Re: [agi] Conway's Game of Life and Turing machine equivalence
Honestly, it seems to me pretty clearly that whatever Richard's thing is with complexity being the secret sauce for intelligence and therefore everyone having it wrong is just foolishness. I've quit paying him any mind. Everyone has his own foolishness. We just wait for the demos. - 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=50720641-1f7528
Re: [agi] a2i2 news update
Not only that, if you work in IT, you might think, considering how poorly adding people to a project works, is he getting desperate or just being foolish? andi On Wed, 25 Jul 2007 18:25:56 -0700 (PDT), Ton Genah wrote Just increasing the number doesnt guarantee a clear path towards increased intelligence . This seems to be the important issue in current AI and not the number! Ton Mike Tintner [EMAIL PROTECTED] wrote: Of course, numerical comparisons are petty, unfair and invidious. But being that sort of person, I can't help noticing that Peter is promising to increase his staff to 24 soon. Will that give him the biggest AGI army? How do the contenders stack up here? - 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=231415id_secret=25415821-b920f1
Re: [agi] News bit: Carnegie Mellon unveils Internet-controlled robots anyone can build
Bob Mottram wrote: I have thought about making a robotic artist in the distant past. Some of the first robots which I remember seeing in the 1980s used the LOGO language to produce sketches using different coloured pens. You could maybe do something similar to that, with a mouse-like body and a few differently coloured pens mounted on servos (there is plenty of scope on the Qwerk to add multiple servos). Alternatively you could build something more like a manipulator arm, and attach pens as if they were fingers on separate servos. Ben wrote: I like the idea of different fingers having different magic markers on the tips of them ;-) Wow, that sounds like a great idea, guys, thanks for bringing it up! One thing, though, I probably wouldn't want to integrate the pens or whatever into the device, since pens dry out and are consumable. I'd prefer a general manipulator. Also, having them all at once seems like it's only about making it faster, but computers and robots are things with inifinite patience, so I would guess one color at a time should not be a problem for them, though of course, I'm sure they could handle all them at once. And all this, too, is reminiscent of the various automatic fabrication robots that have been popping up, with plastic deposition and stuff. None of those integrate vision systems, though, which you might use in a general robot. It might be nice also to have a robot that could handle drills and saws, and other machining tools, for a really productive system. But that's what industrial robots do, I guess. - 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=231415user_secret=fabd7936
Re: Goals of AGI (was Re: [agi] AGI interests)
Not only is each movie different for each person, it is different each time one person sees it. The movie itself is different from the movie-witnessing experience, and there seems to be a feeling that you could compress it by just grabbing the inner experience. But you notice different things each time. And more often than just trying to take away the factual bits of what happened, in any situation we are much more interesting in extracting the meaning than teh simple fact about what just happened. The implications, the point of any particular action. Even in speech, we aren't trying to remember sounds, but which actual word-sound-meaning unit was intended, since it is always ambiguous. And the narrative nature of knowledge I think was mentioned, and I think it's helpful to point out a part of narrative that is often neglected. A narrative is a chronologically ordered telling of a situation that has some moral or point. This moral or point is an important part of the meaning, just as the is the factual content, but it is not nearly so absolutely or clearly defined. Very often, if not mostly at least for TV shows, the moral is just that good triumphs over evil. But if you leave it out in a story, people find themselves not caring and thus not remembering. andi On Wed, 18 Apr 2007 09:24:51 +0200, Kingma, D.P. wrote [Spelling corrected and reworded...] I'm not convinced by this reasoning. First, the way individuals store audiovisual information differs, simply because of slight differences in brain development (nurture). Also, memory is condensed information about the actual high-level sensory/experience information. The actual 45kb memory of a movie is therefore quite personal to the subject. Recall of a photo/video is more like an impressionistic painting then an actual photo. An AGI that reconstructs a movie from 45kb human-ish compressed memory will have to make up 99.99% of video and audio. A very educated guess, but still a guess. Compare it with an extremely talented photorealistic animator human that, purely from memory, creates a reconstruction of a scene from The Matrix. Wouldn't you notice the difference in experience? On 4/18/07, Matt Mahoney [EMAIL PROTECTED] wrote: On 4/17/07, James Ratcliff [EMAIL PROTECTED] wrote: A simple list, or set of goals for an AGI to accomplish reasonably I would find very useful, and something to work for. I think an important goal is to solve the user interface problem. The current approach is for the computer to present a menu of choices (e.g. a set of icons, or automated voicemail press or say 'one'), which is hardly satisfactory. An interface should be more like Google. I tell the computer what I want and it gets it for me. In http://cs.fit.edu/~mmahoney/compression/rationale.html I argue the equivalence of text