I will continue this because I made a couple of errors. The problem with probability networks is that the standard models that are available are not devised to implement frames of reasoning (including reasoning such as Bayesian Reasoning.) So that is left to the imagination of the individual researcher. But, as I have argued (with little response), because critical discernment is not a part of the AI use of statistical methods (as might be contrasted with the vibrant critical scientific application of statistical methods in other fields) this results in a mushing of different kinds of measurements. Now when someone naively goes about applying these methods he is going to find that they only work for certain kinds of applications and that they even have to be adapted for the variations of the kinds of situations that they do work for.
If someone had proven that his methods were truly general then he would be right to talk about them confidently even though they might still be somewhat limited. Nothing is perfect. However, Ben's methods have not been proven to be truly general as can be seen by the lack of a viable demonstration, the talk of the need for "tuning," and declaration of the goodness of "meta-learning." So the question is whether or not the learning methods that Linas and Ben are talking about are truly general, or are they are only narrow methods, or are they like standard Bayesian networks which can be adapted to a lot of situations but are extremely inefficient or whether they are truly general in that they can be used in a lot of different kinds of situations and they are also effective in a lot of different kinds of situations. If they need to be heavily "tuned" for millions of different kinds of situations in order to be effective then they are not general. And if they just do not work for the kinds of purposes that we envision them then they are not feasibly general. Meta-learning is both a super class of learning and a subclass of learning. It is a special kind of learning so it is a subclass of it, but it is also imagined as a kind of management system for learning so it is a super class of it. Although there are implementation details that make this kind of analysis somewhat simplistic, the point is that if a meta-learning system is necessary for the viability and generalization of the learning system, then the learning system is not general and viable to begin with. (Nothing would be perfect, human beings have limitations too. But the issue is whether the learning systems are intelligent enough to begin to show human like skills.) If something is both a super class and a subclass of GL then it is GL. There comes a point when a researcher has to be both mature enough and honest enough to accurately describe his experimentation. To continue to talk about "tuning" and "meta-learning" which "I worked on last year," is disingenuous if there is no evidence that these things actually work. Yes we all use different forms of language to express ourselves but some qualifications are absolutely necessary for clear communication. No learning system is going to be perfect, but the claim that it is has been figured out can only be established by demonstration. Bayesian Networks, for example, are general but, as they stand now they are not effective enough to be comparable to child-like learning because they are inefficient. I believe that the inefficiency can be traced directly to the mush factor (that I have talked about with no apparent traction. Maybe it is one of those things that I sometimes talk about which is so obvious that the geniuses who constitute 98% of this group are saying "well duhhh" to themselves even as they read this.) Jim Bromer On Sun, Sep 30, 2012 at 6:56 PM, Jim Bromer <[email protected]> wrote: > On Tue, Sep 25, 2012 at 11:02 PM, Ben Goertzel <[email protected]> wrote: > What Linas is saying, I think, is that to get a complex multi-component > learning system like OpenCog to work, there are many complex "tuning > /adapting" tasks to be done.... > So he's saying: If we had a "meta-learning" method... > as Linas knows, Nil and I worked on a DARPA-funded project on > meta-learning last year. > Ultimately, I think we can build an OpenCog thinking machine without this > kind of metalearning he alludes to. BUT ... absolutely, the further we can > go in that direction the better... > > > Papers and books about meta-cognition are important for students who have > not thought about things like that. If a computer program was able to > think then it would be able to think about thinking. So while I support > the frame work of a higher management system and its ilk I have to conclude > that the argument that the feasibility of intelligence is dependent on > cognition and the feasibility of cognition is dependent on meta cognition > really doesn't cut it. If a program is not able to 'think' then it will > not be able to 'think about thinking' and the delusion that by creating > some kind of meta-managment system you will finally be able to generate > intelligence is not even slightly reasonable. > > If Linas was wondering about a simple automation system that would be a > different kind of thing. But if he is discovering that the learning > mechanisms that have been detailed are infeasible without an intelligent > entity to shape them for the millions of different kinds of situations that > they have to be adapted for than the talk about 'ftuning' and a > 'meta-learning' system is childish. > > Jim