Ben, I am in absolute agreement with your very eloquent posting. Just one question: Doesn't this also apply equally well to your OpenCog approach to AGI? Indeed, THIS (as described in your posting) has been my primary objection to the past AGI-related efforts that I have seen.
Again, very well said. Steve =============== On Tue, Jun 19, 2012 at 6:31 PM, Ben Goertzel <[email protected]> wrote: > > There's a general fallacy that misleads many AGI people, of the following > form ... > > " > -- Capability or method X, if you could do it incredibly (i.e. > unrealistically) well, would enable arbitrarily great general intelligence > -- Simple versions of X, seem to lead to interesting "narrow AI" behaviors > THEREFORE... > -- By pursuing more and more complex versions of X, we can get high > levels (e.g. human-level) of real-world general intelligence > " > > In the case we're discussing here X = Prediction .. > > In other cases, X = logical reasoning, or pattern recognition, or > automated program learning, or simulation, etc. etc. > > Unfortunately, things just don't work that way ;/ ... > > ben > > > On Tue, Jun 19, 2012 at 9:21 PM, Steve Richfield < > [email protected]> wrote: > >> Jim, >> >> I think we are in agreement here. Computing optimal action without a >> guiding prediction is NOT easy. I mentioned high speed trading because they >> appear to be doing just that, albeit within a narrow domain. I suspect that >> failure to grok this area is just one of many areas where AGI is going to >> have to make progress before it can become "serious". >> >> Steve >> ============= >> On Tue, Jun 19, 2012 at 6:00 PM, Jim Bromer <[email protected]> wrote: >> >>> Steve, >>> High speed trading is something that people are not good at but narrow >>> AI can be. It has to be narrow to keep it efficient. >>> Your idea of optimal action in the absence of prediction is a pretty >>> wild abstraction, and it would be difficult to implement. Just using >>> correlation for example would tell you a lot about the relations of what >>> was obvious and previously identifiable but little about the relations of >>> causation and (ironically) co-occurrence. Correlation can identify perfect >>> co-occurrence but it cannot be relied on -in itself- to identify imperfect >>> or conditional co-occurrence. >>> So for something like correlation to actually work to reliably identify >>> conditional co-occurrences it has to have some way to identify or speculate >>> on interactions that may be hidden. (I think Abram was telling me that >>> hidden Markov processes were capable of doing this but here the guiding >>> form of a Markov process is a conditional premise of the efficacy of the >>> method.) >>> >>> In order then to generalize this example of effectively using >>> correlation just to identify interactive co-occurrence I feel the system >>> would have to be capable of a dealing with a great many possible >>> complications just to identify the few interactive co-occurrences that >>> might have impact on a subsystem that is being observed. First, how do you >>> simultaneously watch multiple subsystems? This is the easy part because >>> our ideas about simple observations are inherently absurd. A simple system >>> can easily be a complicated system. This relativistic view is the first >>> clue then to discovering new ideas about simultaneously watching multiple >>> complex systems. Many of the systems that can be observed easily are >>> complex systems. So the real problem is how do we discriminate or recognize >>> subtle relations that are hidden in the seeming simplicity of an observed >>> event. For example if a cognitive system was deriving insight from a >>> camera, the video of an observation event would contain all the complexity >>> that could be inferred from what the camera could capture. So while a >>> cognitive system might make jump to some real time conclusions, reanalysis >>> of the recording of the event might provide more insight from a more >>> sophisticated cognitive basis. >>> >>> Although this sounds like nothing new, it is still new just because no >>> one has made much progress in identifying how representations of complexity >>> work. >>> >>> Jim Bromer >>> >>> >>> On Tue, Jun 19, 2012 at 5:08 PM, Steve Richfield < >>> [email protected]> wrote: >>> >>>> Just a comment to inject into this discussion: >>>> >>>> Prediction is great when it is possible, but that is rare in our world >>>> of imperfect information. >>>> >>>> Modern economics has brought us the concept of optimal action in the >>>> face if imperfect information. This leans on concepts like volatility, >>>> where tiny (and hence unpredictable) contributions can have huge effects. >>>> >>>> I think the emphasis should NOT be prediction, but rather on the >>>> computation of optimal action in the ABSENCE of prediction. Of course that >>>> is a more complex concept, so it has so far evaded deep discussion here. >>>> >>>> Note that this is the stock and trade of high speed trading software, >>>> so people ARE already making this work in the real world. >>>> >>>> Any thoughts? >>>> >>>> Steve Richfield >>>> ================== >>>> >>>> On Tue, Jun 19, 2012 at 1:50 PM, Jim Bromer <[email protected]>wrote: >>>> >>>>> Logan Streondj <[email protected]> wrote: >>>>> Infering is a form of carrying knowledge from one place and applying >>>>> it in a different place or time. >>>>> >>>>> >>>>> This is the real question but it does not provide us with an answer. >>>>> All the narrow forms of AI do offer solutions to certain kinds of >>>>> problems, >>>>> but is there a general way to work from uncertainty (about most every >>>>> basis to make the determination) toward greater certainty that would allow >>>>> us to say that a particular kind of knowledge that worked in another >>>>> situation could work in this situation? If you base inference on >>>>> similarities then the problem is how do you use automation (in other words >>>>> a program) to detect similarities without some absolute method that some >>>>> of >>>>> the aspects of the two similar events are of a kind? >>>>> >>>>> Jim Bromer >>>>> >>>>> >>>>> >>>>> >>>>> On Tue, Jun 19, 2012 at 3:42 PM, Jim Bromer <[email protected]>wrote: >>>>> >>>>>> Logan, >>>>>> I also cannot predict the kinds of replies that I will get. >>>>>> Of course, prediction is something that human beings can do. I >>>>>> mentioned that in my previous message. However, I do not see this group >>>>>> as >>>>>> swirling around the question of what kinds of things human beings can do >>>>>> but around the question of what can we do to make our computer programs >>>>>> act >>>>>> smarter. In that sense, intelligence is not a product of prediction, >>>>>> prediction is a product of intelligence. >>>>>> If anyone has proof that AGI is a product of prediction then he has >>>>>> solved the problems that are constantly being discussed in this group. >>>>>> >>>>>> I think there was another author who popularized probability and >>>>>> prediction in the 1970s. >>>>>> >>>>>> Your definition of inference is interesting and it is a more >>>>>> sophisticated way of understanding the problem then the usual refrain of >>>>>> the false promise of probability and prediction. >>>>>> Jim Bromer >>>>>> >>>>>> >>>>>> >>>>>> On Tue, Jun 19, 2012 at 3:33 PM, Logan Streondj >>>>>> <[email protected]>wrote: >>>>>> >>>>>>> I agree with Jim that there is too much focus on "prediction". >>>>>>> >>>>>>> Though here the obsession might be related to how Ray Kurzweil, a >>>>>>> man famous for his accurate predictions, chooses to define intelligence >>>>>>> as >>>>>>> the ability to predict. >>>>>>> >>>>>>> If prediction was a normal thing, then Ray wouldn't have gotten >>>>>>> famous for it. >>>>>>> Amongst us ordinary humans, at least around here in Canada, and on >>>>>>> the internet, people rarely if ever mention the word "prediction". And >>>>>>> very >>>>>>> few if anyone claims to make accurate predictions. >>>>>>> >>>>>>> Kurzweil mentions prediction as necessary for things like hunting or >>>>>>> gathering food. >>>>>>> For instance projecting the vector of an animal as is runs away, >>>>>>> or infering that if fruit was gathered in an area at a time of year, >>>>>>> it may be available at same place next year at similar time. >>>>>>> >>>>>>> However projecting, infering, and predicting are 3 different words, >>>>>>> with different meanings. >>>>>>> >>>>>>> Projecting is a form of planning ahead. >>>>>>> Infering is a form of carrying knowledge from one place and >>>>>>> applying it in a different place or time. >>>>>>> Predicting is a form of prophecy or foretelling based on special >>>>>>> knowledge. >>>>>>> >>>>>>> Most people infer and project, but very few predict. >>>>>>> >>>>>>> For instance in New Age cultures, many humans have had their careers >>>>>>> broken by making false predictions. Also no one (but Ray) claims to >>>>>>> make >>>>>>> predictions based on intelligence, but usually based on things like >>>>>>> intuition, and telepathic communications. >>>>>>> >>>>>>> Maybe I'm committing the fallacy of making a distinction without a >>>>>>> difference. Perhaps all those words mean the same thing. >>>>>>> >>>>>>> Though I'd like to state that there are other ways of getting food, >>>>>>> that don't relate to predictive ability. >>>>>>> For instance how does an animal get food, they find a signature of >>>>>>> the food-item, for instance a smell, sound or shape, and then move >>>>>>> towards >>>>>>> it, until their jaws are clenching it. >>>>>>> I guess you could say they "predicted" the food was there, based on >>>>>>> the signature. >>>>>>> As it certainly is possible that following a signature, will lead to >>>>>>> a non-food item, >>>>>>> for instance if the signature is lost, or is being produced by >>>>>>> something else, i.e. carnivorous flower. >>>>>>> >>>>>>> >>>>>>> >>>>>>> It is also possible that everything can be classified as prediction, >>>>>>> for instance when saying the alphabet, we predict what the next >>>>>>> letter is, before saying it. >>>>>>> This would mean that every hello world program uses prediction, as >>>>>>> it loads the hello world string, before printing it to screen. >>>>>>> *shrugs* >>>>>>> >>>>>>> it depends, do we want prediction to lose all meaning, by applying >>>>>>> it just about anything. >>>>>>> Or do we want to be specific with what we are saying, and only use >>>>>>> words like "prediction" in the way that non AGI mailing list people use >>>>>>> it, as a shorter-term version of prophesy, based on special knowledge. >>>>>>> >>>>>>> For instance even though I consider myself intelligence, I can't >>>>>>> predict the content of any reply or if there even will be replies, >>>>>>> though >>>>>>> I can infer that there may be replies, as this is a mailing list, where >>>>>>> people often reply. >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Tue, Jun 19, 2012 at 1:48 PM, Piaget Modeler < >>>>>>> [email protected]> wrote: >>>>>>> >>>>>>>> Jim, >>>>>>>> >>>>>>>> Is this your actual belief or is this disinformation? >>>>>>>> >>>>>>>> ------------------------------ >>>>>>>> Date: Tue, 19 Jun 2012 10:40:22 -0400 >>>>>>>> Subject: [agi] Prediction Did Not Work (except in narrow ai.) >>>>>>>> From: [email protected] >>>>>>>> To: [email protected] >>>>>>>> >>>>>>>> >>>>>>>> The original idea behind the use of "prediction" in AI was that the >>>>>>>> prediction could be compared against the actuality and that comparison >>>>>>>> could be used to test the theory that produced the prediction. (Some >>>>>>>> author popularized that model for AI but it was proposed by academic >>>>>>>> researchers before he did so. Karl Popper used the concept as part of >>>>>>>> his >>>>>>>> model of scientific discovery in the 1930s, but his principles, which >>>>>>>> were >>>>>>>> based on logical positivism, have become more dubious because logical >>>>>>>> certainty has become a more dubious principle of knowledge. And, oh, >>>>>>>> by the >>>>>>>> way, Popper did not believe that AI was possible.) So, continuing >>>>>>>> with the >>>>>>>> march of the use of "prediction" in AI, AI people could see that our >>>>>>>> expectations were like "predictions" and so it did seem that the human >>>>>>>> mind did indeed use a method of prediction. Of course the principle >>>>>>>> that a >>>>>>>> prediction could be compared against an actuality in order to evaluate >>>>>>>> the >>>>>>>> accuracy of a theory only works in narrow AI, and as narrow AI failed >>>>>>>> to >>>>>>>> produce simple AGI that part of the cherished notion of "prediction" >>>>>>>> has >>>>>>>> been gradually eroded. >>>>>>>> >>>>>>>> This group uses the term prediction to simply refer to something >>>>>>>> that is "known" and as such it is a concept which is pretty shallow >>>>>>>> since >>>>>>>> its verification as a mental product is thereby based on the experience >>>>>>>> that when we know something we act as if we were confident that it >>>>>>>> would >>>>>>>> happen. The problem with such concepts like "knowing" or "prediction" >>>>>>>> is >>>>>>>> that they -do not- confirm the efficacy of theories that an AGI program >>>>>>>> might produce, except in those circumstances which would be considered >>>>>>>> narrow AI by this group. >>>>>>>> >>>>>>>> Let me repeat that. >>>>>>>> The problem with such concepts like "knowing" or "prediction" is >>>>>>>> that they -DO NOT- confirm the effaciacy of theories that an AGI >>>>>>>> program >>>>>>>> might produce, except in those circumstances which would be considered >>>>>>>> narrow AI by this group. >>>>>>>> >>>>>>>> So sure, when someone points out that the human mind uses >>>>>>>> "expectation" and expectation is a little like "prediction" I do agree. >>>>>>>> But here the word prediction is just being used to describe "knowing >>>>>>>> something." There is no principle of confirmation or disconfirmation >>>>>>>> of >>>>>>>> the use of "prediction" that can be used to produce AGI, except for >>>>>>>> special >>>>>>>> cases. After years and years of the repetition of the word in these >>>>>>>> types >>>>>>>> of discussions there is still no AGI so that should give you a hint >>>>>>>> about >>>>>>>> how good an idea it was. >>>>>>>> >>>>>>>> If the use of prediction as a confirming method can only be used >>>>>>>> in a limited set of circumstances then its power in these discussions >>>>>>>> has >>>>>>>> been so diminished that it should not be used as if it were a magical >>>>>>>> concept. Without some efficacy the word should not be used as a >>>>>>>> special >>>>>>>> technical term. The word should be used in the way it is usually used. >>>>>>>> >>>>>>>> As I implied, Popper originally used the word the concept in a >>>>>>>> logical model of scientific theory. If a theory could be used to >>>>>>>> predict a >>>>>>>> confirming or disconfirming observable event then the theory could be >>>>>>>> disconfirmed by the failure of the event to occur. (If the event >>>>>>>> occurred >>>>>>>> it still might be caused by a coincidence.) >>>>>>>> >>>>>>>> It is coming back to me. (Or else my creative memory is kicking >>>>>>>> in.) The author who popularized the theory of confirmation through >>>>>>>> prediction had a model of probability and confirmation by prediction. >>>>>>>> That >>>>>>>> model is inherently contradictory. >>>>>>>> >>>>>>>> It amazes me that you guys don't get this. >>>>>>>> >>>>>>>> Jim Bromer >>>>>>>> *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> >>>>>>>> *AGI* | Archives<https://www.listbox.com/member/archive/303/=now> >>>>>>>> <https://www.listbox.com/member/archive/rss/303/5037279-6ef01b0b>| >>>>>>>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>>>>>>> <http://www.listbox.com> >>>>>>>> >>>>>>> >>>>>>> *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> >>>>> <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | >>>>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>>>> <http://www.listbox.com> >>>>> >>>> >>>> >>>> >>>> -- >>>> Full employment can be had with the stoke of a pen. Simply institute a >>>> six hour workday. That will easily create enough new jobs to bring back >>>> full employment. >>>> >>>> >>>> *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> >>> <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | >>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>> <http://www.listbox.com> >>> >> >> >> >> -- >> Full employment can be had with the stoke of a pen. Simply institute a >> six hour workday. That will easily create enough new jobs to bring back >> full employment. >> >> >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/212726-11ac2389> | >> 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/303/10443978-6f4c28ac> | > Modify<https://www.listbox.com/member/?&;>Your Subscription > <http://www.listbox.com> > -- Full employment can be had with the stoke of a pen. Simply institute a six hour workday. That will easily create enough new jobs to bring back full employment. ------------------------------------------- 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
