Re: [agi] interesting Google Tech Talk about Neural Nets
On 03/03/2008, [EMAIL PROTECTED] [EMAIL PROTECTED] wrote: Dont you see the way to go on Neural nets is hybrid with genetic algorithms in mass amounts? I experimented with this combination in the early 1990s, and the results were not very impressive. Such systems still suffered from extremely slow learning and poor scalability. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
On Mon, Mar 3, 2008 at 6:33 AM, [EMAIL PROTECTED] wrote: Thanks for that. Dont you see the way to go on Neural nets is hybrid with genetic algorithms in mass amounts? No, I dont agree with your buzzword-laden statement :) I experimented EA + NN's and its still intractable when scaled up to nontrivial samples. Luckily, there exist more efficient learning methods for NN's then search. For multilayer perceptrons there's standard backpropagation (gradient descent), conjugate gradient descent, newton's method, etc. For RBM's, there's contrastive divergence (CD) or wake-sleep using Gibbs sampling, etc. The great thing about RBM's is that while still slow at learning (can take a few days to converge a complex model), it's a very very simple architecture (just a few matrices) plus very simple learning methods (just a few matrix multiplications) that SEEMS to be exceptionally good at building a * generative* model from (labeled or unlabeled) complex data. With RBM's you can do all kinds of interesting stuff like: - confabulate novel samples from model; - compression (although inherently lossy) - visualisation in 2D (compress to 2 neurons) There's a nice flash demonstration about digit generation/classification http://www.cs.toronto.edu/~hinton/adi/index.htmhttp://www.cs.toronto.edu/%7Ehinton/adi/index.htm Did anyone on this list do experiments with these kind of generative models? I'd can't find much research into this subject outside from the Univ. of Tortonto's CS group, so the information reaching me might be positivily biased. I don't have any affiliation with this group if anyone might ask. Durk --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
On Mon, Mar 3, 2008 at 4:29 PM, Kingma, D.P. [EMAIL PROTECTED] wrote: There's a nice flash demonstration about digit generation/classification http://www.cs.toronto.edu/~hinton/adi/index.htm Did anyone on this list do experiments with these kind of generative models? I'd can't find much research into this subject outside from the Univ. of Tortonto's CS group, so the information reaching me might be positivily biased. I don't have any affiliation with this group if anyone might ask. At this point I see this kind of system as a sufficient substrate for AGI (modulo layers being abandoned, and a 'Perti dish' of feature detectors listening to recent past states of their neighbors used instead). I only converged on this view recently, so I'm just starting to experiment with it. -- Vladimir Nesov [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] Goal Driven Systems and AI Dangers [WAS Re: Singularity Outcomes...]
Kaj Sotala wrote: On 2/16/08, Richard Loosemore [EMAIL PROTECTED] wrote: Kaj Sotala wrote: Well, the basic gist was this: you say that AGIs can't be constructed with built-in goals, because a newborn AGI doesn't yet have built up the concepts needed to represent the goal. Yet humans seem tend to have built-in (using the term a bit loosely, as all goals do not manifest in everyone) goals, despite the fact that newborn humans don't yet have built up the concepts needed to represent those goals. Oh, complete agreement here. I am only saying that the idea of a built-in goal cannot be made to work in an AGI *if* one decides to build that AGI using a goal-stack motivation system, because the latter requires that any goals be expressed in terms of the system's knowledge. If we step away from that simplistic type of GS system, and instead use some other type of motivation system, then I believe it is possible for the system to be motivated in a coherent way, even before it has the explicit concepts to talk about its motivations (it can pursue the goal seek Momma's attention long before it can explicitly represent the concept of [attention], for example). Alright. But previously, you said that Omohundro's paper, which to me seemed to be a general analysis of the behavior of *any* minds with (more or less) explict goals, looked like it was based on a 'goal-stack' motivation system. (I believe this has also been the basis of your critique for e.g. some SIAI articles about friendliness.) If built-in goals *can* be constructed into motivational system AGIs, then why do you seem to assume that AGIs with built-in goals are goal-stack ones? I seem to have caused lots of confusion earlier on in the discussion, so let me backtrack and try to summarize the structure of my argument. 1) Conventional AI does not have a concept of a Motivational-Emotional System (MES), the way that I use that term, so when I criticised Omuhundro's paper for referring only to a Goal Stack control system, I was really saying no more than that he was assuming that the AI was driven by the system that all conventional AIs are supposed to have. These two ways of controlling an AI are two radically different designs. 2) Not only are MES and GS different classes of drive mechanism, they also make very different assumptions about the general architecture of the AI. When I try to explain how an MES works, I often get tangled up in the problem of explaining the general architecture that lies behind it (which does, I admit, cause much confusion). I sometimes use the terms molecular or sub-symbolic to describe that architecture. 