like the error that destroyed that NASA Mars spacecraft several years ago (can't remember which one: they busted so many of them). The one that had one software module calculating in kilometers and the other module calculating in miles, so the results passed from one to the other became meaningless. This could be called a communcation problem, but it is internal, and in the AGI case it is not so simple as just miscalculated numbers. So here is a revised version of the problem: suppose that a system keeps some numbers stored internally, but those numbers are *used* by the system in such a way that their meaning is implicit in the entire design of the system. When the system uses those numbers to do things, the numbers are fed into the using mechanisms in such a way that you can only really tell what the numbers mean by looking at the overall way in which they are used. Now, with that idea in mind, now imagine that programmers came along and set up the *values* for a whole bunch of those numbers, inside the machine, ON THE ASSUMPTION that those numbers meant something that the programmers had decided they meant. So the programmers were really definite and explicit about the meaning of the numbers. Question: what if those two sets of meanings are in conflict? This is effectively what the SGP (symbol grounding problem) is all about. Some AI folks start out by building a program in which they decide ahead of time what the symbols mean, and they insert a whole bunch of actual symbols (AND mechanisms that operate on symbols) into the system on the assumption that their chosen meanings are valid. This becomes a problem because when we say of another person that they meant something by their use of a particular word (say cat), what we actually mean is that that person had a huge amount of cognitive machinery connected to that word cat (reaching all the way down to the sensory perception mechanisms that allow the person to recognise an instance of a cat, and motor output mechanisms that let them interact with a cat). What Stephen Harnad said in his original paper was Hang on a second: if the AI system does not have all that other machinery inside it when it uses a word like cat, surely it does not really mean the same thing by cat as a person would? In effect, he was saying that the very limited machinery inside a simple AI system will have an *implicit* meaning for cat which is very crude because it does not have all that other stuff that we have inside our heads, connected to the cat concept. When you ask the AI Are cats fussy? it will only be able to do something crude like see if it has a memory item recording a fact about cats and fussiness. A person on the other hand (if they know cats) will be able to deploy a huge amount of knowledge about both the [cat] concept and the [fussy] concept, and come to a sophisticated conclusion. What Harnad would say is that the AI does not really have the same meaning attached to cat as people do. He then went on to say that the only way to resolve this problem is to make sure that the system is connected to the real world so it can pick up its own symbols, and only when it has all that real-world connection machinery, and building symbols in the way that we do, will the system really be able to get the meaning of a word like cat. Harnad summarized that by saying that AI systems need to have their symbols grounded in the real world. Now this is where the confusion starts. Lots of people heard him suggest this, and then thought: No problem: we'll attach some video cameras and robot arms to our AI and then it will be grounded! This is a disatrous misunderstanding of the problem. If the AI system starts out with a design in which symbols are designed and stocked by programmers, this part of the machine has ONE implicit meaning for its symbols . but then if a bunch of peripheral machinery is stapled on the back end of the system, enabling it see the world and use robot arms, the processing and symbol building that goes on in that part of the system will have ANOTHER implicit meaning for the symbols. There is no reason why these two sets of symbols should have the same meaning!
Re: [agi] Do we need massive computational capabilities?
Matt,:AGI research needs special hardware with massive computational capabilities. Could you give an example or two of the kind of problems that your AGI system(s) will need such massive capabilities to solve? It's so good - in fact, I would argue, essential - to ground these discussions. - 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=73537387-8fc58e
Re: [agi] Evidence complexity can be controlled by guiding hands
Ed Porter wrote: RICHARD LOOSEMORE= At the cognitive level, on the other hand, there is a strong possibility that what happens when the mind builds a model of some situation, it gets a large nummber of concepts to come together and try to relax into a stable representation, and that relaxation process is potentially sensitive to complex effects (some small parameter in the design of the concepts could play a crucial role in ensuring that the relaxation process goes properly, for example) ED PORTER= Copycat uses a variant of simulated annealing to do its relaxation process, except it is actually a much more chaotic relaxation process than many (e.g., much more than Hecht-Neilsen's Confabulation), because it involves millions of separate codlets being generated to score, decide the value of, and to add or remove elements from a graph, that labels grouping and relationships in the initial string, and between the example initial string and the solution initial string, and between the example initial string and the example changed string, and between the both the solution initial string and the example changed string and the solution changed string, as well as constructing the solution changed string itself during this process. Each of the labelings and mapping links is made by a separate small program called a codelet. Codelets are chosen in a weighted random manner. And one codelet can clobber the work done by another. The ratio of importance between some fitness weighting and pure randomness in the picking of codlets varyies with temperature, which is a measure of overall labeling, mapping, and solution fit, which tends to go down over time as the system moves toward a coherent solution. But it can go up if the system starts settling into a solution that creates a mapping or labeling flaw, at which time more random codelets will be created and randomly change the system, but with the changes being more likely in the parts of the graph or labeling that have the least good fit, and thus requires the least energy to kick apart. Despite this very chaotic process, and the fact this process is sensitive to complex dynamic effects that enable a slight change of state to causes it to settle into different solutions, as Richard mentioned above, the weighting of the system, which varies dynamically in a context sensitive way, causes most of the solutions that it settles into to be appropriate, although they may be quite different. For example, for the copycat problem where the goal is to change ijkk in a manner similar to that in which aabc was changed to produce aabd, which problem can be represented as ex aabc -- aabd ijkk -- ? On one thousand runs the results were # of occurrence result temperature 1 612 wereijll29 2 198 wereijkl49 3 121 werejjkk47 4 47 werehjkk19 5 9 werejkkk42 6 6 wereijkd57 7 3 wereijdd46 8 3 wereijkk69 9 1 was djkk58 ===EXPLANATION OF ANALOGY IN EACH SOLUTION=== ex-last char in string has alphabet number incremented 1-last set of the same chars in each string had alphabet number incremented 2-last char in each string had alphabet number incremented 3-one end char in each string had alphabet number incremented 4-one end char in each string had alphabet number changed by one 5-set of chars in string had alphabet numbers incremented 6-last char in each string is changed to d 7-last set of same chars in each initial string was changed to d 8-last char in each string had alphabet number changed by a value of zero or one 9-one char on end of string was changed to d So you see that each of the changes except solution 8, which had the worst temperature, meaning the system felt it was the worst fit actually captured an analogous change. If temperature were used to filter out the misfits, none of the runs would have produced a non-analogy. So despite the chaotic nature of the system, it almost always settled on a labeling, graphing, and solution that was appropriate, and when it didn't it knew it didn't, because of the systems measure of analogical fit. Although this definitely is a toy problem, it might have as much potential for complexity as the game of life, in terms of its number of components (if you count its codlets), its computations, and its non-linearities. I was told by somebody who worked with Hofstader that individual copycat solutions running on unoptimized LISP code on roughly 1990s Sun work stations normally took between about half hour to a major fraction of a day. The difference between this and the game of life is that has been designed to work. Despite its somewhat chaotic manner of
Re: [agi] Do we need massive computational capabilities?
Thanks. And I repeat my question elsewhere : you don't think that the human brain which does this in say half a second, (right?), is using massive computation to recognize that face? You guys with all your mathematical calculations re the brain's total neurons and speed of processing surely should be able to put ball-park figures on the maximum amount of processing that the brain can do here. Hawkins argues: neurons are slow, so in that half a second, the information entering your brain can only traverse a chain ONE HUNDRED neurons long. ..the brain 'computes' solutions to problems like this in one hundred steps or fewer, regardless of how many total neurons might be involved. From the moment light enters your eye to the time you [recognize the image], a chain no longer than one hundred neurons could be involved. A digital computer attempting to solve the same problem would take BILLIONS of steps. One hundred computer instructions are barely enough to move a single character on the computer's display, let alone do something interesting. IOW, if that's true, the massive computational approach is surely RIDICULOUS - a grotesque travesty of engineering principles of economy, no? Like using an entire superindustry of people to make a single nut? And, of course, it still doesn't work. Because you just don't understand how perception works in the first place. Oh right... so let's make our computational capabilities even more massive, right? Really, really massive. No, no, even bigger than that? Matt,:AGI research needs special hardware with massive computational capabilities. Could you give an example or two of the kind of problems that your AGI system(s) will need such massive capabilities to solve? It's so good - in fact, I would argue, essential - to ground these discussions. For example, I ask the computer who is this? and attach a video clip from my security camera. - 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=73724842-42226f
Re: Human Irrationality [WAS Re: [agi] None of you seem to be able ...]
