#7 - Understanding is learning a new language to the point of fluency. (When the words in the new language activate your language independent concepts, and you have created sufficient behaviors so that you can effortlessly generate expressions in the new language). ~PM
Subject: Re: [agi] What is "understanding"? From: [email protected] Date: Wed, 3 Apr 2013 20:53:44 -0500 To: [email protected] Thanks Aaron! Computer people always want data structures and representation because it would make the problem straightforward, but there is no reason that there have to be any. People don't talk about neural networks having "representations". You guys ever heard of Rodney Brooks. It seems like at the linguistic level, we have representations, but you look closely, no, we don't have retrieval, we have dynamic recreations. As for arguing why propositions are not sufficient, how about all of non-explicit learning, like motor learning for example. Now, it's tricky because representational systems can be made informationally complete, so you might get the idea that you could get such a system to eventually work. But we have many years of constantly working at it. Brittleness is one of the big ones, but there are plenty more. One particular problem I am focusing on is the difficulty in incremental learning for them. You have to attach a completely different mechanism (like weights) onto the proposition to deal with that, so as a simple counterexample, you might see that propositional representations are insufficient. I apologize for not having a proper positive proposal, as that is one of the stated reasons for posting to this forum, but that's such a harder problem. This is still room for criticism of existing approaches. So let me throw in that some of the dynamic network architectures do seem to be promising. I would want the, to address the question of epistemic emotional judgements more directly, but I would say they are capable of it. I haven't looked into them enough I guess. Ben's system (sorry I have already forgotten what he's calling it, and I think I even download bits of it ) and maybe Stan Franklin's Lida. Personally, As I have said, I'm not great when there is a representational level in there anywhere. I'd want to stay closer to pure sensori-motor data, with an openness to deep machine learning for features if you just have to have something like representations. Ok, there. I tried!Andi On Apr 3, 2013, at 12:02 PM, Aaron Hosford <[email protected]> wrote: PM said:One suggestion is that you compile language into a "database of facts" using a propositional representation. In addition, you convert all sensory input to the AGI into the same propositional representation. Then you do inferencing within and generate behaviors from the aforesaid representation. If by "propositional representation", a logical statement with a Boolean true/false value, this will not be sufficient. The reason is, "facts" are never certain, and you never know in advance which ones will later prove wrong. Facts have associated confidence levels, based on supporting and conflicting evidence. Boolean truth values are an idealization of this, throwing away the ongoing accumulation of evidence and giving us only whether a particular proposition is currently accepted as reliable or not. The failure to recognize this has held back many seemingly promising AI projects in the past. Rich said:So, what the heck can we compile NL into that would support prospective AGI operation? This is what I've been describing to you. Semantic networks, properly structured, are up to the task. Any proposition from PM's "propositional representation" can be represented in a semantic network. The advantage that a semantic network then conveys is that the relationships between elements contained within a proposition can themselves be given confidence levels; the analysis of evidential support is no longer limited only to the proposition as a whole. For example, suppose I am looking at the proposition, ate(Billy, Nicky's_Popsicle). In a standard propositional representation like this, I can't analyze where the proposition is wrong, I just have to accept it is either right or wrong as a whole. If I use a semantic network-style representation, Billy<--SUBJECT--ate--OBJECT-->Nicky's_Popsicle, I now have two separate locations where I can attribute the failure of the proposition as a whole to be true: the SUBJECT and OBJECT links. Propositions come so close to doing this, but fail when we attempt to attribute failure to a particular substructure, because they aren't generalized enough to permit full analysis of the relationships of substructures to the parent structure. Andi said: I would go with Todor on this one. More specifically, it's very clear to me that language cannot be the bottom or basis of representation. A language system has to be a piece on top of the basic system. It may be the most important piece to us, because for interaction with us, and ability to use our body of written knowledge and contribute to it, a system will need to use language. But, that need in no way implies that you could ever get any intelligent behavior if you just start at the level of language. There are plenty of reasons to think otherwise. The problem Steve and I both agree needs to be solved is: What, inside the mind, represents the meanings of natural language, and how do we go about designing an analogous structure programmatically? When you say someone understands a sentence, what happens in that person's mind? Is there not some sort of internal structure to which that sentence gets mapped through the act of understanding? In most AI/AGI projects to date, there have been three basic approaches: (1) use it directly in text form, (2) pull out what you need and stash it in "frames", (3) convert it to a parse tree. I think each of these is inadequate to the task. I think there is a more comprehensive data structure used in the human mind to represent what a sentence actually means, and this is the data structure, the lingua franca of the mind, on which the mind operates in the act of thinking. What would that data structure look like, were we to reverse engineer it to work on a computer? Language is useful towards accomplishing this task, not because it is already in the proper form, but because its structure necessarily closely mirrors that form, due to its purpose of communicating knowledge in that form from one mind to another. Once we have a proper understanding of how meaning is represented in the mind, it should be possible to begin mapping sensory information to that format, just as can be done with natural language. On Wed, Apr 3, 2013 at 10:48 AM, Piaget Modeler <[email protected]> wrote: Steve Richfield: "So, what the heck can we compile NL into that would support prospective AGI operation?" One suggestion is that you compile language into a "database of facts" using a propositional representation. In addition, you convert all sensory input to the AGI into the same propositional representation. Then you do inferencing within and generate behaviors from the aforesaid representation. ~PM AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription ------------------------------------------- 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