2(a) I should say something about the architecture difference. In a sub-symbolic architecture you would find that the significant thought events are the result of clouds of sub-symbolic elements interacting with one another across a broad front. This is to be contrasted with the way that symbols interact in a regular symbolic AI, where symbols are single entities that get plugged into well-defined mechanisms like deduction operators. In a sub-symbolic system, operations are usually the result of several objects *constraining* one another in a relatively weak manner, not the result of a very small number of objects slotting into a precisely defined, rigid mechanism. There is a flexibility inherent in the sub-symbolic architecture that is completely lacking in the conventional symbolic system. 3) It is important to understand that in an AI that uses the MES drive system, there is *also* a goal stack, quite similar to what is found in a GS-driven AI, but this goal stack is entirely subservient to the MES, and it plays a role only in the day to day (and moment to moment) thinking of the system. 4) I plead guilty to saying things like ... Goal-Stack motivation system... when what I should do is use the word motivation only in the context of an MES system. A better wording would have been ... Goal-Stack *drive* system Or perhaps ... Goal-Stack *control* system 5) The main thrust of my attack on GS-driven AIs is that goal stacks were invented in the context of planning problems, and were never intended to be used as the global control system for an AI that is capable of long-range development. So, you will find me saying things like A GS drive system is appropriate for handling goals like 'Put the red pyramid on top of the green block', but it makes no sense in the context of goals like 'Be friendly to humans'. Most AI people assume that a GS control system *must* be the way to go, but I would argue that they are in denial about the uselessness of a GS. Also, most conventional AI people assume that a GS is valid simply because they see no alternative ... and this is because the architecture used by most conventional AI does not easily admit of any other type of drive system. In a sense, they have to support the GS idea because
Re: [agi] interesting Google Tech Talk about Neural Nets
Care to state the exact problem you were having? My thought is scalability is to do entirely with speed availability - Original Message - From: Bob Mottram [EMAIL PROTECTED] To: agi@v2.listbox.com Subject: Re: [agi] interesting Google Tech Talk about Neural Nets Date: Mon, 3 Mar 2008 09:48:08 + On 03/03/2008, [EMAIL PROTECTED] [EMAIL PROTECTED] wrote: Dont you see the way to go on Neural nets is hybrid with genetic algorithms in mass amounts? I experimented with this combination in the early 1990s, and the results were not very impressive. Such systems still suffered from extremely slow learning and poor scalability. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Want an e-mail address like mine? Get a free e-mail account today at www.mail.com! --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
that's a great idea Vlad, there are other forms of statistical sampling available. the closer we get to running accelerated evolution to human intelligence the better I beleive. - Original Message - From: Vladimir Nesov [EMAIL PROTECTED] To: agi@v2.listbox.com Subject: Re: [agi] interesting Google Tech Talk about Neural Nets Date: Mon, 3 Mar 2008 17:13:04 +0300 On Mon, Mar 3, 2008 at 4:29 PM, Kingma, D.P. [EMAIL PROTECTED] wrote: There's a nice flash demonstration about digit generation/classification http://www.cs.toronto.edu/~hinton/adi/index.htm Did anyone on this list do experiments with these kind of generative models? I'd can't find much research into this subject outside from the Univ. of Tortonto's CS group, so the information reaching me might be positivily biased. I don't have any affiliation with this group if anyone might ask. At this point I see this kind of system as a sufficient substrate for AGI (modulo layers being abandoned, and a 'Perti dish' of feature detectors listening to recent past states of their neighbors used instead). I only converged on this view recently, so I'm just starting to experiment with it. -- Vladimir Nesov [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Want an e-mail address like mine? Get a free e-mail account today at www.mail.com! --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
[EMAIL PROTECTED] wrote: Care to state the exact problem you were having? My thought is scalability is to do entirely with speed availability The problems with bolting together NN and GA are so numerous it is hard to know where to begin. For one thing, you cannot represent structured information with NNs unless you go to some trouble to add extra architecture. Most NNs can only cope with single concepts learned in isolation, so if you show a visual field containing 5,000 copies of the letter 'A', all that happens is that the 'A' neuron fires. If you do find some way to get around this problem, your solution will end up being the tail that wags the dog: the NN itself will fade into relative insignificance compared to your solution. Richard Loosemore - Original Message - From: Bob Mottram [EMAIL PROTECTED] To: agi@v2.listbox.com Subject: Re: [agi] interesting Google Tech Talk about Neural Nets Date: Mon, 3 Mar 2008 09:48:08 + On 03/03/2008, [EMAIL PROTECTED] [EMAIL PROTECTED] wrote: Dont you see the way to go on Neural nets is hybrid with genetic algorithms in mass amounts? I experimented with this combination in the early 1990s, and the results were not very impressive. Such systems still suffered from extremely slow learning and poor scalability. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
