I am interested in the problem of getting an AGI program to learn how to read using text-based IO.
The question I have is how can an AGI program discover that certain kinds of statements refer to the linguistic markers of relations (in the text) as opposed to a reference to some other conceptual matter. One way is by relying solely on the mechanism designed to crudely learn to define a concept. While trying to find previously learned knowledge related to some central concepts under consideration it might implicitly recognize that some statements can refer to the search process (and other processes that) it itself is conducting. So the processes it uses to find and index references relative to some concept might then be used implicitly to gain a sense that some statements (for example those that a user inputs) might be directed at what it is doing as opposed to some (other) subject matter. But another method might be derived from a stronger mechanism that is capable of producing awareness of what it is doing. I am speaking of awareness in the more ordinary sense of keeping track of what it is doing. As the program interacted with its IO data environment it would record the different actions it takes and because it needs to discover the more important events that occur it will be abstracting and generalizing the recording of its own actions just as it abstracts and generalizes the other concepts it will be learning about. (This also means that it will try to categorize how it responds to different kinds of situations as well.) To do all this it will utilize two important methods. One is that it will be treating the actions that it takes with the same kinds of mechanisms that it will use on other subjects that it is considering. Secondly, it will be explicitly examining its own behavior. So this stronger system will allow it to develop an awareness of what it is doing in parallel to its examination of the conceptual subjects that it makes. However, because this system will always be running it means that it is always tracking its own behavior as it learns about any subject. And this situation can lead to an over identification with the subject matter. It might be hard, for example, to distinguish its own experiences of learning from the subject matter that it is learning about. While human beings do this, they are able to detach their personal experiences in order to focus on the subject matter. But this ability can be explained by making choices about how it approaches some problem. So even though I may have memories of math teachers teaching me elementary mathematical methods, the method of focusing on a problem is a strong enough process to allow me to consider a problem without thinking of that. In fact, the active process of abstraction and generalization also will tend to produce focused concept-based methods which are distinguished from other concepts and various other co-occurring events. This explicit tracking and abstracting of the actions the program takes makes a lot of sense as a method to implement short term memory. An AGI program needs to be able to see if it recognizes something that it had encountered before and a short term memory is part of that step. Although there are extreme limits on how broad this kind of mechanism can be, the fact that it can also be used as a foundation for an awareness of what it is doing supports the notion that something like this should be used. So I think that a mechanism like the one I described will be useful for implementing some kind of self-awareness that can then be used to direct the AGI program to understand that certain kinds of relations in the text can be used to discover linguistic generalizations. Jim Bromer ------------------------------------------- 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