Mike Tintner wrote: Richard: Mike, I think you are going to have to be specific about what you mean by irrational because you mostly just say that all the processes that could possibly exist in computers are rational, and I am wondering what else is there that irrational could possibly mean. I have named many processes that seem to me to fit the irrational definition, but without being too clear about it you have declared them all to be just rational, so now I have no idea what you can be meaning by the word. Richard, Er, it helps to read my posts. From my penultimate post to you: If a system can change its approach and rules of reasoning at literally any step of problem-solving, then it is truly crazy/ irrational (think of a crazy path). And it will be capable of producing all the human irrationalities that I listed previously - like not even defining or answering the problem. It will by the same token have the capacity to be truly creative, because it will ipso facto be capable of lateral thinking at any step of problem-solving. Is your system capable of that? Or anything close? Somehow I doubt it, or you'd already be claiming the solution to both AGI and computational creativity. A rational system follows a set of rules in solving a problem (which can incl. rules that self-modify according to metarules) ; a creative, irrational system can change/break/create any and all rules (incl. metarules) at any point of solving a problem - the ultimate, by definition, in adaptivity. (Much like you, and indeed all of us, change the rules of engagement much of the time in our discussions here). Listen, no need to reply - because you're obviously not really interested. To me that's ironic, though, because this is absolutely the most central issue there is in AGI. But no matter. No, I am interested, I was just confused, and I did indeed miss the above definition (got a lot I have to do right now, so am going very fast through my postings) -- sorry about that. The fact is that the computational models I mentioned (those by Hofstadter etc) are all just attempts to understand part of the problem of how a cognitive system works, and all of them are consistent with the design of a system that is irrational accroding to your above definition. They may look rational, but that is just an illusion: every one of them is so small that it is completely neutral with respect to the rationality of a complete system. They could be used by someone who wanted to build a rational system or an irrational system, it does not matter. For my own system (and for Hofstadter too), the natural extension of the system to a full AGI design would involve a system [that] can change its approach and rules of reasoning at literally any step of problem-solving it will be capable of producing all the human irrationalities that I listed previously - like not even defining or answering the problem. It will by the same token have the capacity to be truly creative, because it will ipso facto be capable of lateral thinking at any step of problem-solving. This is very VERY much part of the design. I prefer not to use the term irrational to describe it (because that has other connotations), but using your definition, it would be irrational. There is not any problem with doing all of this. Does this clarify the question? I think really I would reflect the question back at you and ask why you would think that this is a difficult thing to do? It is not difficult to design a system this way: some people like the trad-AI folks don't do it (yet), and appear not to be trying, but there is nothing in principle that makes it difficult to build a system of this sort. 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=73685934-1acb8b
Re: Human Irrationality [WAS Re: [agi] None of you seem to be able ...]
Richard: Mike, I think you are going to have to be specific about what you mean by irrational because you mostly just say that all the processes that could possibly exist in computers are rational, and I am wondering what else is there that irrational could possibly mean. I have named many processes that seem to me to fit the irrational definition, but without being too clear about it you have declared them all to be just rational, so now I have no idea what you can be meaning by the word. Richard, Er, it helps to read my posts. From my penultimate post to you: If a system can change its approach and rules of reasoning at literally any step of problem-solving, then it is truly crazy/ irrational (think of a crazy path). And it will be capable of producing all the human irrationalities that I listed previously - like not even defining or answering the problem. It will by the same token have the capacity to be truly creative, because it will ipso facto be capable of lateral thinking at any step of problem-solving. Is your system capable of that? Or anything close? Somehow I doubt it, or you'd already be claiming the solution to both AGI and computational creativity. A rational system follows a set of rules in solving a problem (which can incl. rules that self-modify according to metarules) ; a creative, irrational system can change/break/create any and all rules (incl. metarules) at any point of solving a problem - the ultimate, by definition, in adaptivity. (Much like you, and indeed all of us, change the rules of engagement much of the time in our discussions here). Listen, no need to reply - because you're obviously not really interested. To me that's ironic, though, because this is absolutely the most central issue there is in AGI. But no matter. - 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=73661748-adcbd5
Re: [agi] Do we need massive computational capabilities?
--- Mike Tintner [EMAIL PROTECTED] wrote: Thanks. And I repeat my question elsewhere : you don't think that the human brain which does this in say half a second, (right?), is using massive computation to recognize that face? So if I give you a video clip then you can match the person in the video to the correct photo out of 10^9 choices on the Internet in 0.5 seconds, and this will all run on your PC? Let me know when your program is finished so I can try it out. You guys with all your mathematical calculations re the brain's total neurons and speed of processing surely should be able to put ball-park figures on the maximum amount of processing that the brain can do here. Hawkins argues: neurons are slow, so in that half a second, the information entering your brain can only traverse a chain ONE HUNDRED neurons long. ..the brain 'computes' solutions to problems like this in one hundred steps or fewer, regardless of how many total neurons might be involved. From the moment light enters your eye to the time you [recognize the image], a chain no longer than one hundred neurons could be involved. A digital computer attempting to solve the same problem would take BILLIONS of steps. One hundred computer instructions are barely enough to move a single character on the computer's display, let alone do something interesting. Which is why the human brain is so bad at arithmetic and other tasks that require long chains of sequential steps. But somehow it can match a face to a name in 0.5 seconds. Neurons run in PARALLEL. Your PC does not. Your brain performs 10^11 weighted sums of 10^15 values in 0.1 seconds. Your PC will not. IOW, if that's true, the massive computational approach is surely RIDICULOUS - a grotesque travesty of engineering principles of economy, no? Like using an entire superindustry of people to make a single nut? And, of course, it still doesn't work. Because you just don't understand how perception works in the first place. Oh right... so let's make our computational capabilities even more massive, right? Really, really massive. No, no, even bigger than that? Matt,:AGI research needs special hardware with massive computational capabilities. Could you give an example or two of the kind of problems that your AGI system(s) will need such massive capabilities to solve? It's so good - in fact, I would argue, essential - to ground these discussions. For example, I ask the computer who is this? and attach a video clip from my security camera. -- Matt Mahoney, [EMAIL PROTECTED] - 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=73765756-f02c55
RE: [agi] Evidence complexity can be controlled by guiding hands
Richard, With regard to your below post: RICHARD LOOSEMORE ###Allowing the system to adapt to the world by giving it flexible mechanisms that *build* mechanisms (which it then uses), is one way to get the system to do some of the work of fitting parameters (as ben would label it), or reducing the number of degrees of freedom that we have to deal with. But that would be different from *our* efforts, as designers of the system, to design different possible mechanisms, then do tests to establish what kind of system behavior they cause. We have to do this generate and test experimentation in parallel with the system's own attempts to adapt and build new internal mechanisms. They are two different processes, both of which are designed to home in on the best design for an AGI, and they do need to be considered separately. ED PORTER ### I don't understand in exactly what ways you think the experientially learned and the designed features should be treated differently, and how this relates to the potential pitfalls of complexity. Of course they would normally be considered differently (you have to directly design one, the other is learned automatically by a system you design). I think there needs to be joint development of them, because the designed mechanisms are intended to work with the learned ones, and vice versa. In the system I have been thinking of, most of the experientially learned patterns are largely drawn from, or synthesized from, recorded experience in a relatively direct manner, not from some sort of Genetic Algorithm that searches large spaces to find some algorithm which compactly represents large amounts of experiences. This close connection with sensed, behaved, or thought experience tends to make such systems more stable. But it is not clear to me that all experientially learned things are necessarily more safe than designed things. For example, Novamente uses MOSES, which is a genetic algorithm learning tool. I think such a tool is not directly needed for an AGI and probably has no direct analogy in the brain. I think the brain uses something that is a rough combination of copycats type of relaxation type assembly, with something like the superimposed probabilities of hecht-neilsen's confabulation to explore new problem spaces, and that this process is repeated over and over again when trying to solve complex problems with the various good features of successive attempts being remembered as part of an increasing learned vocabulary of patterns from which new synthesis are more likely to be performed (all of which is arguably an analog of GA. I can, however, understand how a Genetic algorithm like MOSES could add tremendous learning, exploratory, and perhaps even representational power to an AGI, particularly for