I thought that your comments about using logic were interesting. To the best of my understanding and memory you started by saying that being an 'element-of' is not the same as being a 'subset-of'. And (I think you were saying) 'equals' is different than 'equivalence'. And it is very difficult to express how they vary using logic even though they are obviously related. So these basic ideas have some characteristics which were simultaneously very different and strongly similar. And I think the fact that these relations might be essential to an application of logic implies that AI-logic might be too difficult to work with.
I think that discrete logic-like AI looked like it held a lot of promise because it seemed like you could start out with some simple ideas and then build on them. The initial ideas did not have to be extremely elementary because you could attach limited meanings to ideas that you used and still build on it as you went. Your experience, however, suggests that this kind of reasoning is illusory because applied logic quickly becomes too complicated. Someone might try using statistics or discrete fuzzy reasoning but these strategies have also turned out to be too weak. Some ideas (or idea-like artificial mentations ) just need to be strong and precisely bound. Although we can claim that a logic-like system could be strong enough representationally, what you found was that the effort was way too complicated. So you want to use neural networks and machine learning. That makes sense but I think there is another way to deal with this problem. Although the essence of the similarity and differences of 'being-an-element-of' and 'being-a-subset-of' may be nearly indescribable for you or me or 99% of the rest of humanity, isn't it really similar to the problems of ambiguity or finding attachments between words in a sentence (like anaphors) or to other words or ideas? (Anaphoric-like connections are too subtle for hard edged automated Boolean Logic and it takes a lot of work for us to analyze them even when we start to identify them.) I use logic in my thinking almost all of the time. But it is not a single totally integrated system of logic. On the other hand, it is not just thousands of totally independent logical statements either. So if I have thought about something carefully (or learned about something through a lot of experience) I can somehow define logical relationships that are well integrated to other issues relevant to the thought. And although this logical thinking is not totally integrated, I can usually find numerous overlaps across subject or sub-subject domains. I think the problem might be (at least partially) resolved through usage patterns. Most people are not writers and most writers are not technical writers so it is difficult to perfectly describe the differences and similarities of provocative abstract features of intelligence. And therefore it is also difficult to program them in as fundamental operators. I may have misunderstood you but regardless, the features 'is an element of' and 'is a subset of' are labels, categorical operations and relations. And they are other things and processes as well. It is easier to say something like that then it is to describe them logically or to make sure that their active definitions work perfectly every time. So rather than trying to divine them as underlying principles I would start out by trying to associate their usage with actual cases and then see if I could get the program to develop abstractions about what is common and what is different in the usages. Then using that information the program could try to make intelligent guesses about other similar cases that it would have to deal with in the future. So understanding that the line between a label and an operator is not always absolute and that some operational rules, including fundamental rules, have to be learned through usage I think that someone might be able to figure a way develop a mostly discrete AI program that can start off with extreme simplifications of sophisticated ideas and build on that. And one other thing. Usage-based learning (that can also look for abstractions of similarities and differences) will tend to build distributed systems of (mostly) discrete knowledge. Distributed systems of strongly related discrete knowledge can be too complicated to manage but if the management of something like that proves to be feasible then the distribution of related knowledge should tend to strengthen the knowledge base of the AI program. I don't know if I can effectively use these ideas in an actual AI program but I think that I have some well-founded ideas to start with. Jim Bromer On Thu, Feb 25, 2016 at 1:23 PM, YKY (Yan King Yin, 甄景贤) < [email protected]> wrote: > This 8-minute video (now with Chinese and English subtitles) explains my > latest AGI theory: > > https://www.youtube.com/watch?v=c9HWcYd36E8 > > The main idea is > > *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> > <https://www.listbox.com/member/archive/rss/303/24379807-653794b5> | > 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-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
