RE: [agi] Early Apps.
Gary Miller wrote: *** I guess I'm still having trouble with the concept of grounding. If I teach/encode a bot with 99% of the knowledge about hydrogen using facts and information available in books and on the web. It is now an idiot savant in that it knows all about hydrogen and nothing about anything else and it is not grounded. But if I then examine the knowledge learned about hydrogen for other mentioned topics like gases, elements, water, atoms, etc... And teach/encode 99% of of the knowledge on these topics to the bot. Then the bot is still an idiot savant but less so isn't it better grounded? A certain amount of grounding I think has occurred by providing knowledge of related concepts. If we repeat this process again, we may say the program is an idiot savant in chemistry. ... I will agree that today's bots are not grounded because they are idiot savants and lack the broad based high level knowledge with which to ground any given fact or concept. But if I am correct in my thinking this is the same problem that Helen Keller's teacher was faced with in teaching Helen one concept at a time until she had enough simple information or knowledge to build more complex knowledge and concepts upon. *** What you're describing is the Expert System approach to AI, closely related to the common sense approach to AI. Cycorp takes this point of view, and so have a whole lot of other AI projects in the last few decades... I certainly believe there's some truth to it. If you encoded a chemistry textbook in formal logic, fed it into an AI system, and let the AI system do a lot of probabilistic reasoning and associating on the information, then you'd have a lot of speculative uncertain intuitive knowledge generated in the system, complementing the hard knowledge that was explicitly encoded. If you encoded a physics textbook and a bio textbook as well, you could have the system generate uncertain, intuitive cross-domain knowledge in the same way. In fact, we are doing something like this in Novamente now, for a bioinformatics application. We're feeding in information from a dozen different bio databases and letting the system reason on the integrated knowledge right now we're at the feeding in stage. Unlike some anti-symbolic-AI extremists, I think this sort of thing can be *useful* for AGI. But I think it can only be a part of the picture. Whereas I think experience-based learning is a lot more essential I don't think that a pragmatically-achievable amount of formally-encoded knowledge is going to be enough to allow a computer system to think deeply and creatively about any domain -- even a technical domain about science. What's missing, among other things, is the intricate interlinking between declarative and procedural knowledge. When humans learn a domain, we learn not only facts, we learn techniques for thinking and problem-solving and experimenting and information-presentation .. and we learn these in such a way that they're all mixed up with the facts In theory, I believe, all this stuff could be formalized -- but the formalization isn't pragmatically possible to do, because we humans don't explicitly know the techniques we use for thinking, problem-solving, etc. etc. In large part, we do them tacitly, and we learn them tacitly.. When we learn a new domain declaratively, we start off by transferring some of our tacit knowledge from other domains to that new domain. Then, we gradually develop new tacit knowledge of that domain, based on experience working in the domain... I think that this tacit knowledge (lots of uncertain knowledge, mixing declarative procedural) has got to be there as a foundation, for a system to really deploy factual knowledge in a creative fluent way... *** I think we cut and paste what we are trying to say into what we think is the correct template and then read it back to ourselves to see if it sounds like other things we have heard and seems to make sense. *** I think this is a good description of one among many processes involved in language generation... I also think there's some more complex unconscious inference going on, than is implied by your statement. It's not a matter of cutting and pasting into a template, it's a matter of recursively applying a bunch of syntactic rules that build up complex linguistic forms from simpler ones. The syntactic buildup process has parallels to the thought-buildup process, and the two sometimes proceed in synchrony, which is one of the reasons formulating thoughts in language can help clarify them. I dealt with some of these issues -- on a conceptual, not an implementational level - in a chapter in my book from complexity to creativity, entitled Fractals and Sentence Production: http://www.goertzel.org/books/complex/ch9.html If I were to rewrite that chapter now, it would have a lot of stuff on probabilistic inference unification grammars -- richer and better details, enhanced by the particular
RE: [agi] Early Apps.