certain classes of problems. BUT I HAVE LITTLE UNDERSTANDING FOR EXACTLY WHAT TYPE OF COMPLEXITY DANGERS SUCH GENETIC ALGORITHM PRESENTS. GAs have been successfully used for multiple purposes, particularly where one has a clearly defined and measurable fitness function. But it is not clear to me what happens if you use GAs to control an AGI's relatively high levels of behavior in a complex environment for which there would often not be any simply applicable fitness function. Nor is it clear to me what happens if you have large number of GA controlled systems interacting with each other. It would seem to me they would have much more potential for knarliness than my more experientially based learning systems, but I really don't know. Ben would probably know much more about this than most. RICHARD LOOSEMORE ###The other major comment that I have is that the *main* strategy that I have for reducing the number of degrees of freedom (in the design) is to keep the design as close as possible to the human cognitive system. This is where my approach and the Novamente approach part company in a serious way. I believe that the human design has already explored the space of possible solutions for us (strictly speaking it is evolution that did the exploration when it tried out all kinds of brain edsigns over the eons). I believe that this will enable us to drastically reduce the number of possibilities we have to explore, thus making the project feasible. My problem is that it may be tempting to see a ground-up AGI design (in which we just get a little inspiration from the human system, but mostly we ignore it) as just as feasible when in fact it may well get bogged down in dead ends within the space of possible AGI designs. ED PORTER ### You might be right, you might be wrong. It is my intuition that you do not need to reverse engineer the human brain to build AGI's. I think some of the types of design mistakes you envision from not waiting until we get the whole picture on how the brain works, will probability require some significant software revisions, but such revisions are common in development of complex systems of a new type. I think we
RE: Distributed search (was RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research])
Hi Matt, Wonderful idea, now it will even show the typical human trait of lying...when i ask it do you still love me? most answers in its database will have Yes as an answer but when i ask it 'what's my name?' it'll call me John? However, your approach is actually already being implemented to a certain extent. Apparantly (was it newsweek, time?) the No 1 search engine in (Singapore? Hong Kong? Taiwan? - sorry I forgot) is *not* Google but a local language QA system that works very much the way you envisage it (except it collects the answers in its own SAN i.e. not distributed over the user machines) =Jean-Paul On 2007/12/07 at 18:58, in message [EMAIL PROTECTED], Matt Mahoney [EMAIL PROTECTED] wrote: Hi Matt You call it an AGI proposal but it is described as a distributed search algorithms that (merely) appears intelligent i.e. design for an Internet-wide message posting and search service. There doesn't appear to be any grounding or semantic interpretation by the AI system? How will it become more intelligent? Turing was careful to make no distinction between being intelligent and appearing intelligent. The requirement for passing the Turing test is to be able to compute a probability distribution P over text strings that varies from the true distribution no more than it varies between different people. Once you can do this, then given a question Q, you can compute answer A that maximizes P(A|Q) = P(QA)/P(Q). This does not require grounding. The way my system appears intelligent is by directing Q to the right experts, and by being big enough to have experts on nearly every conceivable topic of interest to humans. A lot of AGI research seems to be focused on how to represent knowledge and thought efficiently on a (much too small) computer, rather than on what services the AGI should provide for us. -- Research Associate: CITANDA Post-Graduate Section Head Department of Information Systems Phone: (+27)-(0)21-6504256 Fax: (+27)-(0)21-6502280 Office: Leslie Commerce 4.21 - 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=73912948-7bb204
Re: Re[2]: [agi] Do we need massive computational capabilities?
On Dec 7, 2007 7:41 PM, Dennis Gorelik [EMAIL PROTECTED] wrote: No, my proposal requires lots of regular PCs with regular network connections. Properly connected set of regular PCs would usually have way more power than regular PC. That makes your hardware request special. My point is - AGI can successfully run on singe regular PC. Special hardware would be required later, when you try to scale out working AGI prototype. I believe Matt's proposal is not as much about the exposure to memory or sheer computational horsepower - it's about access to learning experience. A supercomputer atop an ivory tower (or in the deepest government sub-basement) has an immense memory and speed (and dense mesh of interconnects, etc., etc.) - but without interaction from outside itself, it's really just a powerful navel-gazer. Trees do not first grow a thick trunk and deep roots, then change to growing leaves to capture sunlight. As I see it, each node in Matt's proposed network enables IO to the us [existing examples of intelligence/teachers]. Maybe these nodes can ask questions, What does my owner know of A? - the answer becomes part of its local KB. Hundreds of distributed agents are now able to query Matt's node about A (clearly Matt does not have time to answer 500 queries on topic A) During the course of processing the local KB on topic A, there is a reference to topic B. Matt's node automatically queries every node that previously asked about topic A (seeking first likely authority on the inference) - My node asks me, What do you know of B? Is A-B? I contribute to my node's local KB, and it weights the inference for A-B. This answer is returned to Matt's node (among potentially hundreds of other relative weights) and Matt's node strengthen the A-B inference based on received responses. At this point, the distribution of weights for A-B are all over the network depending on the local KB of each node and the historical traffic of query/answer flow. After some time, I ask my node about topic C. It knows nothing of topic C, so it asks me directly to deposit information to the local KB (initial context) - through the course of 'conversation' with other nodes, my answer comes back as the aggregate of the P2P knowledge within a query radius. On a simple question I may only allow 1 hour of think time, for a deeper research project that radius of query may be allowed to extend 2 weeks of interconnect. During my research, my node will necessarily become interested in topic C - and will likely become known among the network as the local expert. (local expert for a topic would be a useful designation to weigh each node for primary query targets as well as 'trusting' the weight of the answers from each node) I don't think this is vastly different from how people (as working examples of intelligence nodes) gather knowledge from peers. Perhaps this approach to intelligence is not an absolute definition as much as a best effort/most useful answer to date intention. Even if this schema does not extend to emergent AGI, it builds a useful infrastructure that can be utilized by currently existing intelligences as well as whatever AGI does eventually come into existence. Matt, is this coherent with your view or am I off base? - 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=73898638-6a4fad
Re: Re[2]: [agi] Interpreting Brain damage experiments
Hippocampus damage and resulting learning deficiencies are very interesting phenomena. They probably show how important high-level control of learning is in efficient memorization, particularly in memorization of regularities that are presented only few times (or just once, as in the case of episodic memories) and are successfully memorized by healthy people but not by people with damaged hippocampus. People with damaged hippocampus are still able to memorize regularities that pass sufficiently many times through their perception (which is how low-level subsystems probably learn normally). They can compensate for regularities that they can deliberatively recite, like text, but not whole episodic memories. It shows a limitation of Hebbian learning, of balance between gathering information about regularity and applying it to reinforce the regularity, and of importance of high-level mechanism that is able to compensate for this property. This I think can be a useful observation for AGI design. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73897738-7ea5fd
[agi] High-level brain design patterns
Derek, Low level design is not critical for AGI. Instead we observe high level brain patterns and try to implement them on top of our own, more understandable, low level design. I am curious what you mean by high level brain patterns though. Could you give an example? 1) All dependencies we may observe between inputs or outputs. For example, conditional reflex and unconditional reflex. 2) Activation of neuron A that happens _consistently_ with activation of neuron B. 3) Richard Loosemore already gave his example: http://www.dennisgorelik.com/ai/2007/12/reducing-agi-complexity-copy-only-high.html For example, much to our surprise we might see waves in the U values. And every time two waves hit each other, a vortex is created for exactly 20 minutes, then it stops. - 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=73895601-d02434
Re[2]: [agi] Interpreting Brain damage experiments
Richard, Let's save both of us time and wait when somebody else read this Cognitive Science book and will come here to discuss it. :-) Though interesting, interpreting Brain damage experiments is not the most important thing for AGI development. In both cases vision module works good. Vision-to-identification works fine in both cases. In this case identified object cannot produce proper actions, because connection with action module was damaged. In another case identified object cannot be resolved into language concept, because connection with language module was damaged. Agree? I don't think this works, unfortunately, because that was the first simple explanation that people came up with, and it did not match up with the data at all. I confess I do not have time to look this up right now. You wouldn't be able to read one of the latest cognitive neuropsychology books (not cognitive neuroscience, note) and let me know would you? ;-) - 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=73890290-7fa0bf
Re[2]: [agi] Solution to Grounding problem