On Mon, Mar 3, 2008 at 6:39 PM, Richard Loosemore [EMAIL PROTECTED] wrote: The problems with bolting together NN and GA are so numerous it is hard to know where to begin. For one thing, you cannot represent structured information with NNs unless you go to some trouble to add extra architecture. Most NNs can only cope with single concepts learned in isolation, so if you show a visual field containing 5,000 copies of the letter 'A', all that happens is that the 'A' neuron fires. If you do find some way to get around this problem, your solution will end up being the tail that wags the dog: the NN itself will fade into relative insignificance compared to your solution. Well, you could achieve that (5000 registration of the letter 'A' with their corresponding position in the image) by using a sliding window over multiple rescaled (and maybe other transformations) transformations of the input image. This way, you get image patches for each window and scale (and maybe other transformations), and each patch can be a given a corresponding position in multidimensional space (e.g., an image patch with X and Y position and scale S has is a point in 3-dimensional space). For each of the produced points (patches) in the space, run the neural net to produce a lower-dimensional code and corresponding energy (= reconstruction quality). Now filter this space by let the points have local battles for salience using some heuristic (e.g. lower energy means higher salience) and filter out the low-salient points. This produces a filtered space with fewer points then the previous one, and each point containing a lower-dimensional code. In the example of the letter 'A', the above method would recognize all 5000 versions while remembering their individual input position. This presumes the neural net is properly trained on the letter 'A' and can properly reconstuct them (using Hinton's method). This should produce 5000 registrations of the letter 'A', while filtering out unimportant information. But you could take it a step further. For each image input, the above method creates a filtered, 3-dimensional space with points containing low-dimensional codes. This space can then again be harvested by taking patches with each patch containing *n* points, each point containing an *m *dimensional code, so each patch being (*m***n*).* *A neural net can be trained on lowering the dimension of these patches from (*m***n*) to something lower-dimensional. This process is quite similar to the one in the previous paragraph. What could *possibly *go wrong? :) Regards, Durk Kingma Richard Loosemore - Original Message - From: Bob Mottram [EMAIL PROTECTED] To: agi@v2.listbox.com Subject: Re: [agi] interesting Google Tech Talk about Neural Nets Date: Mon, 3 Mar 2008 09:48:08 + On 03/03/2008, [EMAIL PROTECTED] [EMAIL PROTECTED] wrote: Dont you see the way to go on Neural nets is hybrid with genetic algorithms in mass amounts? I experimented with this combination in the early 1990s, and the results were not very impressive. Such systems still suffered from extremely slow learning and poor scalability. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
I'm increasingly convinced that the human brain is not a statistical learner, but a logical learner. There are many examples of humans learning concepts/rules from one or two examples, rather than thousands of examples. So I think that at a high level, AGI should be logic-based. But it would be interesting to integrate NN-based techniques to logic-based AI, especially in vision. (NN is also very weak at language processing.) YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
Kingma, D.P. wrote: On Mon, Mar 3, 2008 at 6:39 PM, Richard Loosemore [EMAIL PROTECTED] mailto:[EMAIL PROTECTED] wrote: The problems with bolting together NN and GA are so numerous it is hard to know where to begin. For one thing, you cannot represent structured information with NNs unless you go to some trouble to add extra architecture. Most NNs can only cope with single concepts learned in isolation, so if you show a visual field containing 5,000 copies of the letter 'A', all that happens is that the 'A' neuron fires. If you do find some way to get around this problem, your solution will end up being the tail that wags the dog: the NN itself will fade into relative insignificance compared to your solution. Well, you could achieve that (5000 registration of the letter 'A' with their corresponding position in the image) by using a sliding window over multiple rescaled (and maybe other transformations) transformations of the input image. This way, you get image patches for each window and scale (and maybe other transformations), and each patch can be a given a corresponding position in multidimensional space (e.g., an image patch with X and Y position and scale S has is a point in 3-dimensional space). For each of the produced points (patches) in the space, run the neural net to produce a lower-dimensional code and corresponding energy (= reconstruction quality). Now filter this space by let the points have local battles for salience using some heuristic (e.g. lower energy means higher salience) and filter out the low-salient points. This produces a filtered space with fewer points then the previous one, and each point containing a lower-dimensional code. In the example of the letter 'A', the above method would recognize all 5000 versions while remembering their individual input position. This presumes the neural net is properly trained on the letter 'A' and can properly reconstuct them (using Hinton's method). This should produce 5000 registrations of the letter 'A', while filtering out unimportant information. But you could take it a step further. For each image input, the above method creates a filtered, 3-dimensional space with points containing low-dimensional codes. This space can then again be harvested by taking patches with each patch containing /n/ points, each point containing an /m /dimensional code, so each patch being (/m/*/n/)./ /A neural net can be trained on lowering the dimension of these patches from (/m/*/n/) to something lower-dimensional. This process is quite similar to the one in the previous paragraph. What could /possibly /go wrong? :) Regards, Durk Kingma Excellent! Sounds like a perfect solution ;-). Oh, wait! What about. if the scene is structured in such a way that the 5,000 copies of the letter 'A' were actually scattered around in such a way that most (but not all) of them were arranged to form a huge letter 'A'? Would it then count 5,001 copies? Oh, and one more thing I forgot to mention that is in the same scene (how could I forget this one?): there are also a couple of women standing side by side, leaning against each other with their shoulders touching and keeping their bodies stiff and straight, forming the two sides of a letter 'A', and holding a model of a horizontally reclining woman between them at waist height, to form the crossbar of a letter 'A'. Could we get the NN to recognize, in the context of the overall scene, that here were actually 5,002 copies of the letter 'A'..? And if the scene had one single, rather small letter B over in the corner, would the NN find this funny? You have 30 minutes to devise an algorithm, Durk... :-). Richard Loosemore --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
i stumbled upon this project recently. it adresses the connectivity in a neural network. pretty interresting stuff. could be its a known thing but i just wanted to share this. http://oege.ie.hva.nl/~bergd/ im sorta new to this agi development but as far as i understand, couldn't this speed up the performance/effectivity to agi software?(keeping in mind the current topic) -- youri lima, Developer --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
RE: [agi] Thought experiment on informationally limited systems
How intelligent would any human be if it couldn't be taught by other humans? Could a human ever learn to speak by itself? The few times this has happened in real life, the person was permanently disabled and not capable of becoming a normal human being. If humans can't become human without the help of other humans, why should this is a criteria for AGI? David Clark PS I am not suggesting that explicitly programming 100% of an AGI is either doable or desirable but some degree of detailed teaching must be a requirement for all on this list who dream of creating an AGI, no? -Original Message- From: Mike Tintner [mailto:[EMAIL PROTECTED] Sent: March-02-08 5:36 AM To: agi@v2.listbox.com Subject: Re: [agi] Thought experiment on informationally limited systems Jeez, Will, the point of Artificial General Intelligence is that it can start adapting to an unfamiliar situation and domain BY ITSELF. And your FIRST and only response to the problem you set was to say: I'll get someone to tell it what to do. IOW you simply avoided the problem and thought only of cheating. What a solution, or merest idea for a solution, must do is tell me how that intelligence will start adapting by itself - will generalize from its existing skills to cross over domains. Then, as my answer indicated, it may well have to seek some instructions and advice - especially and almost certainly if it wants to acquire a whole new major skill, as we do, by taking courses etc. But a general intelligence should be able to adapt to some unfamiliar situations entirely by itself - like perhaps your submersible situation. No guarantee that it will succeed in any given situation, (as there isn't with us), but you should be able to demonstrate its power to adapt sometimes. In a sense, you should be appalled with yourself that you didn't try to tackle the problem - to produce a cross-over idea. But since literally no one else in the field of AGI has the slightest cross-over idea - i.e. is actually tackling the problem of AGI, - and the whole culture is one of avoiding the problem, it's to be expected. (You disagree - show me one, just one, cross-over idea anywhere. Everyone will give you a v. detailed,impressive timetable for how long it'll take them to produce such an idea, they just will never produce one. Frankly, they're too scared). Mike Tintner [EMAIL PROTECTED] wrote: You must first define its existing skills, then define the new challenge with some degree of precision - then explain the principles by which it will extend its skills. It's those principles of extension/generalization that are the be-all and end-all, (and NOT btw, as you suggest, any helpful info that the robot will receive - that,sir, is cheating - it has to work these things out for itself - although perhaps it could *ask* for info). Why is that cheating? Would you never give instructions to a child about what to do? Taking instuctions is something that all intelligences need to be able to do, but it should be attempted to be minimised. I'm not saying it should take instructions unquestioningly either, ideally it should figure out whether the instructions you give are any use for it. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; 724342 Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 2/28/08, Mark Waser [EMAIL PROTECTED] wrote: I think Ben's text mining approach has one big flaw: it can only reason about existing knowledge, but cannot generate new ideas using words / concepts There is a substantial amount of literature that claims that *humans* can't generate new ideas de novo either -- and that they can only build up new ideas from existing pieces. That's fine, but the way our language builds up new ideas seems to be very complex, and it makes natural language a bad knowledge representation for AGI. For example: An apple pie is a pie with apple fillings. A door knob is a knob attached to a door. A street prostitute is prostitute working in the streets. So the meaning of AB depends on the *interactions* of A and B, and it violates the principle of compositionality -- where the meaning of AB would be somehow combined from A and B in a *fixed* way. An even more complex example: spread the jam with a knife draw a circle with a knife cut the cake with a knife rape the girl with a knife stop the train with a knife (with unclear meaning) So the simple concept do X with a knife can be interpreted in myriad ways -- it generates new ideas in complex ways. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
