No one has yet created an AGI program (a holistic program that is capable of true learning) so the argument that some theory is holistic therefore it is adequate is just not necessarily the case. Perhaps with some other advancements in computer science these methods might turn out to be adequate. I might be wrong. My point of view is that methods which are hyper abstract tend to be methods in which some overly broad method is used in the dream that it would be feasible, given enough computing power and a few details that need to be worked out. But any AGI paradigm (one that is holistic and capable of some genuine learning) could be said to be hyper-abstract in the way that I meant it. So then I have to find a better way to construct my criticism of the supposed adequacy of some hyper-abstract method. The only way I could do that is by explaining that a hyper-abstract method (like relying on weighted reasoning) that is too limited will produce limited results. What we have seen so far is that even if someone has an AGI program that works really well for some species of problems it always turns that it is inadequate to demonstrate recognizable continued general learning. (The AGI programs that work are never incrementally scalable. They can work on sub-classes of problems but they are not capable of the conceptual integration that is so obvious in human beings.) So I do agree that we use theories which are approximately correct. But I disagree, for example, with a theory that details all learning as a reduction of the range of error of some numerical approximation that represents some knowledge. Does your theory see all learning as some kind of numerical evaluation problem? If you think that then I would say that even though your theory might be demonstrably general (it can be applied to a variety of problems), it is not adequately general (it does not demonstrate continued learning beyond those sub-species of problems that it works well on). It may turn out that with enough computing power an great many AGI paradigms could be finished to demonstrate true general learning that can actually get off the ground. But even if that is the case it would still probably turn out that there would be something more that is needed to make the demo. I am interested in discovering what those missing parts are. Jim Bromer Date: Sun, 30 Jun 2013 14:08:59 -0500 Subject: Re: [agi] Probably Approximately Correct: Nature's Algorithms for Learning and Prospering in a Complex World [By Leslie Valiant] From: [email protected] To: [email protected]
There are 2 approaches for programming AI: 1. Reductionist AI, in which the programmer hardwires a reductionist solution to a specific kind of problems. This approach is brittle when you change the kind of problems to solve. This is what Narrow AI is all about. 2. Holistic AI, in which the meta-programmer meta-programs a learning network capable of adapting its topology to fit the causal hyper-geometries of all kinds of problems. In other words, the Holistic AI system does the reduction process the programmer is supposed to do in the Narrow AI approach. This is what General AI is all about. Holistic AI is the approach of Monica Anderson and me. The book "Probably Approximately Correct: Nature's Algorithms for Learning and Prospering in a Complex World" is about Holistic AI. On Sun, Jun 30, 2013 at 8:34 AM, Jim Bromer <[email protected]> wrote: Date: Sat, 29 Jun 2013 23:17:15 -0500 Subject: [agi] Probably Approximately Correct: Nature's Algorithms for Learning and Prospering in a Complex World [By Leslie Valiant] From: [email protected] To: [email protected] Probably Approximately Correct: Nature's Algorithms for Learning and Prospering in a Complex World [By Leslie Valiant] http://www.amazon.com/dp/B00BE650IQ/ref=cm_sw_r_tw_ask_bQunF.0CDBG0V ------------------------------------------------------- I am just guessing about what the book is about based on the blurb, but the idea that we can muddle through without needing to understand what is going on is either poorly stated or nonsense. Although our theories are usually pretty weak, they are none the less theories. I do not believe that we are just basing our interest according to a coincidental correlation between three objects which can then be used to create chains and fences of correlations. I believe that the imagination is extremely important both in discovering objects of interest and in generating theories to explain the mechanisms behind the objects. That does not mean that we never rely on the linkages of ternary correlations it is just that a computational explanation of consciousness which goes that since a computer is not "conscious" of what it is doing then the potential for higher computational intelligence must prove that human beings are not "conscious" of what they are doing, just does not work for me. We are conscious of some of what we do even if these theories are not very good ones. One thing that I have been talking about for a number of years now is the importance of structural integration of concepts. Even if our theories and knowledge about a subject of interest are not that great we can begin to develop different ways to think about the subject and then use these different vantages to begin building better responsive insight about the subject. I think this can be done in AGI programs. Weak theories do not (always) need to be disposed but their influence in deriving conclusions about a subject matter can be modified so that they are used when more appropriate for the conditions. - This is only one possible presentation about my theories of structural integration. This is one way that I have to think about the subject. Parts of this presentation should seem very familiar to people who have thought about the subject and I am sure that there are people who would seize on the part where I said, the "influence [of weak theories] in deriving conclusions about a subject matter are modified so that they can be used when more appropriate for the conditions," as referring to the exact same thing as they have thought about when they try to design AI methods capable of producing improvement over time or after training. However, this effort to interpret what someone else says only on the basis of whether or not -I- have thought about things like this before can produce extremely insipid conclusions. (I do it all the time so I am not claiming some kind of superiority.) One reason my thoughts about this subject are a little different than the typical machine learning paradigm of learning-based improvement is that I explicitly emphasize the use of theories during learning. I was not just talking about the theories of people (who programmed some learning mechanism) but about the theories that the AGI program might generate through artificial imagination. So, had someone misread my statement believing that his machine learning theory was already imbued with a system where, 'conclusions about a subject matter were modified so that they would be used when more appropriate for the conditions,' he might have missed the point of my message entirely. That is one of the most serious problems with egomaniac-driven theorization. If you read everything only in the terms of how it is right or wrong according to your own theories you may end up missing central points of some reasonable remarks. So even though computers may not be conscious like we are I believe that we have to use meta-awareness in our AGI programs in order to make them act more reasonably. The theories that they will generate may not be that great but by using conceptual structural integration I believe that it should be feasible for them to use imagination and reason to build better analysis and response methods. So as they learn, some of their weak theories will be strengthened by making them more conditional and by extending their range of implementation slightly. This is only one part of my structural integration theories. Jim Bromer 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
