I just realized that I said something that is counterintuitive. Specialization is necessary for creating a -general- AI program. Of course many people have considered something like that. Perhaps if there were thousands of AI algorithms (which were specializations) then the program would have a great deal of general range. However, my thought on this is a little different. Perhaps if the AGI program kept track of thousands and thousands of specialized cases (linguistic specialization is only one example) then perhaps the program could achieve greater general abilities without having to rely on narrow AI algorithms (that were specialized for data types and so on.) So this is somewhat counterintuitive. Generalization is dependent on specialization. When I say it like that it makes perfect sense. But to make the point a little stronger: Learned Generalization is Dependent on Learned Specialization. Jim Bromer
On Sat, Mar 30, 2013 at 9:12 PM, Jim Bromer <[email protected]> wrote: > I am not sure that anyone is actually interested in my opinion on this but > to repeat my views on this argument again. > I am not actually interested in designing a human clone. I am interested > in writing an AGI program with an actual electronic computer. While having > real world events from sensors which were similar to human senses to tie a > referent to would be nice once in a while I just do not think that approach > would simplify the problem of designing a more advanced AGI program at this > time. So far there has been no outstanding evidence that multiple sensory > modalities would actually solve the contemporary problems of AGI. As I > have said, I believe that the main problem is one of complexity and while > the use of multiple sensory types might make the problem simpler in some > cases, I do not think it would not make it simpler in general. There is > however, outstanding evidence that text-only AGI is possible, and that is > the success of Watson. Watson may not have been a true AGI program, but it > represented a major milestone in AI. Few people foresaw that an > encyclopedic AI program would be achieved before an AI program exhibited > more typical human like reasoning but in retrospect most insightful > futurists did foresee a computer program that would have encyclopedic > knowledge. So Watson may not fully explain human reasoning but it does > provide a part in the stream of evidence that text-only AGI is possible. > > Reasoning is not something that relies only on real world facts. The idea > that visual input would somehow give an AI program reliable stream of real > world facts is surprisingly naïve. If it did, then visual based AI would > have already worked and the problem already solved. The problems of > understanding how thinking takes place has not been sidestepped by > declaring that visual-based AI and robotic AI is necessary for advanced > AGI. I realize that progress has been made in visual AI and robotic AI but > those problems are as difficult as text based reasoning has been. The > problem as I see it is that insight must be based on a complicated process > that is a little beyond our skills at this time. Is there some kind of > short cut that we might use. Ok, maybe. By using trial and error methods, > perhaps some future AGI program will be able to achieve genuine general > learning. But at this time there is no, or very little, evidence that some > simplistic system - like a traditional logical system - is sufficient. So > instead, I am wondering if perhaps a great deal of specialization is what > is needed to provide the basis for simplifying the complexity of AGI. > Perhaps by keeping track of all the specializations that the program can > conjecture it might be able to advance without getting bogged down in > complexity. > 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