Too easy ;) One of the points in patch-space corresponds to X=center, Y=center, Scale=huge, so this patch is a rescaled version (say 20x20) of the whole image (say 1000x1000). In this 20x20 patch, the letter 'A' emerges naturally and can be reconstructed by the NN, and therefore be recognized. It will probably be salient, since it's far away in patch-space from the small A's in the Scale dimension. Far-away points in patch-space dont battle for salience. Your second example is solved analogously. Okay, time for diner now. Vision solved :) Regards, Durk On Mon, Mar 3, 2008 at 7:59 PM, Richard Loosemore [EMAIL PROTECTED] wrote: Kingma, D.P. wrote: On Mon, Mar 3, 2008 at 6:39 PM, Richard Loosemore [EMAIL PROTECTED] mailto:[EMAIL PROTECTED] wrote: The problems with bolting together NN and GA are so numerous it is hard to know where to begin. For one thing, you cannot represent structured information with NNs unless you go to some trouble to add extra architecture. Most NNs can only cope with single concepts learned in isolation, so if you show a visual field containing 5,000 copies of the letter 'A', all that happens is that the 'A' neuron fires. If you do find some way to get around this problem, your solution will end up being the tail that wags the dog: the NN itself will fade into relative insignificance compared to your solution. Well, you could achieve that (5000 registration of the letter 'A' with their corresponding position in the image) by using a sliding window over multiple rescaled (and maybe other transformations) transformations of the input image. This way, you get image patches for each window and scale (and maybe other transformations), and each patch can be a given a corresponding position in multidimensional space (e.g., an image patch with X and Y position and scale S has is a point in 3-dimensional space). For each of the produced points (patches) in the space, run the neural net to produce a lower-dimensional code and corresponding energy (= reconstruction quality). Now filter this space by let the points have local battles for salience using some heuristic (e.g. lower energy means higher salience) and filter out the low-salient points. This produces a filtered space with fewer points then the previous one, and each point containing a lower-dimensional code. In the example of the letter 'A', the above method would recognize all 5000 versions while remembering their individual input position. This presumes the neural net is properly trained on the letter 'A' and can properly reconstuct them (using Hinton's method). This should produce 5000 registrations of the letter 'A', while filtering out unimportant information. But you could take it a step further. For each image input, the above method creates a filtered, 3-dimensional space with points containing low-dimensional codes. This space can then again be harvested by taking patches with each patch containing /n/ points, each point containing an /m /dimensional code, so each patch being (/m/*/n/)./ /A neural net can be trained on lowering the dimension of these patches from (/m/*/n/) to something lower-dimensional. This process is quite similar to the one in the previous paragraph. What could /possibly /go wrong? :) Regards, Durk Kingma Excellent! Sounds like a perfect solution ;-). Oh, wait! What about. if the scene is structured in such a way that the 5,000 copies of the letter 'A' were actually scattered around in such a way that most (but not all) of them were arranged to form a huge letter 'A'? Would it then count 5,001 copies? Oh, and one more thing I forgot to mention that is in the same scene (how could I forget this one?): there are also a couple of women standing side by side, leaning against each other with their shoulders touching and keeping their bodies stiff and straight, forming the two sides of a letter 'A', and holding a model of a horizontally reclining woman between them at waist height, to form the crossbar of a letter 'A'. Could we get the NN to recognize, in the context of the overall scene, that here were actually 5,002 copies of the letter 'A'..? And if the scene had one single, rather small letter B over in the corner, would the NN find this funny? You have 30 minutes to devise an algorithm, Durk... :-). Richard Loosemore --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription:
Re: [agi] interesting Google Tech Talk about Neural Nets
On Mon, Mar 3, 2008 at 9:50 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: I'm increasingly convinced that the human brain is not a statistical learner, but a logical learner. There are many examples of humans learning concepts/rules from one or two examples, rather than thousands of examples. So I think that at a high level, AGI should be logic-based. But it would be interesting to integrate NN-based techniques to logic-based AI, especially in vision. (NN is also very weak at language processing.) One doesn't preclude another: if AGI can learn finite state machines statistically, it can then use them to carry out more 'logical' kinds of reasoning. Fast learning is also possible: it just takes a more similar state to evoke episodic memories than to evoke strong semantic memories. -- Vladimir Nesov [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] Thought experiment on informationally limited systems
