My idea about Conceptual Structure is really a return to basics. While it is related to semantics, it is not the same as the old application of semantic methods. Since I believe that concepts have to interact with other concepts and that they can be relativistic that means that a simplistic semantic representation, or one that is based on simple vectors, just won't work. My program will rely on learned relations between concepts and those relations will tend to be unchanging once they are shown to be useful, but over time they will need to be modified - at least slightly - as they encounter new applications including applications that might superficially look very similar to a previous application.
One problem is that some evidence for their utility or insight has to be established. (I don't mean quantified measures of utility although quantification might be used.) The method I want to try to use to do this is conceptual integration. As conjectures of concepts are developed for particular situations, the utility or insight of the of the conjecture might be based on how well the features of the conjecture can be integrated with other conceptual structures (or conceptual models). Truly novel conjectures may be stretched to fit into existing conceptual structures, so new conjectures won't be relied on until there has been many more attempts to examine and shape the new idea. Jim Bromer On Tue, Aug 6, 2013 at 10:49 PM, Jim Bromer <jimbro...@gmail.com> wrote: > My idea that an AGI program has to have an executive function or > process that it is very simple but it has to be capable of AGI seems > obvious enough. It has to be lightweight or simple because the more > complicated it gets the greater the potential it will have to create > logjams. It might turn out that a lot of the potential for logjams > may be due to programming errors, but just as every little detail can > add some greater complexity to the programming, so can each detail add > to the complexity of the AGI program as it runs. > > Secondly, the recognition that the integration of Conceptual Structure > is the key to making it work is also a potential key to making the AGI > part relatively simple. Conceptual Structure is not a blanket > abstraction that the programmer completely details with his program > but a more creative structure that the program must create. So, yes, > Conceptual Structure is an abstract system - or more accurately > Conceptual Structure will consist of multiple implementations of > abstract systems - but it will be systems that are generated by the > AGI program as it is running. This idea of the Conceptual Structure, > which is based on the fact that concepts play roles when integrated > with other concepts, has to be kept simple or else it will be too > complicated and too slow for the program to manage it. > > Finally the program has to use rational creativity and it has to use > some kind of trial and error method. But the interesting thing about > this theory is that now I that I have an initial conjecture about > Conceptual Structure I should be able to craft it with as much control > as I need. Presuming that at first I will need to find a way to input > many of the details of how concepts should be integrated means that my > first endeavors would not really be AI or AGI even if my current > theory works. But at some point I hope to be able to figure out a way > for the program to learn how to determine more of the steps to > intelligently integrate conceptual structures. > > One theory that was never established, even weakly, in experiment was > that once you figured out how to create an AI program that it should > eventually become more adept at learning new things. I believe that > the theory of Conceptual Structures would make that feasible - if the > theory is any good at all. And this is how you could test the program > to compare it against competing AGI programs. It could learn new > things and integrate it as long as you could teach it. > > Jim Bromer ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
