I guess there were some list processing projects which used both process of elimination and accumulation in word interpretation done in the 1970s, but it was done with the technology of the period and, I believe that for the most part they worked with collections of extensions of synonyms of words and some specialized relational meanings. With modern computers these collections could, for example, be more extensive, more varied and include different kinds of specializations. For example, I would not be restricting the collections to synonyms and metaphors but I also envision an extensive collections of reactions and other specialized relations. An AGI program needs to know how to react to situations as they are expressed in the IO data environment. The goal in sentence recognition is not to work with collections of synonyms and to only eliminate those possible synonyms which are not appropriate for the particular sentence, but to be able to look beyond and create reactions that are insightful. The work done 40 years ago was mostly aimed at choosing the 'meaning' of a particular word as it was used in a sentence, and while there was some effort to build on the relations between sentences I think it was extremely restricted.
On Sun, Nov 24, 2013 at 11:39 AM, Jim Bromer <[email protected]> wrote: > There is another case for avoiding a discussion of a plan. The actual > implementation of a plan might be so different from the imagined > implementation that in the end the strength of the project might have > little to do with the main features of the plan. That is also a good > reason to avoid a prolonged contemplation about a plan. The basic > nature of day to day work of writing a computer program is pretty > consistent, at least across a project that uses one programming > language. So they may be more important to the implementation than the > inspiration behind the plan. But still some kinds of problems do tend > to bubble up in long planned thought which can be seen in the terms of > programming problems. For example, I can relate the problem of > representing multiplicity of possibilities to my thoughts about > cross-categorization and cross-generalization. When the program is > trying to 'recognize' a kind of situation from the input, it needs to > work from less detail to greater detail. If the features of the > (data) 'objects' it has to work with are extensively cross-generalized > (if the associations via similar features are extensively > cross-related) then the recognition stage of the process might be able > to traverse those relations more quickly. However, if recognition is > determined one feature at a time then the program will encounter > search complexity over and over again. This has been one of the main > problems in AI and it can be seen in a wide range of AI and AGI > implementations. So instead of traversing from one final recognition > object to another via the similarity of features I think it would > probably be more efficient to refer to collections of objects that > share some sets of features. (I am using a text based method but a > reference to a collection does not have to be expressed only using > text since I am talking about some kind of internal processing during > recognition.) So here I am saying that ideas like > cross-categorization can help you get what I mean when I talk about > cross-generalization. And the idea of a cross-generalization matrix > of features can help you understand what I mean when I talk about a > traversal of possibilities via the similarities of features. But we > know from the experiences of other programmers that if the program is > traversing possible end-product recognition objects then the search > process can be so slow as to make the search impossible. In the past > people have tried ideas like ideological vectors and weighted > references and reasoning in an effort to make more subtle decisions > but this hasn't worked because there is still not a straightforward > step by step process that can work for all cases (or even most cases). > Lists of collections were tried in the early days but these were > typically simple step by step elimination methods. What I am saying > is that the only way around this is to discover some effective > holistic method (neural networks are too inefficient) or to work with > intermediate collections of possible objects without resorting to an > overly simplistic step-by-step process of elimination. I believe that > the efficiency of modern computers can help us develop novel > approaches to finding effective solutions that weren't possible in the > last 25 years of the twentieth century. So how could I avoid using > the elimination approach that did not seem to work in the nineteen > seventies? I can't avoid an elimination approach entirely because I > need to end up with a final recognition object. But we can take > something that Wittgenstein realized to see why the step by step > elimination process might not work when dealing with generalization > collections. The generalizing principle in language is the use of > categorical families where each object in a category shares some > familial trait but that does not mean that two objects from the > category necessarily have some traits in common. In language we use > general terms to describe a specific referent. But since these terms > may be based on familial traits we cannot eliminate the possible > traits that may apply to the specified object simply because it is not > common to two terms used in the description. I'm sorry but this is > basic stuff. If you have thought about this then it should be > obvious. (I may not have expressed it in the simplest terms but if > you have thought about it before you should be able to figure it out. > If you haven't thought about it before then you should start now.) > And the end product of a recognition does not have to be a specific > object because the terms we use in language and in thought are > generalizations. So I don't have it all figured out but to make this > simple: you cannot use a simplistic step-by-step elimination method to > narrow the possibilities down. But there is a way around this. -- 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
