Ken, Wow. I was going to say, this is one of the most interesting posts I have read on the AGI list in a while, until I realized it wasn't on the AGI list. Too bad. I have copied this response and your original email (below) to the AGI list to share the inspiration.
In the following I have copied certain parts of your post and followed them by questions or comments. >KEN LAWS=====> And much of the advanced robotic planning software developed at NASA Ames is based on particle filters, a method of representing probability distributions as they pass through various nonlinear transformations. (It remains to be seen whether any of this software will find use in actual missions, but I'm betting it will be used in the next Mars rovers.) ED PORTER=====> Probability particle filters sounds cool. I assume it means you only consider or transmit information about probability (or probabilistic implication) distributions or changes in such distributions that have over a certain concentration in a given portion of space/time to those locations in space/time. Is that correct? And what sort of non-linear transforms are you talking about? >KEN LAWS=====> Artificial neural networks, like humans, have a remarkable ability to deal with noise inputs and under-constrained models, but the learning is very slow. That's why evolution has provided us with a priori brain structuring, instead of a tabula rasa mind. It makes the learning tractable. ED PORTER=====> Other than the way sensory, homeostatic, body sensation, and other info is mapped into our brains, the cortico-basil-ganglia-thalamic feedback loop, the cortico-cerebellum-thalamic feedback loop, and the other pre-designed plug and play interface to the reptile brain (all of which establish a certain type of architecture and control structure), what are your talking about? I would be very interesting in knowing what type of constraint, other than these basic architectural constraints are involved. I just attended a lecture this weekend where a Harvard researcher on the unconscious mind said that one-day-old babies have been shown to be able to mimic a few basic facial expressions, such as sticking out their tongue and putting their mouth into a small circle, as when saying "who". This is hard to understand, because one would imaging that by this age a child fresh from the fog of the womb would not have had time to build up the visual patterns enabling them to recognize facial features, and further more would not have had time to map the correspondence between that blob of pink sticking out of the hole in somebody's face and the baby's own tongue, or own mouth. (I don't know how much evidence this study has.) I have not had anybody explain to me how such instinctual programming could be represented in advance of the learned, experientially derived patterns out of which most mind patterns are represented. The one exception is Sam Adam's explained in an off-stage discussion at the Singularity Conference, about how new-borns are designed to visually focus on the female areola because tests have shown their visual system is pre-wired to detect circular patterns.) >KEN LAWS=====>For those who prefer fine-scale brain/mind modeling, look into the decades of theoretical and simulation work by the SOAR community, and by the APEX community, for human sensor/manipulator learning simulations. ED PORTER=====> I haven't read about SOAR for ten years. It struck me as a generalized expert system (if-then rules), but one with a relatively enlightened goal structure and learning structure for an expert system. Has it grown into a real contender for human level AGI, and what sort of tasks its is currently actually capable of. Re APEX, I have never heard it. Have you any good URLs for summarizing the nature and capabilities of each. >KEN LAWS=====> ... Early pattern recognition researchers had high hopes for statistical learning, but eventually realized that the magic is almost always in feature extraction rather than the statistical back end. Represent a problem well and it may be easy to solve; badly and you'll need more computing power than you can afford. .. ED PORTER=====> I think most intelligent AGI models envision a system that has many representations which compete for existence based on usefulness. This is one way of addressing this problem. Another is to understand that for certain complex problems, such as the those of the type we who are trying to design AGI often face, part of the problem is creating the proper novel representation, and that can involve a lot of trial and error and exploration that hopefully tends to build up patterns representing partial or not quite right solutions that over time probabilistically increase the chance of synthesizing a better representation. The Novamente-OpenCog approach should be able to use both of these methods to find proper representations, although the system should be biased toward learning how to learn, which includes learning how to select appropriate representations. Do you agree? With regard to your general discussion about different levels of sensory input, my feeling is that one can develop and test many of the aspects of a generalized AGI architecture on many different types and sizes of problems, although certain additional issue may have to be addressed as you change the nature of the problems, and the diversity and size of them. If one takes the Novamente/OpenCog approach, it seems to me that many of the same issues would be involved in creating an NL understanding and generating system as in creating a visual understanding and outputting system, as would be involved in creating a robotic environmental/mechanical/goal state recognition and generation system. In my mind purely cerebral AGI's are mind robots, involving goal, hierarchical behaviors, task feedback monitoring, etc, just like a robot. Thus I think the basic idea behind OpenCog makes sense, if it can get a large enough community to give it some momentum, the basic architecture should emerge, and the basic architectures should be flexible enough to adapt itself to many different types of problems. Do you agree? If not, what other approaches would you suggest? Thanks for your post. Ed Porter -----Original Message----- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Ken Laws Sent: Tuesday, December 04, 2007 4:13 PM To: [EMAIL PROTECTED] Subject: Re: A global approach to AI in virtual, artificial and real worlds > -- The car from > Stanford that won the original DARPA Grand Challenge was controlled > based on probabilistic robotics.... And much of the advanced robotic planning software developed at NASA Ames is based on particle filters, a method of representing probability distributions as they pass through various nonlinear transformations. (It remains to be seen whether any of this software will find use in actual missions, but I'm betting it will be used in the next Mars rovers.) However, as Ben said, probabilistic methods have computational tractability issues. You can't