I don't have any details on how it would actually operate because it is a fairly wild model. I would have to control it using a somewhat precise special language to direct it so I could test the basic ideas out without having it be a full fledged AGI program.
Let's say that the program was trying to interpret what a sentence meant. "What time is the train arriving?" Suppose that it had recognized the words but now was trying to make sense of them. (I am not going to write a program that has a vocabulary at the start by the way.) It would know that trains depart and arrive at train stations if those concepts were already associated with the concept of a train (through previous learning). If it knew that departures and arrivals were made according to a schedule which was based on time and station, then it should be able to interpret that the sentence was concerned with the arrival time of a train at some station. It might not be absolutely certain of this interpretation. But it would be able to make that interpretation if those kinds of relations had been associated with the concept of a train. Other possible interpretations, like an odd one that inferred that a train was a kind of time piece would not be confirmed by the knowledge that it had about trains. Suppose however, that it had knowledge of a clock that was shaped like a model train for example. Then there might be some confusion about what the sentence meant. However, even in this special case the program could learn that arrival times were a more common issue when talking about trains than the much rarer case of a clock that was made to look like a train. So even though the program might be exposed to a lot of odd cases, it could also have a way to designate more common conceptual relations in its conceptual network. But this idea goes beyond associating facts with a particular concept. Conceptual relations can also be used to shape how ideas work. In fact, even this simple case demonstrates one way this can occur. Jim Bromer On Sat, Oct 6, 2012 at 9:45 PM, Dimitry Volfson <[email protected]> wrote: > Jim, > > I'm trying to understand. Could you show how your conceptual network would > ~~ see how parts are being used and see how much sense that makes to the > central concept ~~. And what the result would be depending on how much > sense was made. A hypothetical example is what I'd like to see. > > Thanks, > Dimitry > > > On 10/6/2012 7:12 AM, Jim Bromer wrote: > > I am presenting a rough idea of a conceptual network as a potential > advancement from earlier ideas like semantic networks. Looking on Wikipedia > I found some examples of semantic networks. In a semantic network the nodes > are the "concepts" and the edges are "relations between concepts". A > semantic network was usually defined with a conveniently finite number of > definitions of the edges (as types of relations between concepts) and a lot > of nodes (which were the concepts). One difference then is that the > conceptual network that I envision will not be limited by the number of > relations between concepts. This initial presentation, however, is a little > misleading because, as can easily be deduced from an inspection of a > semantic network, it is obvious that the edges, which are called "relations > between the concepts," are concepts themselves. So in the conceptual > network, a relation could become a concept itself. And the conceptual > network that I am thinking of does not have a single systematic method of > being 'activated' in some way (although searches would be made through it). > Furthermore, the network does not have to be envisioned as a single > network, but since different kinds of concepts may be associated > arbitrarily the potential for interrelations would tend to be extensive. > > Since this network is not as simple as a semantic network, the > utilization of the parts of the conceptual network would probably be > defined as they are used. So the different parts would not all work just > the same way. (However, the underlying methodology of how the different > parts are used might be drawn from a standard system). Finally, since the > network is not used in one simple way, deduction (derived from conceptual > knowledge) would also rely on what I call structural relations. Different > concepts would have different structural relations when used with other > concepts. This way an expectation of structural relations concerning a > central concept can help to derive meaning from a sentence or an > observation. So if the central concepts of a sentence (for example) were > recognized then other parts of the sentence that were directly related to > the central concepts could be found by fitting them to some of the > potential structural relationships that had been previously defined for > those central concepts. > > Different people have different kinds of knowledge about things, so the > structural relations that I am talking about are not (usually) normative. > For instance, a causal relation is a structural relation, but different > people will believe different kinds of things so there would be no > pre-defined underlying normative system of causality for the AGI program. > However, the program would be interested in trying to understand what other > people are describing and if this model of structural relations could be > used as a successful basis for an AGI program then it would learn something > about how people structure their own conceptual relations. Many other kinds > of relations between concepts could be considered as structural; I > mentioned causality only because it is such a familiar concept. > > The structural concept thing that I am thinking about is distinctly > different than (what I call) the funneling AGI models. Conclusions are not > derived through a funneling of deductions or weight-based reasoning. Yes, I > would use deduction and weight-based reasoning and yes the reaching of a > conclusion would have a terminal point, but the structural concept method > means that you don't just try to smush a measurement of the validity of all > ideas that are related to some central concept into a common hopper even > when the conclusion would not be homogenous for that combination of things. > Instead the program would look to see how the parts are being used and > whether or not that makes sense for the kind of central concepts that are > being considered at that moment. (I am using the term "structural" to > denote the fact that interrelated concepts should not all be funneled > through one single circuit of reasoning). > > While many people have come to the conclusion that my ideas about > conceptual structure only represented a high-level form of GOFAI or that > they were the same as the desired high level products of machine learning, > my theory is that that the structural relations between (individuated and > instanced) concepts have to be seen as part of the basis of reasoning, not > just the resultant of it. So while the individuated structural relations > between concepts in a particular instance would (usually) be learned, the > underlying programming has to take their usage into account. I believe that > the use of conceptual structure concerning some central idea that is to be > considered has to be a part of the foundational process of artificial > intelligence. And this idea can be used as an explanation of how we can > derive meaning from combinations of ideas that are somewhat novel. > > This is not an easy model but I believe it could be developed and at > least tested with some simple cases. > > Jim Bromer > > > On Fri, Oct 5, 2012 at 2:16 PM, Piaget Modeler > <[email protected]>wrote: > Sure. >> >> >> ~PM >> ------------------------------ >> I am curious about something. Is anyone interested in discussing my >> ideas about conceptual structure? >> Jim Bromer >> > *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> > <https://www.listbox.com/member/archive/rss/303/10215994-5ed4e9d1> | > 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> > > > ____________________________________________________________ > *Woman is 53 But Looks 25* > Mom reveals 1 simple wrinkle trick that has angered doctors... > <http://thirdpartyoffers.juno.com/TGL3142/5070decacaac15ec94a3est01duc> > ConsumerLifestyleMag.com<http://thirdpartyoffers.juno.com/TGL3142/5070decacaac15ec94a3est01duc> ------------------------------------------- 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