Yes, an AGI will have to be able to do narrow AI. What you are doing here - and everyone is doing over and over and over - is saying: Yes, I know there's a hard part to AGI, but can I please concentrate on the easy parts - the narrow AI parts - first? If I give you a problem, I don't want to know whether you can take dictation and spell, I just want to know whether you can solve the problem - and not make excuses, or create distractions. It's simple - do you have any ideas about the problem of AGI - ideas for generalizing skills (see below) - cross-over ideas - or not? David: How intelligent would any human be if it couldn't be taught by other humans? Could a human ever learn to speak by itself? The few times this has happened in real life, the person was permanently disabled and not capable of becoming a normal human being. If humans can't become human without the help of other humans, why should this is a criteria for AGI? David Clark PS I am not suggesting that explicitly programming 100% of an AGI is either doable or desirable but some degree of detailed teaching must be a requirement for all on this list who dream of creating an AGI, no? -Original Message- From: Mike Tintner [mailto:[EMAIL PROTECTED] Sent: March-02-08 5:36 AM To: agi@v2.listbox.com Subject: Re: [agi] Thought experiment on informationally limited systems Jeez, Will, the point of Artificial General Intelligence is that it can start adapting to an unfamiliar situation and domain BY ITSELF. And your FIRST and only response to the problem you set was to say: I'll get someone to tell it what to do. IOW you simply avoided the problem and thought only of cheating. What a solution, or merest idea for a solution, must do is tell me how that intelligence will start adapting by itself - will generalize from its existing skills to cross over domains. Then, as my answer indicated, it may well have to seek some instructions and advice - especially and almost certainly if it wants to acquire a whole new major skill, as we do, by taking courses etc. But a general intelligence should be able to adapt to some unfamiliar situations entirely by itself - like perhaps your submersible situation. No guarantee that it will succeed in any given situation, (as there isn't with us), but you should be able to demonstrate its power to adapt sometimes. In a sense, you should be appalled with yourself that you didn't try to tackle the problem - to produce a cross-over idea. But since literally no one else in the field of AGI has the slightest cross-over idea - i.e. is actually tackling the problem of AGI, - and the whole culture is one of avoiding the problem, it's to be expected. (You disagree - show me one, just one, cross-over idea anywhere. Everyone will give you a v. detailed,impressive timetable for how long it'll take them to produce such an idea, they just will never produce one. Frankly, they're too scared). Mike Tintner [EMAIL PROTECTED] wrote: You must first define its existing skills, then define the new challenge with some degree of precision - then explain the principles by which it will extend its skills. It's those principles of extension/generalization that are the be-all and end-all, (and NOT btw, as you suggest, any helpful info that the robot will receive - that,sir, is cheating - it has to work these things out for itself - although perhaps it could *ask* for info). Why is that cheating? Would you never give instructions to a child about what to do? Taking instuctions is something that all intelligences need to be able to do, but it should be attempted to be minimised. I'm not saying it should take instructions unquestioningly either, ideally it should figure out whether the instructions you give are any use for it. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; 724342 Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.516 / Virus Database: 269.21.3/1308 - Release Date: 3/3/2008 10:01 AM --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
YKY: the way our language builds up new ideas seems to be very complex, and it makes natural language a bad knowledge representation for AGI. An even more complex example: spread the jam with a knife draw a circle with a knife cut the cake with a knife rape the girl with a knife stop the train with a knife (with unclear meaning) So the simple concept do X with a knife can be interpreted in myriad ways -- it generates new ideas in complex ways. YKY, Good example, but how about: language is open-ended, period and capable of infinite rather than myriad interpretations - and that open-endedness is the whole point of it?. Simple example much like yours : handle. You can attach words for objects ad infinitum to form different sentences - handle an egg/ spear/ pen/ snake, stream of water etc. - the hand shape referred to will keep changing - basically because your hand is capable of an infinity of shapes and ways of handling an infinity of different objects. . And the next sentence after that first one, may require that the reader know exactly which shape the hand took. But if you avoid natural language, and its open-endedness then you are surely avoiding AGI. It's that capacity for open-ended concepts that is central to a true AGI (like a human or animal). It enables us to keep coming up with new ways to deal with new kinds of problems and situations - new ways to handle any problem. (And it also enables us to keep recognizing new kinds of objects that might classify as a knife - as well as new ways of handling them - which could be useful, for example, when in danger). - Original Message - From: YKY (Yan King Yin) To: agi@v2.listbox.com Sent: Monday, March 03, 2008 7:14 PM Subject: Re: [agi] would anyone want to use a commonsense KB? On 2/28/08, Mark Waser [EMAIL PROTECTED] wrote: I think Ben's text mining approach has one big flaw: it can only reason about existing knowledge, but cannot generate new ideas using words / concepts There is a substantial amount of literature that claims that *humans* can't generate new ideas de novo either -- and that they can only build up new ideas from existing pieces. That's fine, but the way our language builds up new ideas seems to be very complex, and it makes natural language a bad knowledge representation for AGI. For example: An apple pie is a pie with apple fillings. A door knob is a knob attached to a door. A street prostitute is prostitute working in the streets. So the meaning of AB depends on the *interactions* of A and B, and it violates the principle of compositionality -- where the meaning of AB would be somehow combined from A and B in a *fixed* way. An even more complex example: spread the jam with a knife draw a circle with a knife cut the cake with a knife rape the girl with a knife stop the train with a knife (with unclear meaning) So the simple concept do X with a knife can be interpreted in myriad ways -- it generates new ideas in complex ways. YKY -- agi | Archives | Modify Your Subscription -- No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.516 / Virus Database: 269.21.3/1308 - Release Date: 3/3/2008 10:01 AM --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
On Mon, Mar 3, 2008 at 11:30 PM, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: Can you explain a bit more, your terms are too vague. I think statistical learning and logical learning are fundamentally quite different. I'd be interested in some hybrid approach, if it exists. Bayesian logic becomes something like Aristotelian logic when probability tends to 1. If statistical learning observes a perfect regularity, it forms a strong link, and classification becomes logical inference. Classification is performed in time, so that act of classification is an event that takes place after the events that were classified, and logical inference becomes a deterministic algorithm. These algorithms build up and help in learning other regularities and other algorithms. Maybe you mean something specific by logical learning that can't be supported by this kind of algorithm imitation? -- Vladimir Nesov [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
On 3/4/08, Mike Tintner [EMAIL PROTECTED] wrote: Good example, but how about: language is open-ended, period and capable of infinite rather than myriad interpretations - and that open-endedness is the whole point of it?. Simple example much like yours : handle. You can attach words for objects ad infinitum to form different sentences - handle an egg/ spear/ pen/ snake, stream of water etc. - the hand shape referred to will keep changing - basically because your hand is capable of an infinity of shapes and ways of handling an infinity of different objects. . And the next sentence after that first one, may require that the reader know exactly which shape the hand took. But if you avoid natural language, and its open-endedness then you are surely avoiding AGI. It's that capacity for open-ended concepts that is central to a true AGI (like a human or animal). It enables us to keep coming up with new ways to deal with new kinds of problems and situations - new ways to handle any problem. (And it also enables us to keep recognizing new kinds of objects that might classify as a knife - as well as new ways of handling them - which could be useful, for example, when in danger). Sure, AGI needs to handle NL in an open-ended way. But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. Otherwise, the knowledge stored in episodic memory would be open to interpretations and may need to errors in recall, and similar problems. YKY --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] would anyone want to use a commonsense KB?
Sure, AGI needs to handle NL in an open-ended way. But the question is whether the internal knowledge representation of the AGI needs to allow ambiguities, or should we use an ambiguity-free representation. It seems that the latter choice is better. Otherwise, the knowledge stored in episodic memory would be open to interpretations and may need to errors in recall, and similar problems. Rather, I think the right goal is to create an AGI that, in each context, can be as ambiguous as it wants/needs to be in its representation of a given piece of information. Ambiguity allows compactness, and can be very valuable in this regard. Guidance on this issue is provided by the Lojban language. Lojban allows extremely precise expression, but also allows ambiguity as desired. What one finds when speaking Lojban is that sometimes one chooses ambiguity because it lets one make ones utterances shorter. I think the same thing holds in terms of an AGI's memory. An AGI with finite memory resources must sometimes choose to represent relatively unimportant information ambiguously rather than precisely so as to conserve memory. For instance, storing the information A is associated with B is highly ambiguous, but takes little memory. Storing logical information regarding the precise relationship between A and B may take one or more orders of magnitude more information. -- Ben --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] interesting Google Tech Talk about Neural Nets