just "use a physical stereoscopic array of cameras and microphones to build a basic object library and vocabulary via stationary video and audio sensors" because of the problem's dimensionality. Artificial neural networks, like humans, have a remarkable ability to deal with noise inputs and under-constrained models, but the learning is very slow. That's why evolution has provided us with a priori brain structuring, instead of a tabula rasa mind. It makes the learning tractable. > We, humans, are what we are because we have 5 senses > (6 if you addproprioception). And many more, if you care to include them. See http://en.wikipedia.org/wiki/Senses . And even that leaves out the social senses that we are discovering in the form of mirror neurons. One reason we have so many senses is that they help keep pattern recognition problems from being so massively under-determined. The only question that a situated agent really has to answer, moment by moment, is "What do I do now?" The more potentially relevant inputs, the easier and more reliable the solution and the less need there is for high-level reasoning. Consider that man on the beach, for instance. He doesn't need to compile thousands of years of experience with near and distant dogs in various crowd situations. He can look around and see that no one else is worried about the distant dog, hence he needn't worry either. The "wisdom of crowds" has already compiled and presented the probability that he needs. One problem with simulated worlds is that they lack this physical and social richness. This forces attention away from the immediate "What do I do now?" problem toward a more abstract level of reasoning. However, the toy worlds also provide a pre-parsed experience of reality, much like the simplified world that parents present to their children. This can be helpful. Early pattern recognition researchers had high hopes for statistical learning, but eventually realized that the magic is almost always in feature extraction rather than the statistical back end. Represent a problem well and it may be easy to solve; badly and you'll need more computing power than you can afford. (A classic example is the question of whether you can cover a chess board with dominoes if you've first removed the two chess board squares at opposite corners.) The advantage of a simulated world is that the coarse-scale learning is easier; the disadvantage is that you can only learn the world structure that is presented. One of the lessons of the expert systems fad was that the world is seldom fully represented from any one perspective, and so one can't expand the capabilities of an expert system merely by adding more knowledge. It may be necessary to use multiple representations, plus meta-level arbiters to combine them. (Exactly what human brain centers seem designed to do. For instance, we humans use something like 20 different analogies to reason about love, according to George Lakoff. Love is a road; love is a carriage; etc. Robert J. Sternberg has likewise identified at least two dozen conflicting models of marriage.) Roboticists learned, many decades ago, that success in simulation is only a tiny step toward success in the real world. Labs that could afford real hardware had no interest in simulations (even when I, as a funder, urged that approach). I see that Stanford now has impressive simulation software for robot dynamics, so perhaps the earlier lesson was over-learned. Still, beware the idea that simulation addresses real-world problems. It may, or it may not. Ben commented that he can't afford a robotic approach for this project, and I agree. However, those who really want to go that route shouldn't despair; there are ways to do it. Remember Seymour Papert's turtle graphics, and the robot turtles that resulted? A more modern distributed educational robotics project was developed by Prof. Illah Nourbakhsh as the Personal Rover Project, http://www.cs.cmu.edu/~personalrover/, which has now become the TeRK project, http://www.terk.ri.cmu.edu/ . I haven't been following it, but Nourbakhsh did develop a roving robotic platform affordable by individual children, plus ways of running robots remotely and of collaborating on the development of software libraries. For those who prefer fine-scale brain/mind modeling, look into the decades of theoretical and simulation work by the SOAR community, and by the APEX community, for human sensor/manipulator learning simulations. I believe that these and OpenCog have much in common, though SOAR and APEX are more closely tied to human emulation at the level of memory chunking and millisecond response times. OpenCog is focused on a much coarser level of Turing test. It can be taken in many directions. I'm a bit skeptical of the dog-parrot-monkey trajectory though, because I'm not convinced that interest can be sustained or that the results will teach us much of anything. (But what do I know? Not much about it, yet.) One concern is that the whole attraction of training a dog is that it is difficult to do well, hence challenging to attempt and impressive when accomplished. Designing an AI to be difficult to train seems counter-productive, but necessary for the application. (I acknowledge, of course, that even "difficult to train" is an advance over the recent state of the art, "impossible to train.") It's a baby step, but at least it's a step. One final note: > I think that issues like that would be better discussed on the AGI list > [email protected] > than on this list which is supposed to be devoted to the OpenCog project > in particular ;-) Wow, that takes me back to the old Usenet days, and to my time as the AIList Arpanet bboard moderator. We techies keep organizing societies around the Dewey Decimal System, but the truth is that each discussion has a social dimension. We are gathered in this virtual time and place to discuss specific issues and implementations, but also to discuss whatever the hell we feel like discussing. The same topic will develop differently in different lists, entailing differing contributions and effects. I'm not a member of AGI, for instance, so this contribution would be absent there. The best way to handle discussion overlap is for interested people to report on any discussions that need cross-communication, summarizing relevant questions and conclusions and also pointing people to other conversations they might wish to join. Indeed, such reporting can also take place across time, as in some of my comments above. Let a thousand blogs bloom. -- Ken Laws --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "OpenCog.org (Open Cognition Framework)" group. To post to this group, send email to [EMAIL PROTECTED] To unsubscribe from this group, send email to [EMAIL PROTECTED] For more options, visit this group at http://groups.google.com/group/opencog?hl=en -~----------~----~----~----~------~----~------~--~--- ----- 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=8660244&id_secret=72059876-578332
<<attachment: winmail.dat>>