Ben Goertzal wrote: I don't think that a pragmatically-achievable amount of formally-encoded knowledge is going to be enough to allow a computer system to think deeply and creatively about any domain -- even a technical domain about science. What's missing, among other things, is the intricate interlinking between declarative and procedural knowledge. When humans learn a domain, we learn not only facts, we learn techniques for thinking and problem-solving and experimenting and information-presentation .. and we learn these in such a way that they're all mixed up with the facts What you're describing is the Expert System approach to AI, closely related to the common sense approach to AI. ... I agree that as humans we bring a lot of general knowledge with us when we learn a new domain. That is why I started off with the general conversational domain and am now branching into science, philosophy, mathematics and history. And of course the AI can not make all the connections without being extensively interviewed on a subject and having a human help clarify it's areas of confusion just as a parent answers questions for a child or a teacher for a student. I am not in fact trying to take the exhaustive approach one domain at a time approach but rather to teach it the most commonly known and requested information first. My last email just used that description to identify my thoughts on grounding. I am hoping that by doing this and repeating the interviewing process in an iterative development cycle that eventually the bot will eventually be able to discuss many different subjects at a somewhat superficial level much as same as most humans are capable of. This is a lot different from the exhaustive definition that Cyc provides for each concept. I view what I am doing distinct from expert systems because I do not yet use either a backward or forward inference engine to satisfy a limited number of goal states. The knowledge base is not in the form of rules but rather many matched patterns and encoded factoids of knowledge many of which are transitory in nature and track the context of the conversation. Each pattern may trigger a request for additional information like an expert system. But the bot does not have a particular goal state in mind other that learning new information unless a specific request of it is made by the user. I also differ from Cyc in that realizing the importance of English as a user interface from the beginning, all internal thoughts and goal states occur as an internal dialog in English. This eliminates the requirement to translate an internal knowledge representation to an external natural language other than providing one or response patterns to specific input patterns. It also makes it easy to monitor what the bot is learning and whether it is making proper inferences because it's internal thought process is displayed in English while in debug mode.. The templates which generate the responses in some cases do have conditional logic to determine which output template is appropriate response based on the AI's personality variables and the context of the current conversation. Variables are also set conditionally to maintain metadata for context. If the references a male in it's response [He] and [Him] get set vs. [Her] and [She] if a female is referenced. [CurrentTopic], [It], [There] and [They] are all set to maintain backward contextual references. I was able to find a few references to the Common Sense approach to AI on google and some of the difficulties in achieving it. And I must admit I have not yet implemented non-monotonic reason or probabilistic reasoning as of yet. I am not under the illusion that I am necessarily inventing or implementing anything that has not been conceived of before. As Newton says if I achieve great heights it will be because I have stood on the shoulders of giants. I just see the current state of the art and think that it can be made much better. I do not actually know how far I can take it while staying self-funded, but hopefully by the time my money runs out it will demonstrate enough utility and potential to be of value to someone. I think I like the sound of the Common Sense Approach to AI though. I can't remember the last time anyone accused me of having common sense, but I like the sound of it! I don't think AI is absent sufficient theory, just sufficient execution. I feel like the Cyc Project's heart was in the right place and the level of effort was certainly great, but perhaps the purity of their vision took priority over usability of the end result. Is any company actually using Cyc as anything other than a search engine yet? That being said other than Cyc I am at a loss to name any serious AI efforts which are over a few years in duration and have more that 5 man years worth of effort (not counting promotional and fundraising). The Open Source efforts are interesting and have some utility but are
[agi] Thinking may be overrated.
Perhaps thinking is overrated. It sometimes seems the way progress is made, and lessons learned, is predominately by trial and error. Thomas Edison's light bulb is good example, especially since it is the very symbol of idea. From what I know, Edison's contribution was his desire to make the light bulb, previously invented by other's, into a commercially successful product. His approach was to try this and try that until he finally succeeded. Benjamin Franklin invented the rocking chair. Why had no one invented it before? Surely ancient Chinese, Egyptian, and Sumerian civilizations would have loved this bit of easy low-tech entertainment. Perhaps we think a little too highly of our intellectual ability. Native Americans did not discover the three-finger (index, middle, ring) method of archery, even though they spent dozens of generations developing their archery skills. The more natural thumb and index finger method reduces the effective range by a factor of three. Lucky thing for the Pilgrims I guess. Random evolution resulted in our fantastic technology-using brains. No planned design using calculus or any other type of logic seems to have been needed. Nervous systems developed for one purpose randomly morphed to perform others. Some of the more complex organisms had evolutionary advantages that allowed them to propagate. But evolution largely failed to take advantage of basic technologies like fire, wheels, and metallurgy. It is ironic that we have succeeded doing a lot of technology the evolutionary computer failed to develop, but we are struggling to duplicate much of the technology it did. Thinking in humans, much like genetic evolution, seems to involve predominately trial and error. Even the logic we like to use is more often than not faulty, but can lead us to try something different. And example of popular logic that is invariably faulty is reasoning by analogy. It is attractive, but always breaks down on close examination. But this type of reasoning will lead to a trial that may succeed, possibly because of the attractive similarities, but more likely in spite of them. When the Wright brothers made the first airplane, they used a lot of different technologies. There was no single silver bullet, except for a determination to accomplish their goal. Like any technological advancement, the road to AGI will be paved with a variety techniques, technologies, trials, and errors. This seems doubly true since thinking as we know it is apparently a hodgepodge of methods. Catch you all later . . . Kevin C. --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