compression with AI. I would therefore set a goal of matching humans at text prediction (about 1 bit per character). Humans use vast knowledge and reasoning to predict strings like All men are mortal. Socrates is a man. Therefore . An AGI should be able to make predictions as accurately as humans given only a 1 GB corpus of text, about what a human could read in 20+ years. I would go further and include lossy compression tests. In theory, you could compress speech to 10 bits per second by converting it to text and using text compression. The rate at which the human brain can remember video is not much greater, probably less than 50 bps*. Therefore, as a goal, an AGI ought to be able to compress a 2 hour movie to a 45 KB file, such that when a person views the original and reconstructed movie on consecutive days (not side by side), the viewer will not notice any differences. It should be able to do this after training on 20 years of video. The purpose of this goal is that such an AGI could also perform useful tasks such as reduce a video to a verbal description understandable by humans, or given a script, produce a movie. These tasks would be trivial extensions of the compression process, which would probably consist of describing a movie using text and augmenting with some nonverbal data such as descriptions of faces and voices in terms that humans cannot easily express. *50 bps is probably high. Tests of image recall by Standing [1] suggest that a picture viewed for 5 seconds is worth about 30 bits. [1] Standing, L. (1973), Learning 10,000 Pictures, Quarterly Journal of Experimental Psychology (25) pp. 207-222. -- 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/?; --- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to:
Re: Goals of AGI (was Re: [agi] AGI interests)
It occurs to me the problem I'm having with this definition of AI as compression. There are two different tasks here, recognition of sensory data and reproduction of it. It sounds like this definition proposes that they are exactly equivalent, or that any recognition system is automatically invertable. I simply doubt that this can be true, using a principle (I have no proof for but I hold) that meaning--something we use to recognize equivalence-- is just not the same for different peceptual events. An another example I use to think about it is how difficult it is trying to draw a reproduction of a picture from memory, and how different the task is from drawing a copy is from analyzing the elements in a picture. Reproducing visual information is different from conceptual scene decomposition. On Wed, 18 Apr 2007 16:45:04 -0700 (PDT), Matt Mahoney wrote --- Matt Mahoney [EMAIL PROTECTED] wrote: 3. Standing [3] had subject memorize 10,000 pictures, one every 5.6 seconds over 5 days. Two days later they could recall about 80% in tests. This is about the result you would get if you reduced each picture to a 16 bit feature vector and checked for matches. This is a memory rate of 0.3 bits per second. That should be 3 bits per second. -- 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=231415user_secret=fabd7936
Re: [agi] Why evolution? Why not neuroscience?
Eugen discussed evolution as a development process. I just wanted to comment about what Minsky said in his talk (and I have to thank this list for pointing out that resource). He said that the problem with evolution is that it throws away the information about why bad solutions failed. That really has affected my thinking about it, since I was thinking that it at least sounded like a pretty good idea. But it is really a very terrible waste, and I no longer really think it is such a great model to use. I'm not sure what adaptations could be made to make up fo that loss, but surely there could be an improvement over evolution, even in a sytem of random generation and recombination with competitive survival. andi - 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/?list_id=303
Re: [agi] general weak ai
I can't speak for Minsky, but I would wonder what advantage would there be for having only one agent? I think he talks about the disadvantages. How is it going to deal with naturally different sorts of management problems and information? It seems like it's just a better aproach to have a system that has several different resources working together. BTW, Minsky has gone from calling them agents to calling them resources in _The Emotion Machine_. andi On Fri, 9 Mar 2007 19:31:38 -0500, J. Storrs Hall, PhD. wrote Not at all. the agent that does the pointing is just a build a deck agent (or, more likely, a society of deck) that gets activated when deck- building is the thing to do. I don't know Minsky's ultimate take on the subject, but I don't see any problem with putting one agent in charge of the whole business, especially for the duration of a specific task, as long as it isn't supposed to have any more capabilities per se than any other agent. Josh On Friday 09 March 2007 07:36, Pei Wang wrote: On 3/9/07, J. Storrs Hall, PhD. [EMAIL PROTECTED] wrote: If I understand Minsky's Society of Mind, the basic idea is to have the tools be such that you can build your deck by first pointing at the saw and saying you do your thing and then pointing at the hammer, etc. The tools are then in turn made of little guys who do the same to their tools, ad infinitum (or at least ad neuronium). This understanding assumes a you who does the pointing, which is a central controller not assumed in the Society of Mind. To see intelligence as a toolbox, we would have to assume that somehow the saw, hammer, etc. can figure out what they should do in building the deck all by themselves. Pei - 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/?list_id=303 - 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/?list_id=303