Richard, This could be called a communcation problem, but it is internal, and in the AGI case it is not so simple as just miscalculated numbers. Communication between subsystems is still communication. So I suggest to call it Communication problem. So here is a revised version of the problem: suppose that a system keeps some numbers stored internally, but those numbers are *used* by the system in such a way that their meaning is implicit in the entire design of the system. When the system uses those numbers to do things, the numbers are fed into the using mechanisms in such a way that you can only really tell what the numbers mean by looking at the overall way in which they are used. That's right approach of doing things. Concepts gaining meaning by connecting to other concepts. The only exception - concepts that are directly connected to hardcoded sub-systems (dictionary, chat client, web browser, etc). Such directly connected concepts would have some predefined meaning. This predefined meaning would be injected by AGI programmers. Now, with that idea in mind, now imagine that programmers came along and set up the *values* for a whole bunch of those numbers, inside the machine, ON THE ASSUMPTION that those numbers meant something that the programmers had decided they meant. So the programmers were really definite and explicit about the meaning of the numbers. Question: what if those two sets of meanings are in conflict? How could they be in conflict, if one set is predefined, and another set gained meaning from predefined set? If you are talking about inconsistencies within predefined set -- that's problem of design development team. Do you want to address this problem? So far I can suggest one tip: keep the set of predefined concept as small as possible. Most of mature AGI intelligence should come from concepts (and their relations) acquired during system life time. If the AI system starts out with a design in which symbols are designed and stocked by programmers, this part of the machine has ONE implicit meaning for its symbols . but then if a bunch of peripheral machinery is stapled on the back end of the system, enabling it see the world and use robot arms, the processing and symbol building that goes on in that part of the system will have ANOTHER implicit meaning for the symbols. There is no reason why these two sets of symbols should have the same meaning! Here's my understanding of your problem: We have an AGI, and now we want to extend it by adding new module. We afraid that new module will have problems communicating with other modules, because the meaning of some symbols is different. If I understood your correctly, here're two solutions: Solution #1: Connect modules through Neural Net. Under Neural Net I mean set of concepts (nodes) connected with other concepts by relations. Concepts can be created and deleted dynamically. Relations can be created and deleted dynamically. When we connect new module to the system - it will introduce its own concepts into Neural Net. Initially these concepts are not connected with existing concepts. But then some process will connect these new concepts with existing concepts. One example of such process could be: if concepts are active at the same time -- connect them. There could be other possible connecting processes. In any case, eventually system would connect all new concepts, and that connections would define how input from new module is interpreted by the rest of the system. Solution #2: Connect new module into another hardcoded modules directly. In this case it's responsibility of AGI development team to make sure that both hardcoded modules talk the same language. That's typical module integration task for developers. In fact, it turns out (when you think about it a little longer) that all of the problem has to do with the programmers going in and building any symbols using THEIR idea of what the symbols should mean: the system has to be allowed to build its own symbols from the ground up, without us necessarily being able to interpret those symbols completely at all. We might nevcer be able to go in and look at a system-built symbol and say That means [x], because the real meaning of that symbol will be implicit in the way the system uses it. In summary: the symbol grounding problem is that systems need to have only one interpretation of their symbols, Not sure what you mean by one interpretation. Symbol can be have multiple interpretations in different contexts. Our goal is to make sure that different systems and different modules has ~same understanding of the symbols at the time of communication. (Under symbols here I mean data that is passed through interfaces) and it needs to be the one built by the system itself as a result of a connection to the external world. So it seems you already have a solution (I propose the same solution) to the Real Grounding Problem. Can
Re: [agi] AGI communities and support
On Dec 8, 2007 2:10 AM, Ed Porter [EMAIL PROTECTED] wrote: Vlad, The Russians have traditionally had more than their share of math whizzes, so I am surprised there isn't more interest in this subject there. I don't understand I wonder where your question has a positive answer and how it can look like. Perhaps you mean, you wonder where one would be able to positively answer such a question. The answer to that is that I know of no place that is funding AGI, proper, like you think they would. The US government is funding a fair amount of traditional AI, but not yet real AGI. Yes, it's what I mean. But communities can exist irrespective of funding, like this one and in previous years on SL4. I know of no other online community that is focused on AGI (although I speak only Russian and English, so there can be some in other languages; Japanese, anyone?) -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73868031-fb0782
RE: [agi] AGI communities and support
Vlad, The Russians have traditionally had more than their share of math whizzes, so I am surprised there isn't more interest in this subject there. I don't understand I wonder where your question has a positive answer and how it can look like. Perhaps you mean, you wonder where one would be able to positively answer such a question. The answer to that is that I know of no place that is funding AGI, proper, like you think they would. The US government is funding a fair amount of traditional AI, but not yet real AGI. Of course we do not know what might be being done in secret in various countries' militaries, intelligence services, or corporations. We might wake up tomorrow to find out the Google already has one up and running. You never know. Ed Porter -Original Message- From: Vladimir Nesov [mailto:[EMAIL PROTECTED] Sent: Friday, December 07, 2007 5:35 PM To: agi@v2.listbox.com Subject: Re: [agi] AGI communities and support On Dec 8, 2007 1:08 AM, Ed Porter [EMAIL PROTECTED] wrote: Vlad, What country are you in? And what is the level of web-comunity, academic, commercial, and governmental support AGI in your country? Ed Porter I live in Moscow. AGI-related activities are nonexistent here; there's a small web community, but I don't follow its discussions, archives for recent years don't show anything interesting. I wonder where your question has a positive answer and how it can look like. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73864995-76310d
RE: [agi] Do we need massive computational capabilities?
Mike Tintner # Yes, I understood that (though sure, I'm capable of misunderstanding anything here!) ED PORTER # Great, I am glad you understood this. Part of what you said indicated you did. BTW, we are all capable of misunderstanding things. Mike Tintner # Hawkins' basic point that the brain isn't a computer at all - which I think can be read less controversially as is a machine that works on very fundamentally different principles to those of currently programmed computers - especially when perceiving objects - holds. You're not dealing with that basic point, and I find it incredibly difficult to get anyone here squarely to face it. People retreat into numbers and millions. ED PORTER # I think most of us understand that and are not disputing it. A Novamente-like approach to AGI is actually quite similar to Hawkins' in many way. For example, it uses hierarchical representation. So few of us are talking about Old Fashioned AI as the major architecture for our systems (although OFAI has its uses in certain areas). Mike Tintner # P.S. You also don't answer my question re: how many neurons in total *can* be activated within a half second, or given period, to work on a given problem - given their relative slowness of communication? Is it indeed possible for hundreds of millions of messages about that one subject to be passed among millions of neurons in that short space (dunno-just asking)? Or did you pluck that figure out of the air? ED PORTER # I was not aware I had been asked this question. If you are asking where I got the it-probably-takes-hundreds-of-millions-of-steps-to-recognize-a-face, I was sort of picking it out of the air, but I don't think it is unreasonable pick. I was including each synaptic transmission as a step. Assume the average neuron has roughly 1K active synapses (some people say several thousand some say only about 100), and lets say an active cell fires at least ten times during the 100 step process, and since you assume 100 levels of activation, that would only be assuming an average of 100 neurons activated on average at each of your 100 levels, which is not a terribly broad search. If a face were focused on, so that it took up just the size of your thumbnail with you thumb sticking up with your arm extended fully in front of your that would activate a portion of your foviated retna having a resolution of roughly 10k pixels (if I recollect correctly from a conversation with Tomaso Poggio). Presumably this would include 3 color inputs, a BW input, and with mangocellur and pravocellur inputs from each eye, so you may well be talking about 100k neurons actived at just the V1 level. If each has 1K neurons firing 10 times, that's 10k x 100K or 100M synaptic firings right there, in just one of your 100 steps. Now some of that activation might be filtered out by the thalamus, but then you would have to include all its activations used for such filtering, which according to Stephen Grossberg involves multi-level activations in the cortico-thalamic feedback loop, which probably would require roughly at least 100M synaptic activations. And when you recognize a face you normally are seeing it substantially larger than your thumb nail at its furthest extension from your face. If you saw it as large as the length of your entire thumb, rather than just your thumbnail, it would project on to about 10 times as many neurons in your thalamus and V1. So, yes, I was guessing, but I think hundreds of millions of steps. AKA synaptic activations, was a pretty safe guess. Ed Porter -Original Message- From: Mike Tintner [mailto:[EMAIL PROTECTED] Sent: Friday, December 07, 2007 5:08 PM To: agi@v2.listbox.com Subject: Re: [agi] Do we need massive computational capabilities? ED PORTER # When you say It only takes a few steps to retrieve something from memory. I hope you realize that depending how you count steps, it actually probably takes hundreds of millions of steps or more. It is just that millions of them are performed in parallel, such that the longest sequence of any one causal path among such steps is no longer than 100 steps. That is a very, repeat very, different thing that suggesting that only 100 separate actions were taken. Ed, Yes, I understood that (though sure, I'm capable of misunderstanding anything here!) But let's try and make it simple and as concrete as possible - another way of putting Hawkins' point, as I understand, is that at any given level, if the brain is recognising a given feature of the face, it can only compare it with very few comparable features in that half second with its 100 operations - whereas a computer will compare that same feature with vast numbers of others. And actually ditto, for that useful Hofstadter example you quoted, of proceeding from aabc: aabd to jjkl: ??? (although this is a somewhat more complex operation which may take a couple of seconds