Bromer > > > > > On Tue, Sep 25, 2012 at 11:02 PM, Ben Goertzel <[email protected]> wrote: > >> What Linas is saying, I think, is that to get a complex multi-component >> learning system like OpenCog to work, there are many complex "tuning >> /adapting" tasks to be done.... Each time you want to apply a certain >> learning algorithm to a certain sort of problem (such as, the problem of >> helping another learning algorithm do its stuff better), you have to tweak >> that learning algorithm in certain ways. He's afraid that, in a system >> based on the interactions of a lot of different learning algorithms for a >> lot of different purposes, this leads to a very large number of >> moderately-complex algorithm tuning/adaptation problems. >> >> So he's saying: If we had a "meta-learning" method for automatically >> tuning/adapting a learning algorithm to a new domain and a new sort of >> problem, *then* an architecture like OpenCog could be made to work a lot >> more tractably. Otherwise, getting it to work is gonna be a long slog of >> tuning/tweaking each component for its interaction with each other >> component. >> >> At least, that's my quick gloss of his point. >> >> I'm unsure of my reaction. Indeed, metacognition is important -- as >> Linas knows, Nil and I worked on a DARPA-funded project on meta-learning >> last year. (We have an in-progress paper on "meta-learning for feature >> selection" in a MOSES context.) On the other hand, the human brain was >> honed incrementally by evolution over a long period of time, and evolution >> is exactly the sort of algorithm that is able to repetitively carry out the >> long slog of tuning/tweaking each component of a complex system for its >> interaction with each other component... >> >> Ultimately, I think we can build an OpenCog thinking machine without this >> kind of metalearning he alludes to. BUT ... absolutely, the further we can >> go in that direction the better... >> >> ben g >> >> >> On Tue, Sep 25, 2012 at 9:52 PM, Piaget Modeler < >> [email protected]> wrote: >> >>> >>> Kindly advise, which component is "The Magic Happens Here"? I didn't >>> recognize it on Ben's diagram or mine. >>> >>> >>> https://www.facebook.com/photo.php?fbid=460040220684646&set=a.217364258285578.55073.203359906352680&type=1&theater >>> >>> >>> Please advise. >>> >>> ~PM. >>> >>> >>> ------------------------------ >>> From: [email protected] >>> Date: Tue, 25 Sep 2012 20:44:11 -0500 >>> Subject: [agi] Magic Happens Here [was Re: [opencog-dev] Uber big scary >>> monster OpenCog diagram >>> To: [email protected] >>> CC: [email protected]; [email protected] >>> >>> >>> >>> >>> On 12 September 2012 15:36, Ben Goertzel <[email protected]> wrote: >>> >>> >>> http://goertzel.org/WholeBigOpenCogDiagram.pdf >>> >>> >>> So I got to thinking about the "Magic Happens Here" part of the diagram, >>> and I think maybe it got left out of the diagram. >>> >>> So, for example, I was recently thinking that I could use MOSES to >>> automatically learn new link-grammar parse relations. Now, it doesn't >>> explicitly have to be MOSES that would do the learning; I suppose that many >>> modern-day, reasonably competent learning systems might do the trick: they >>> would need only to be able to do some basic modelling of the innards of a >>> black box. And it doesn't have to be link-grammar either: some other >>> reasonably flexible NLP system would do. >>> >>> As it became clear as to how to hook these two up, so as to learn new >>> rules, I also thought about what it couldn't do ... and realized that I'd >>> need some other variants and modifications to handle those cases. So, my >>> plan went from hooking up the two things, to realizing that I would need to >>> hook them up in three distinct ways. Each different way handles a certain >>> generic kind of learning problem, but with a different focus and a >>> different output/outcome. And I started grasping that in fact, maybe >>> there are 4 or 5 or 6 different other situations to deal with. >>> >>> And at that point, it gets out of hand -- all of a sudden, I need to >>> manually handle a bunch of different learning tasks. Each task may take >>> months to code up and make operational. And what if there are more than 6? >>> That's when I realized that, in fact, I have a meta-learning problem. >>> That is, what I really need is a system that can learn how to learn >>> specific learning tasks. *That* is the "Magic happens Here" bubble. >>> >>> At the moment, I have no clue how to do this meta-learning. But I do >>> know where to start: experimental trial and error. Go ahead, hook up moses >>> to link-grammar. See what happens. Plug the holes, See if a general pattern >>> or paradigm emerges. If, after much work, some meta-pattern becomes >>> apparent, then, well, that was the magic part. After that .. who knows. >>> But one cannot find out, until one starts trying to build these things. >>> >>> --linas >>> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >>> <https://www.listbox.com/member/archive/rss/303/19999924-5cfde295> | >>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>> <http://www.listbox.com> >>> >> >> >> >> -- >> Ben Goertzel, PhD >> http://goertzel.org >> >> "My humanity is a constant self-overcoming" -- Friedrich Nietzsche >> >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/10561250-164650b2> | >> Modify<https://www.listbox.com/member/?&;>Your Subscription >> <http://www.listbox.com> >> > > ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