Re: [agi] general weak ai
On Tue, 6 Mar 2007 09:49:47 +, Bob Mottram wrote Some of the 3D reconstruction stuff being done now is quite impressive (I'm thinking of things like photosynth, monoSLAM and Moravec's stereo vision) and this kind of capability to take raw sensor data and turn it into useful 3D models which may then be cogitated upon would be a basic prerequisite for any AGI operating in the real world. I'm sure that these and other similar methods are soon destined to be fall into the bracket of being no longer AI, instead being considered as just another computational tool. In the past I've tried many ad-hoc vision experiments, which would certainly come under the narrow AI label, but I now no longer believe that this kind of approach is a good way to proceed. Far more straightforward, albeit more computationally demanding, techniques give a general solution to the vision problem which is not highly specific to any particular kind of domain or environment. Under this system applications which are often treated separately, such as visual navigation and object recognition, actually turn out to be the same algorithm deployed on different spatial scales (maybe a classic case of physics envy!). Well what is intelligence if not a collection of tools? One of the hardest problems is coming up with such tools that are generalizable across domains, but can't that just be a question of finding more tools that work well in a computer environment, instead of just finding the ultimate principle. Ideas like gofai symbolic symbol manipulation and Bayesian decision networks seem to me to naturally just fit into the idea of part of an AI kit, but I personally would want this kit to be more compatible with the post AI techniques. Another example, that someone is using AI is often recognized by them using some kind of search instead of some algorithm, like gradient ascent or resolution, but there's not reason why a system can't throw multiple approaches at a problem, and maybe fall back on some general search when needed. And maybe that's why I think an AI's proper world is controlling a computer (ie. a PC), so it can just run programs whenever it needs to get things done. andi - 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/?list_id=303
Re: [agi] SOTA
On Fri, 20 Oct 2006 22:15:37 -0400, Richard Loosemore wrote Matt Mahoney wrote: From: Pei Wang [EMAIL PROTECTED] On 10/20/06, Matt Mahoney [EMAIL PROTECTED] wrote: It is not that we can't come up with the right algorithms. It's that we don't have the computing power to implement them. Can you give us an example? I hope you don't mean algorithms like exhaustive search. For example, neural networks which perform rudamentary pattern detection and control for vision, speech, language, robotics etc. Most of the theory had been worked out by the 1980's, but applications have been limited by CPU speed, memory, and training data. The basic building blocks were worked out much earlier. There are only two types of learning in animals, classical (association) and operant (reinforcement) conditioning. Hebb's rule for classicical condioning proposed in 1949 is the basis for most neural network learning algorithms today. Models of operant conditioning date back to W. Ross Ashby's 1960 Design for a Brain where he used randomized weight adjustments to stabilize a 4 neuron system build from vacuum tubes and mechanical components. Neural algorithms are not intractable. They run in polynomial time. Neural networks can recognize arbitrarily complex patterns by adding more layers and training them one at a time. This parallels the way people learn complex behavior. We learn simple patterns first, then build on them. I initially wrote a few sentences saying what was wrong with the above, but I chopped it. There is just no point. What you said above is just flat-out wrong from beginning to end. I have done research in that field, and taught postgraduate courses in it, and what you are saying is completely divorced from reality. Richard Loosemore I have simply taken maybe one and say a half (because it seems like every ai survey class has to touch upon neural nets again) graduate classes on the subject, and not taught or done research in the area, but I recognized that most of that was wrong. I at least hold out the possibility that neural nets can be made useful with some greater theory about architectures and much greater computing power. I think it would be worthwhile for you to take the time to list what you think the flaws were, if only to open the possibility for some positive recomendations for research directions. Even thought you may be completely disillusioned, maybe not everyone is. - 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/[EMAIL PROTECTED]
Re: [agi] Computer monitoring and control API
I wrote: I just had a notion. The proper sensory input and motor output for an AI is the computer screen (and sound input and regular keyboard and mouse input). One thing that needs to exist is a freely available standard API for these things, so people can work on them, plus implementations for the platforms that people use. My hope is that it would give different researchers, especially all those lone wolves out there, something intercompatible to work with. It also seems possible that this could be a common mechanism for the different systems to work together, in a sort of extension of the Blackboard model. And, as a lighter element of it, I'd really like it if these projects could use video games, because they more and more have become very sophisticated real-world modelling tools. andi And Richard Loosemore asked for clarification: Can you be more specific about what this would entail? I can think