Re: [agi] AGI communities and support
AGI related activities everywhere are minimal right now. Even people interested in AI often have no idea what the term AGI means. The meme hasn't spread very far beyond a few technologists and visionaries. I think it's only when someone has some amount of demonstrable success with an AGI system that things will really begin to move. On 07/12/2007, Vladimir Nesov [EMAIL PROTECTED] wrote: On Dec 8, 2007 1:08 AM, Ed Porter [EMAIL PROTECTED] wrote: Vlad, What country are you in? And what is the level of web-comunity, academic, commercial, and governmental support AGI in your country? Ed Porter I live in Moscow. AGI-related activities are nonexistent here; there's a small web community, but I don't follow its discussions, archives for recent years don't show anything interesting. I wonder where your question has a positive answer and how it can look like. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73856876-ce6de8
RE: [agi] Complexity in AGI design
Dennis Gorelik writes: Derek, I quoted this Richard's article in my blog: http://www.dennisgorelik.com/ai/2007/12/reducing-agi-complexity-copy-only-high.html Cool. Now I'll quote your blogged response: So, if low level brain design is incredibly complex - how do we copy it? The answer is: we don't copy low level brain design. Low level design is critical for AGI. Instead we observe high level brain patterns and try to implement them on top of our own, more understandable, low level design. I'm not sure for myself what I think of this complexity argument, so I don't have anything to say about your answer except to wish you luck (if Richard is right, you'll need a lot of it; if many paths lead up the hill then you might not need much at all). I am curious what you mean by high level brain patterns though. Could you give an example? - 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=73825873-cc7440
Re: [agi] Evidence complexity can be controlled by guiding hands
On Dec 7, 2007 10:54 PM, Ed Porter [EMAIL PROTECTED] wrote: Vlad, So, as I understand you, you are basically agreeing with me. Is this correct? Ed Porter I agree that high-level control allows more chaos at lower level, but I don't think that copycat-level stochastic search is necessary or even desirable. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73812575-6ea12f
RE: [agi] Evidence complexity can be controlled by guiding hands
Vlad, So, as I understand you, you are basically agreeing with me. Is this correct? Ed Porter -Original Message- From: Vladimir Nesov [mailto:[EMAIL PROTECTED] Sent: Friday, December 07, 2007 2:24 PM To: agi@v2.listbox.com Subject: Re: [agi] Evidence complexity can be controlled by guiding hands On Dec 7, 2007 7:42 PM, Ed Porter [EMAIL PROTECTED] wrote: Yes, there would be a tremendous number of degrees of freedom, but there would be a tremendous number of sources of guidance and review from the best matching prior experiences of the past successes and failures of the most similar perceptions, thoughts, or behaviors in the most similar contexts. With such guidance, there is reason to believe that even a system large enough to compute human-level world knowledge would stay largely within the realm of common sense and not freak out. It should have enough randomness to fairly often think strange new thoughts, but it should have enough common-sense from its vase experiences to judge roughly as well as a human when to, and when not to, act on such strange new ideas. Ed, I believe that high-level control is instrumental not only to deliberation-level decision-making, but to very formation of system's low-level knowledge and behavior. Hebbian learning needs sufficient time to collect evidence before starting to reliably activate inferential link, lest it risks to disturb system's dynamics and create a bias with positive feedback. It makes fast learning and learning from few examples problematic. Learning can work much faster if it's assisted by recitation loops, which are triggered only for regularities that are deemed reasonable by higher-level processes (these processes can be just of slightly more higher level, I'm not talking about overall control). Also mechanism you describe is why I think it's OK to activate everything that activates: higher-level control (based on regeneration of typical patterns in recitation loops) should remove nonsense and at the same time teach system not to produce it again. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73798921-cd6358
RE: [agi] Solution to Grounding problem
Richard Loosemore writes: This becomes a problem because when we say of another person that they meant something by their use of a particular word (say cat), what we actually mean is that that person had a huge amount of cognitive machinery connected to that word cat (reaching all the way down to the sensory perception mechanisms that allow the person to recognise an instance of a cat, and motor output mechanisms that let them interact with a cat). What Stephen Harnad said in his original paper was Hang on a second: if the AI system does not have all that other machinery inside it when it uses a word like cat, surely it does not really mean the same thing by cat as a person would? [...] Thanks, Richard. That post was a terrific bit of writing. On a related note, I think those that are uneasy with the idea of grounding symbols in experience with a virtual world wonder whether the (current) thin and skewed sensory experiece of cats or any other concept-friendly regularities in such worlds are sufficiently similar to provide enough of the same meaning for communication with humans using the resulting concepts. For that matter, one wonders even when concepts are grounded in the real world whether the resulting concepts and their meanings can be similar enough for communication if the concept formation machinery is not quite similar to our own sometimes even individual human conceptualizations are barely similar enough to allow conversation. Very interesting stuff. - 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=73797095-e37936
Re: [agi] Evidence complexity can be controlled by guiding hands
On Dec 7, 2007 7:42 PM, Ed Porter [EMAIL PROTECTED] wrote: Yes, there would be a tremendous number of degrees of freedom, but there would be a tremendous number of sources of guidance and review from the best matching prior experiences of the past successes and failures of the most similar perceptions, thoughts, or behaviors in the most similar contexts. With such guidance, there is reason to believe that even a system large enough to compute human-level world knowledge would stay largely within the realm of common sense and not freak out. It should have enough randomness to fairly often think strange new thoughts, but it should have enough common-sense from its vase experiences to judge roughly as well as a human when to, and when not to, act on such strange new ideas. Ed, I believe that high-level control is instrumental not only to deliberation-level decision-making, but to very formation of system's low-level knowledge and behavior. Hebbian learning needs sufficient time to collect evidence before starting to reliably activate inferential link, lest it risks to disturb system's dynamics and create a bias with positive feedback. It makes fast learning and learning from few examples problematic. Learning can work much faster if it's assisted by recitation loops, which are triggered only for regularities that are deemed reasonable by higher-level processes (these processes can be just of slightly more higher level, I'm not talking about overall control). Also mechanism you describe is why I think it's OK to activate everything that activates: higher-level control (based on regeneration of typical patterns in recitation loops) should remove nonsense and at the same time teach system not to produce it again. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73782244-94fe20
RE: [agi] Do we need massive computational capabilities?
Bob, I agree. I think we should be able to make PC based AGI's. With only about 50 million atoms they really wouldn't bea ble to have much world knowledge, but they should be able to understand, say the world of a simple video game, such as pong or PacMan. As Richard Loosemore and I have just discussed in our last several emails on the Evidence complexity can be controlled by guiding hands thread, to achieve powerful AGI's we will need very large complex systems and we need to start experimenting with how to control the complexity of such larger systems. So building AGI's on a PC is a good start, which will hopefully start happening ofter OpenCog comes out, but we also need to start building and exploring larger system. It is my very rough guess that human level AGI will need within several orders of magnitude of 10TBytes of RAM or approximately as fast memory, 10T random RAM accesses/sec, and global x sectional bandwidth of 100G 64 Byte messages/sec. So you won't have that on your desktop any time soon. But in twenty years you might. We should be exploring constantly bigger and bigger machines between a PC AGI and human level AGI's to learn more and more about the problem of scaling up large systems. Ed Porter -Original Message- From: Bob Mottram [mailto:[EMAIL PROTECTED] Sent: Friday, December 07, 2007 10:21 AM To: agi@v2.listbox.com Subject: Re: [agi] Do we need massive computational capabilities? If I had 100 of the highest specification PCs on my desktop today (and it would be a big desk!) linked via a high speed network this wouldn't help me all that much. Provided that I had the right knowledge I think I could produce a proof of concept type AGI on a single PC today, even if it ran like a tortoise. It's the knowledge which is mainly lacking I think. Although I do a lot of stuff with computer vision I find myself not being all that restricted by computational limitations. This certainly wasn't the case a few years ago. Generally even the lowest end hardware these days has enough compute power to do some pretty sophisticated stuff, especially if you include the GPU. - 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=73637520-930b42
Re: [agi] Do we need massive computational capabilities?