Kingma, D.P. wrote: Too easy ;) One of the points in patch-space corresponds to X=center, Y=center, Scale=huge, so this patch is a rescaled version (say 20x20) of the whole image (say 1000x1000). In this 20x20 patch, the letter 'A' emerges naturally and can be reconstructed by the NN, and therefore be recognized. It will probably be salient, since it's far away in patch-space from the small A's in the Scale dimension. Far-away points in patch-space dont battle for salience. Your second example is solved analogously. Okay, time for diner now. Vision solved :) Regards, Durk Yeah, modulo a few implementation details, that sounds about right. We can probably do language the same way tomorrow morning. ;-) Richard Loosemore On Mon, Mar 3, 2008 at 7:59 PM, Richard Loosemore [EMAIL PROTECTED] mailto:[EMAIL PROTECTED] wrote: Kingma, D.P. wrote: On Mon, Mar 3, 2008 at 6:39 PM, Richard Loosemore [EMAIL PROTECTED] mailto:[EMAIL PROTECTED] mailto:[EMAIL PROTECTED] mailto:[EMAIL PROTECTED] wrote: The problems with bolting together NN and GA are so numerous it is hard to know where to begin. For one thing, you cannot represent structured information with NNs unless you go to some trouble to add extra architecture. Most NNs can only cope with single concepts learned in isolation, so if you show a visual field containing 5,000 copies of the letter 'A', all that happens is that the 'A' neuron fires. If you do find some way to get around this problem, your solution will end up being the tail that wags the dog: the NN itself will fade into relative insignificance compared to your solution. Well, you could achieve that (5000 registration of the letter 'A' with their corresponding position in the image) by using a sliding window over multiple rescaled (and maybe other transformations) transformations of the input image. This way, you get image patches for each window and scale (and maybe other transformations), and each patch can be a given a corresponding position in multidimensional space (e.g., an image patch with X and Y position and scale S has is a point in 3-dimensional space). For each of the produced points (patches) in the space, run the neural net to produce a lower-dimensional code and corresponding energy (= reconstruction quality). Now filter this space by let the points have local battles for salience using some heuristic (e.g. lower energy means higher salience) and filter out the low-salient points. This produces a filtered space with fewer points then the previous one, and each point containing a lower-dimensional code. In the example of the letter 'A', the above method would recognize all 5000 versions while remembering their individual input position. This presumes the neural net is properly trained on the letter 'A' and can properly reconstuct them (using Hinton's method). This should produce 5000 registrations of the letter 'A', while filtering out unimportant information. But you could take it a step further. For each image input, the above method creates a filtered, 3-dimensional space with points containing low-dimensional codes. This space can then again be harvested by taking patches with each patch containing /n/ points, each point containing an /m /dimensional code, so each patch being (/m/*/n/)./ /A neural net can be trained on lowering the dimension of these patches from (/m/*/n/) to something lower-dimensional. This process is quite similar to the one in the previous paragraph. What could /possibly /go wrong? :) Regards, Durk Kingma Excellent! Sounds like a perfect solution ;-). Oh, wait! What about. if the scene is structured in such a way that the 5,000 copies of the letter 'A' were actually scattered around in such a way that most (but not all) of them were arranged to form a huge letter 'A'? Would it then count 5,001 copies? Oh, and one more thing I forgot to mention that is in the same scene (how could I forget this one?): there are also a couple of women standing side by side, leaning against each other with their shoulders touching and keeping their bodies stiff and straight, forming the two sides of a letter 'A', and holding a model of a horizontally reclining woman between them at waist height, to form the crossbar of a letter 'A'. Could we get the NN to recognize, in the context of the overall scene, that here were actually 5,002 copies of the letter 'A'..? And if the scene had one single, rather small letter B over in the corner, would
Re: [agi] Thought experiment on informationally limited systems
Will:Is generalising a skill logically the first thing that you need to make an AGI? Nope, the means and sufficient architecture to acquire skills and competencies are more useful early on in an agi development Ah, you see, that's where I absolutely disagree, and a good part of why I'm hammering on the way I am. I don't think many (anyone?) will agree with David, but many if not everyone will agree with you. Yes, the problem of generalising is the very first thing you tackle, and should shape everything you do - at least once you have moved beyond idle thought to serious engagement. If you're trying to develop a new electric battery, you look for that new chemical first (assuming that's what you reckon you'll need) - you don't start looking at the casing or other aspects of the battery. Anything peripheral you do first may be rendered totally irrelevant later on when you do discover that chemical and a total waste of time. And such, I'm sure, is the case with AGI. That central problem of generalising demands a total new mentality - a sea-change of approach. (You saw an example in my exchange with YKY. I think - in fact, I'm just about totally certain - that generalising demands a system of open-ended concepts like ours. Because he isn't directly concerned with the generalising problem, he wants a closed-ended, unambiguous language - which is in fact only suitable for narrow AI and, I would argue, a waste of time). P.S. It's a bit sad - you started this thread with a generalising problem, now you're backtracking on it. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com