of several interpretations of what you say, but am not sure which you mean. Well, if you can think of several interpretations, then why don't you pick one you like? I was thinking along the lines of java.awt.Robot. I only had a vague recollection of it, and I never used it, and looking at it again now, I think it is exactly what I was thinking of. Another reason I thought of it is that Stan Franklin's Ida model uses e-mail as a sort of sensory-motor and that's a kind of subset of this notion. It seems like the standard reactions people have when they wonder what an artificial intelligence is going to do is either sit in a box and answer questions or control a physical robot clunking around the world. I would simply propose that one other useful answer is to control and use a computer the way a person might control a computer. This would mean that it could use all manner of existing tools to multiply whatever power its additional intelligence adds. But one of the tricky bits of the idea is having something sufficiently general and useful enough to make a contribution. As I mentioned, there is a Java class that does the kind of thing I'm interested in. And it's probably straightforward to have this kind of thing in other imperative languages. But how would you have a neural network system interface to it? I don't know, maybe the API idea is foolish. I've never tried to design one, so I don't particularly know what's involved or if it's even a good idea. The really basic functions I would expect are an ability to capture a piece of the screen, to control the mouse, and input keyboard events. I think a very valuable addition would be to discover a character (or piece of text) that's at a particular location, so reading in text from a screen would be easier. We have to do this to use a computer and any agent using a computer would need to do this anyway, so it would be just more useful to add that in at the beginning. Unfortunately, that could be a tricky bit of code, but it is miles away from OCR, so it isn't unreasonable. I also mentioned having access to the sound streams. People can get away with not using the sound on a computer, so clearly it wouldn't be necessary for an artificial agent using it to use it, but it might make a valuable addition. And it might a useful feature if part of this interface enabled an AI to simply watch what a person (or conceivably another agent) was doing, which could open opportunities for some kind of instruction or learning. andi - 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/[EMAIL PROTECTED]
[agi] Computer monitoring and control API
I just had a notion. The proper sensory input and motor output for an AI is the computer screen (and sound input and regular keyboard and mouse input). One thing that needs to exist is a freely available standard API for these things, so people can work on them, plus implementations for the platforms that people use. My hope is that it would give different researchers, especially all those lone wolves out there, something intercompatible to work with. It also seems possible that this could be a common mechanism for the different systems to work together, in a sort of extension of the Blackboard model. And, as a lighter element of it, I'd really like it if these projects could use video games, because they more and more have become very sophisticated real-world modelling tools. andi - 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/[EMAIL PROTECTED]
RE: [agi] Failure scenarios
Peter Voss mentioned trying to solve the wrong problem is the first place as a source for failure in an AGI project. This was actually this first thing that I thought of, and it brought to my mind a problem that I think of when considering general intelligence theories--object permanence. Now, I think it's established that babies have to learn the concept of object permanence. They are probably genetically inclined to do so, but they still have to acquire the concept. You don't have to have an anthromorphic system, certainly, but to me this says profound things about what intelligence itself could possibly be if you could be intelligent before having such a simple concept, and then have some way that you develop it. One of the implications for me is that intelligence almost certainly requires a some kind of causal, sensory-motor interaction with the world. Object permanence itself is an abstract notion from the various practical behaviors involved with it, so I would also not expect it to be just a piece of knowledge that was added to a system. It's a hard question of what it actually is, speaking to the nature of generalization of knowledge. And while this is only one concept of many, it and others like it are the kind of problems that I see as getting missed in the sorts of general intelligence theories that I see. - 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/[EMAIL PROTECTED]
RE: [agi] Why so few AGI projects?
PS. http://adaptiveai.com/company/opportunities.htm This also reminds me of something, and I know it's true of myself, and I think it might be generally true. It seems like people tend to have their own ideas of what they want to be done, and they are just not very interested in working on someone else's idea or concept. I know that's why I am not working on Stan's project. It could also be why I haven't been aggressive enough to really go after working on one of the other projects that are out there, a2i2 included. It seems like there are quite a few lone AI hackers out there. And this is a specific case of something I have found: nobody likes to be told what to do--some people tolerate it more than others, but nobody likes it. andi - 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/[EMAIL PROTECTED]