--- Dennis Gorelik [EMAIL PROTECTED] wrote: Matt, For example, I disagree with Matt's claim that AGI research needs special hardware with massive computational capabilities. I don't claim you need special hardware. But you claim that you need massive computational capabilities [considerably above capabilities of regular modern PC], right? That means special. No, my proposal requires lots of regular PCs with regular network connections. It is a purely software approach. But more hardware is always better. http://www.mattmahoney.net/agi.html -- Matt Mahoney, [EMAIL PROTECTED] - 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=73635143-725e61
Re: [agi] Do we need massive computational capabilities?
Clearly the brain works VASTLY differently and more efficiently than current computers - are you seriously disputing that? It is very clear that in many respects the brain is much less efficient than current digital computers and software. It is more energy-efficient by and large, as Read Montague has argued ... but OTOH sometimes it is wy less algorithmically efficient For instance, in spite of its generally high energy efficiency, my brain wastes a lot more energy calculating 969695775755/ 8884 than my computer does. And e.g. visual cortex, while energy-efficient, is horribly algorithmically inefficient, involving e.g. masses of highly erroneous motion-sensing neurons whose results are averaged together to give reasonably accurate values.. -- 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=73893310-401039
Re[2]: [agi] Do we need massive computational capabilities?
Matt, No, my proposal requires lots of regular PCs with regular network connections. Properly connected set of regular PCs would usually have way more power than regular PC. That makes your hardware request special. My point is - AGI can successfully run on singe regular PC. Special hardware would be required later, when you try to scale out working AGI prototype. It is a purely software approach. But more hardware is always better. Not always. More hardware costs money and requires more maintenance. http://www.mattmahoney.net/agi.html - 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=73892920-985965
[agi] Worst case scenario
Here's the worst case scenario I see for ai: that there has to be hardware complexity to the extent that generally nobody is going to be able to get the initial push. Indeed, there's Moore's law to take account of, but the economics might just prevent us from accumulating enough nodes, enough connections, and so on. So, worst case, maybe some gazillionair will have to purchase/make his own semiconductor manufacturing facility and have it completely devoted to building additional microprocessors to add to a giant cluster, supercomputer, or computation cloud, whatever you want to call it. A first step on the way to such a setup might be purchasing supercomputer time and trying to wire up a few different supers, then trying to see if even a percentage of the computational power predicted yields results remotely ressembling ai. Over time, ai will improve and so the semiconductor facility can recover costs by hosting a very large digital work force, but this is all or nothing and so what arguments might there be to persuade a gazillionair into doing this? - 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=73878310-b41ab6
Re: [agi] Do we need massive computational capabilities?
On Friday 07 December 2007, Mike Tintner wrote: P.S. You also don't answer my question re: how many neurons in total *can* be activated within a half second, or given period, to work on a given problem - given their relative slowness of communication? Is it indeed possible for hundreds of millions of messages about that one subject to be passed among millions of neurons in that short space (dunno-just asking)? Or did you pluck that figure out of the air? I suppose that the number of neurons that are working on a problem at a moment will have to expand exponentially based on the number of synaptic connections per neuron as well as the number of hits/misses per neuron that are receiving the signals, viewed as if an expanding light-cone sphere in the brain (it's of course, a neural activity cone / sphere, not light). I am sure this rate can be made into a model. - 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=73877309-9727c9
Re: [agi] AGI communities and support
The robotics revolution is already happening. Presumably, as some kind of roboticist, you would agree? The robotics revolution has already happened. There has been a quiet revolution in some manufacturing industries with large amounts of human labour being replaced by automation. However, this isn't something which most people ever see or are aware of and it doesn't catch media attention since this is mostly dull, repetitive and unglamourous work. Much of our modern lifestyles with cheap consumer goods is actually supported and enabled by robotic labour. Cheap human labour remains competitive, but there will come a time within the next few decades when no human labour - however inexpensive - will be able to compete economically against automated factories. However, this is only one largely unseen revolution. The next robotics revolution is yet to begin, and this will see another wave of automation moving out of factories and into homes and offices. Don't be fooled by the showy humanoids that you might see strutting around or playing violins. I very much doubt that consumer robotics is going to look like this. It's going to be far more utilitarian. Think Roomba rather than ASIMO. This revolution will begin once there is some cheap and easy way of taking regular PC hardware and making it mobile by adding legs, wheels and arms. At the moment doing this involves a good deal of expertise, and we're waiting for off-the-shelf standardised systems to replace the current perpetual re-inventions of the wheel. Once you have standards and the price is right then the vast pool of IT developers who previously had no involvement with robots will be able to apply their expertise to robotics problems. If you're interested in what needs to be done before the second wave can begin see Matt Trossen's recent talk. http://www.trossenrobotics.com/tutorials/trossenroboticssystem.aspx - 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=73876643-6ac08c
Re: [agi] AGI communities and support
On Dec 8, 2007 1:08 AM, Ed Porter [EMAIL PROTECTED] wrote: Vlad, What country are you in? And what is the level of web-comunity, academic, commercial, and governmental support AGI in your country? Ed Porter I live in Moscow. AGI-related activities are nonexistent here; there's a small web community, but I don't follow its discussions, archives for recent years don't show anything interesting. I wonder where your question has a positive answer and how it can look like. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73849598-ab4225
Re: [agi] Solution to Grounding problem
Derek Zahn wrote: Richard Loosemore writes: This becomes a problem because when we say of another person that they meant something by their use of a particular word (say cat), what we actually mean is that that person had a huge amount of cognitive machinery connected to that word cat (reaching all the way down to the sensory perception mechanisms that allow the person to recognise an instance of a cat, and motor output mechanisms that let them interact with a cat). What Stephen Harnad said in his original paper was Hang on a second: if the AI system does not have all that other machinery inside it when it uses a word like cat, surely it does not really mean the same thing by cat as a person would? [...] Thanks, Richard. That post was a terrific bit of writing. On a related note, I think those that are uneasy with the idea of grounding symbols in experience with a virtual world wonder whether the (current) thin and skewed sensory experiece of cats or any other concept-friendly regularities in such worlds are sufficiently similar to provide enough of the same meaning for communication with humans using the resulting concepts. For that matter, one wonders even when concepts are grounded in the real world whether the resulting concepts and their meanings can be similar enough for communication if the concept formation machinery is not quite similar to our own sometimes even individual human conceptualizations are barely similar enough to allow conversation. That is a very good point, and one to which I don't have a ready answer. This question will attract a good deal of attention when we get nearer to the point of being able to test real candidate AGI systems. It is another reason to stay close to the human design, I believe. 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=73905368-2fdc72
Re[2]: [agi] Do we need massive computational capabilities?
Matt, Matt,:AGI research needs special hardware with massive computational capabilities. Could you give an example or two of the kind of problems that your AGI system(s) will need such massive capabilities to solve? It's so good - in fact, I would argue, essential - to ground these discussions. For example, I ask the computer who is this? and attach a video clip from my security camera. Why do you need image recognition in your AGI prototype? You can feed it with text. Then AGI would simply parse text [and optionally - Google it]. No need for massive computational capabilities. - 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=73892756-356b26
Re[4]: [agi] Solution to Grounding problem
Mike, 1. Bush walks like a cowboy, doesn't he? The only way a human - or a machine - can make sense of sentence 1 is by referring to a mental image/movie of Bush walking. That's not the only way to make sense of the saying. There are many other ways: chat with other people, or look on Google: http://www.google.com/search?q=Bush+walks+cowboy http://images.google.com/images?q=grundchen Merely referring to more words won't cut it. It would. Meaning - is connection between concepts. If proper words are referred, then meaning is there. Oh, just to make your day, if you don't have a body, you can't understand the images either How is that don't have a body remark relevant? Computers have body and senses (such as keyboard and Internet connection). Is all that clear? No. You didn't describe what grounding problem is about. - 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=73839347-3bb442
RE: [agi] Do we need massive computational capabilities?
Mike, MIKE TINTNER # Hawkins' point as to how the brain can decide in a hundred steps what takes a computer a million or billion steps (usually without much success) is: The answer is the brain doesn't 'compute' the answers ; it retrieves the answers from memory. In essence, the answers were stored inmemory a long time ago. It only takes a few steps to retrieve something from memory. Slow neurons are not only fast enough to do this, but they constitute the memory themselves. The entire cortex is a memory system. It isn't a computer at all.[ON INtelligence - Chapter on Memory] ED PORTER # When you say It only takes a few steps to retrieve something from memory. I hope you realize that depending how you count steps, it actually probably takes hundreds of millions of steps or more. It is just that millions of them are performed in parallel, such that the longest sequence of any one causal path among such steps is no longer than 100 steps. That is a very, repeat very, different thing that suggesting that only 100 separate actions were taken. You many already know and mean this, but from a quick read of your argument it was not clear you did. So I don't know which side of the Do we need massive computational capabilities? you are on, but we do need massive computational capabilities. That 100 step task you referred, which often involves recognizing a person at a different scale, angle, body position, facial expression, and lighting, than we have seen them before, would probably require many hundreds of millions of neuron to neuron messages in the brain, and many hundreds of millions of computations in a computer. I hope you realize that Hawkin's theory of Hierarchical memory means that images are not stored as anything approaching photographs or drawings. They are stored as distributed hierarchical representations, in which a match would often require parallel computing involving matching and selection at multiple different representational levels. The answer is not retrieved from memory by any simple process, like vectoring into a look-up table, and hopping to an address where the matching image is simply retrieved like a jpg file. The quote retrieval is a relatively massively parallel operation. You may already understand all of this, but it was not obvious from your below post. Some parts of your post seemed to reflect the correct understanding, others didn't, at least from my quick read. Ed Porter -Original Message- From: Mike Tintner [mailto:[EMAIL PROTECTED] Sent: Friday, December 07, 2007 3:26 PM To: agi@v2.listbox.com Subject: Re: [agi] Do we need massive computational capabilities? Matt, First of all, we are, I take it, discussing how the brain or a computer can recognize an individual face from a video - obviously the brain cannot match a face to a selection of a billion other faces. Hawkins' answer to your point that the brain runs masses of neurons in parallel in order to accomplish facial recognition is: if I have many millions of neurons working together, isn't that like a parallel computer? Not really. Brains operate in parallel parallel computers operate in parallel, but that's the only thing they have in common.. His basic point, as I understand, is that no matter how many levels of brain are working on this problem of facial recognition, they are each still only going to be able to perform about ONE HUNDRED steps each in that half second. Let's assume there are levels for recognising the invariant identity of this face, different features, colours, shape, motion etc - each of those levels is still going to have to reach its conclusions EXTREMELY rapidly in a very few steps. And all this, as I said, I would have thought all you guys should be able to calculate within a very rough ballpark figure. Neurons only transmit signals at relatively slow speeds, right? Roughly five million times slower than computers. There must be a definite limit to how many neurons can be activated and how many operations they can perform to deal with a facial recognition problem, from the time the light hits the retina to a half second later? This is the sort of thing you all love to calculate and is really important - but where are you when one really needs you? Hawkins' point as to how the brain can decide in a hundred steps what takes a computer a million or billion steps (usually without much success) is: The answer is the brain doesn't 'compute' the answers ; it retrieves the answers from memory. In essence, the answers were stored inmemory a long time ago. It only takes a few steps to retrieve something from memory. Slow neurons are not only fast enough to do this, but they constitute the memory themselves. The entire cortex is a memory system. It isn't a computer at all.[ON INtelligence - Chapter on Memory] I was v. crudely arguing something like this in a discussion with Richard about massive parallel computation. If
Re: [agi] Do we need massive computational capabilities?
Matt, First of all, we are, I take it, discussing how the brain or a computer can recognize an individual face from a video - obviously the brain cannot match a face to a selection of a billion other faces. Hawkins' answer to your point that the brain runs masses of neurons in parallel in order to accomplish facial recognition is: if I have many millions of neurons working together, isn't that like a parallel computer? Not really. Brains operate in parallel parallel computers operate in parallel, but that's the only thing they have in common.. His basic point, as I understand, is that no matter how many levels of brain are working on this problem of facial recognition, they are each still only going to be able to perform about ONE HUNDRED steps each in that half second. Let's assume there are levels for recognising the invariant identity of this face, different features, colours, shape, motion etc - each of those levels is still going to have to reach its conclusions EXTREMELY rapidly in a very few steps. And all this, as I said, I would have thought all you guys should be able to calculate within a very rough ballpark figure. Neurons only transmit signals at relatively slow speeds, right? Roughly five million times slower than computers. There must be a definite limit to how many neurons can be activated and how many operations they can perform to deal with a facial recognition problem, from the time the light hits the retina to a half second later? This is the sort of thing you all love to calculate and is really important - but where are you when one really needs you? Hawkins' point as to how the brain can decide in a hundred steps what takes a computer a million or billion steps (usually without much success) is: The answer is the brain doesn't 'compute' the answers ; it retrieves the answers from memory. In essence, the answers were stored inmemory a long time ago. It only takes a few steps to retrieve something from memory. Slow neurons are not only fast enough to do this, but they constitute the memory themselves. The entire cortex is a memory system. It isn't a computer at all.[ON INtelligence - Chapter on Memory] I was v. crudely arguing something like this in a discussion with Richard about massive parallel computation. If Hawkins is right, and I think he's at least warm, you guys have surely got it all wrong. (although you might still argue like Ben that you can it do your way not the brain's - but hell, the difference in efficiency is so vast it surely ought to break your engineering heart). Matt/ MT: Thanks. And I repeat my question elsewhere : you don't think that the human brain which does this in say half a second, (right?), is using massive computation to recognize that face? So if I give you a video clip then you can match the person in the video to the correct photo out of 10^9 choices on the Internet in 0.5 seconds, and this will all run on your PC? Let me know when your program is finished so I can try it out. You guys with all your mathematical calculations re the brain's total neurons and speed of processing surely should be able to put ball-park figures on the maximum amount of processing that the brain can do here. Hawkins argues: neurons are slow, so in that half a second, the information entering your brain can only traverse a chain ONE HUNDRED neurons long. ..the brain 'computes' solutions to problems like this in one hundred steps or fewer, regardless of how many total neurons might be involved. From the moment light enters your eye to the time you [recognize the image], a chain no longer than one hundred neurons could be involved. A digital computer attempting to solve the same problem would take BILLIONS of steps. One hundred computer instructions are barely enough to move a single character on the computer's display, let alone do something interesting. Which is why the human brain is so bad at arithmetic and other tasks that require long chains of sequential steps. But somehow it can match a face to a name in 0.5 seconds. Neurons run in PARALLEL. Your PC does not. Your brain performs 10^11 weighted sums of 10^15 values in 0.1 seconds. Your PC will not. IOW, if that's true, the massive computational approach is surely RIDICULOUS - a grotesque travesty of engineering principles of economy, no? Like using an entire superindustry of people to make a single nut? And, of course, it still doesn't work. Because you just don't understand how perception works in the first place. Oh right... so let's make our computational capabilities even more massive, right? Really, really massive. No, no, even bigger than that? Matt,:AGI research needs special hardware with massive computational capabilities. Could you give an example or two of the kind of problems that your AGI system(s) will need such massive capabilities to solve? It's so good - in fact, I would argue, essential -
Re: [agi] Evidence complexity can be controlled by guiding hands
On Dec 7, 2007 7:05 PM, Richard Loosemore [EMAIL PROTECTED] wrote: You are asking good questions about the mechanisms, which I am trying to explore emprically. No good answers to this yet, although I have many candidate solutions, some of which (I think) look like your above model. I certainly agree with the sentiment that not *all* of the process can be as fluid as the higher level parts (if that is what you are meaning). Or maybe they even shouldn't and can't be too fluid: it would be a challenge to precisely implement procedures otherwise (like playing piano). -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73783598-bfca32
Re[2]: [agi] How to represent things problem
Richard, the instance nodes are such an important mechanism that everything depends on the details of how they are handled. Correct. So, to consider one or two of the details that you mention. You would like there to be only a one-way connection between the generic node (do you call this the pattern node?) 1) All nodes are equal. 2) Nodes can point to each other. Yes, connection should be one way. (E.g.: You know George Bush, but he doesn't know you :-)) Two-way connection can be easily implemented by two separate connections. For instance, we are able to see a field of patterns, of different colors, and then when someone says the phrase the green patterns we find that the set of green patterns jumps out at us from the scene. It is as if we did indeed have links from the generic concept [green pattern] to all the instances. Yes, that's good way to store links: All relevant nodes are connected. Another question: what do we do in a situation where we see a field of grass, and think about the concept [grass blade]? Field or grass concept and grass blade concept are obviously directly connected. This link was formed, because we saw Field of grass and Grass blade together many times. Are there individual instances for each grass blade? If you remember individual instances -- then yes. Are all of these linked to the generic concept of [grass blade]? Some grass blades may be directly connected to field of grass. Other may be connected only through other grass blade instances. It would depend on if it's useful for brain to keep these direct associations. What is different is that I see many, many possible ways to get these new-node creation mechanisms to work (and ditto for other mechanisms like the instance nodes, etc.) and I feel it is extremely problematic to focus on just one mechanism and say that THIS is the one I will implement because I think it feels like a good idea. The reason I think this is a problem is that these mechanisms have system-wide consequences (i.e. they give rise to global behaviors) that are not necessarily obvious from the definition of the mechanism, so we need to build a simulation to find out what those mechanisms *really* do when they are put together and allowed to interact. I agree -- testing is important. In fact, it's extremely important. Not only we need to test several models [of creating updating nodes and links), but within single model we should try several settings values (such as if node1 and node2 were activated together -- how much should we increase the strength of the link between them). That's why it's important to carefully design tests. Such tests should work reasonably fast and be able to indicate how good did system work. What is good and what is not good -- has to be carefully defined. Not trivial, but quite doable task. I can show you a paper of mine in which I describe my framework in a little more detail. Isn't this pager public yet? - 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=73487197-bbb1fa
Re: [agi] Do we need massive computational capabilities?
--- Mike Tintner [EMAIL PROTECTED] wrote: Matt,:AGI research needs special hardware with massive computational capabilities. Could you give an example or two of the kind of problems that your AGI system(s) will need such massive capabilities to solve? It's so good - in fact, I would argue, essential - to ground these discussions. For example, I ask the computer who is this? and attach a video clip from my security camera. -- Matt Mahoney, [EMAIL PROTECTED] - 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=73639920-0e69de
RE: [agi] Evidence complexity can be controlled by guiding hands
Vlad, Agreed. Copycat is a lot more wild and crazy at the low level than my system would be. But my system might operate more like it at a higher more deliberative level. For example, this might be the case if I were trying to attack a difficult planning problem, such as how to write an answer to a complex question in a most convincing manner. (Of course there I would have the words on my computer screen to help me keep track of a significant part of the problem space.) But the fact that Copycat's craziness can be relatively effectively harnessed to do what it is supposed to is an encouraging sign that the potential pitfalls of complexity can be significantly avoided I say significantly because Richard has a point, once a system gets really complex, it get increasingly more difficult to feel you truly understand it, and thus that you can truly trust it. Of course that goes for people too. Every so often one of them goes postal. Ed Porter -Original Message- From: Vladimir Nesov [mailto:[EMAIL PROTECTED] Sent: Friday, December 07, 2007 3:07 PM To: agi@v2.listbox.com Subject: Re: [agi] Evidence complexity can be controlled by guiding hands On Dec 7, 2007 10:54 PM, Ed Porter [EMAIL PROTECTED] wrote: Vlad, So, as I understand you, you are basically agreeing with me. Is this correct? Ed Porter I agree that high-level control allows more chaos at lower level, but I don't think that copycat-level stochastic search is necessary or even desirable. -- Vladimir Nesovmailto:[EMAIL PROTECTED] - 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=73828093-9a54e5
Re: [agi] Do we need massive computational capabilities?
RE: [agi] Do we need massive computational capabilities?ED PORTER # When you say It only takes a few steps to retrieve something from memory. I hope you realize that depending how you count steps, it actually probably takes hundreds of millions of steps or more. It is just that millions of them are performed in parallel, such that the longest sequence of any one causal path among such steps is no longer than 100 steps. That is a very, repeat very, different thing that suggesting that only 100 separate actions were taken. Ed, Yes, I understood that (though sure, I'm capable of misunderstanding anything here!) But let's try and make it simple and as concrete as possible - another way of putting Hawkins' point, as I understand, is that at any given level, if the brain is recognising a given feature of the face, it can only compare it with very few comparable features in that half second with its 100 operations - whereas a computer will compare that same feature with vast numbers of others. And actually ditto, for that useful Hofstadter example you quoted, of proceeding from aabc: aabd to jjkl: ??? (although this is a somewhat more complex operation which may take a couple of seconds for the brain), again a typical intelligent brain will almost certainly consider v. few options, compared with the vast numbers of options considered by that computer. Ditto, for godsake, a human chessplayer like Kasparov's brain considers an infinitesimal percentage of the moves considered by Big Blue in any given period - and yet can still win (occasionally) because of course it's working on radically different principles. Hawkins' basic point that the brain isn't a computer at all - which I think can be read less controversially as is a machine that works on very fundamentally different principles to those of currently programmed computers - especially when perceiving objects - holds. You're not dealing with that basic point, and I find it incredibly difficult to get anyone here squarely to face it. People retreat into numbers and millions. Clearly the brain works VASTLY differently and more efficiently than current computers - are you seriously disputing that? P.S. You also don't answer my question re: how many neurons in total *can* be activated within a half second, or given period, to work on a given problem - given their relative slowness of communication? Is it indeed possible for hundreds of millions of messages about that one subject to be passed among millions of neurons in that short space (dunno-just asking)? Or did you pluck that figure out of the air? P.P. S. A recent book by Read Montague on neuroeconomics makes much the same point from a v. different angle - highlighting that computers have a vastly wasteful search heritage which he argues has its roots in Turing and Bletchley Park's attempts to decode Engima. - 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=73842990-515452
Re: [agi] AGI communities and support
Bob : AGI related activities everywhere are minimal right now. Even people interested in AI often have no idea what the term AGI means. The meme hasn't spread very far beyond a few technologists and visionaries. I think it's only when someone has some amount of demonstrable success with an AGI system that things will really begin to move. Bob, Just to confirm - after Richard mentioned hearing AGI on the radio. I was following up a brain machine interface story, on the New Scientist website. At the bottom it offered me: Robots - Learn more about the robotics revolution in our continually updated special report - which led to a mass of robotics stories. So I searched the site for Artificial General Intelligence and AGI. Nothing. The robotics revolution is already happening. Presumably, as some kind of roboticist, you would agree? - 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=73864470-0e335b
Re: Human Irrationality [WAS Re: [agi] None of you seem to be able ...]
Mike Tintner wrote: Richard:For my own system (and for Hofstadter too), the natural extension of the system to a full AGI design would involve a system [that] can change its approach and rules of reasoning at literally any step of problem-solving it will be capable of producing all the human irrationalities that I listed previously - like not even defining or answering the problem. It will by the same token have the capacity to be truly creative, because it will ipso facto be capable of lateral thinking at any step of problem-solving. This is very VERY much part of the design. There is not any problem with doing all of this. Does this clarify the question? I think really I would reflect the question back at you and ask why you would think that this is a difficult thing to do? Richard, Fine. Sounds interesting. But you don't actually clarify or explain anything. Why don't you explain how you or anyone else can fundamentally change your approach/rules at any point of solving a problem? Why don't you, just in plain English, - in philosophical as opposed to programming form - set out the key rules or principles that allow you or anyone else to do this? I have never seen such key rules or principles anywhere, nor indeed even adumbrated anywhere. (Fancy word, but it just came to mind). And since they are surely a central problem for AGI - and no one has solved AGI - how on earth could I not think this a difficult matter? I have some v. rough ideas about this, which I can gladly set out. But I'd like to hear yours - you should be able to do it briefly. But please, no handwaving. I will try to think about your question when I can but meanwhile think about this: if we go back to the analogy of painting and whether or not it can be used to depict things that are abstract or non-representational, how would you respond to someone who wanted exact details of how come painting could allow that to be possible.? If someone asked that, I couldn't think of anything to say except ... why *wouldn't* it be possible? It would strike me as just not a question that made any sense, to ask for the exact reasons why it is possible to paint things that are not representational. I simply cannot understand why anyone would think it not possible to do that. It is possible: it is not easy to do it right, but that's not the point. Computers can be used to program systems of any sort (including deeply irrational things like Microsoft Office), so why would anyone think that AGI systems must exhibit only a certain sort of design? This isn't handwaving, it is just genuine bafflement. 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=73903282-a471b6
