WHAT ARE THE MISSING CONCEPTUAL PIECES IN AGI? --- recent input and responses
BELOW ARE THE MOST RECENT DISCUSSIONS CONCERNING POSSIBLE MISSING CONCEPTUAL PROBLEMS THAT MIGHT STAND BETWEEN US AND AGI. THESE COMMENTS ALL RELATE TO IMPORTANT ISSUES TO BE DEALT WITH --- BUT IT IS NOT CLEAR ANY OF THEM REPRESENT A MAJOR CONCEPTUAL PROBLEM FOR WHICH WE HAVE NO REASONABLE POTENTIAL SOLUTION. ========================================================== ====Richard Loosemore Sat 4/19/2008 7:57 PM ========================================================== Richard Loosemore [EMAIL PROTECTED] Sun 4/20/2008 4:20 P Ed Porter wrote: > RICHARD, > > I can't provide "concrete reasons" why Novamente and roughly similar > approaches will work --- precisely because they are designed to > operate in the realm Wolfram calls computationally irreducible --- > meaning it cannot be modeled properly by anything substantially less > complex than itself. Thus, whether or not it will work cannot be > formally proven in advance. Just a few small details: As I understand it, Ben has argued vehemently that Novamente is not subject to the computational irreducibility (complex systems) issue. And complexity does mean that "it cannot be modeled properly by anything substantially less complex than itself". What it does mean is that it cannot be explained in an "analytic" manner. ====<ED PORTER= According to the def of computational irreducibility in wikipedia it is a concept that applies in vary degrees to differing levels and types of description of a computation in physical reality or in a computer. A large Novamente system would have so many complex things going on that it would probably take more effort to devise model that would accurately predict its behavior under the various types of complex states it would develop than to actually build the system itself. I have a hunch that Ben would agree. I think what Ben was saying is that his system, like Hofstadter's Copycat would be able to avoid potentially disasterous effects of complexity that you have often ominously warned about.> </ED PORTER>==== I would not ask anyone to formally prove that the systems at AGI 2008 will work (me, of all people!). Not formal proof, just something other than a hunch. ====<ED PORTER= Glad you of all people would not demand formal proof. I think Ben and people like him have something more than just a hunch. Ben has built a lot of AI systems, he's not an amateur. Joscha Bach claims his system, which in some ways more similar to some of my thinking than Novamete says he has run tests on it that shows it scales efficiently well (although, admittedly, still in small systems), and had actually performed well at automatically learning and created hierarchical memory representations that automatically learn to perform desired functionality in the toy world his hardware budget can support. So these are not just hunches > </ED PORTER>==== > > I assume it would have been equally hard to provide "concrete reasons" > why Hofstadter's Copycat and his similar systems would work before > they were built. But they did work. Not really germane: Copycat was unbelievably simple compared to AGI systems, and Hofstadter would nover have claimed ahead of time that it would do anything, because it was an experimental system. ====<ED PORTER= I don't know how much or how little Hofstadter predicted about his systems ahead of time. Agreed AGI is much more complex than Copycat, but not as much more complex as you may think, because one of the basic concepts in most of AGI approaches is to apply a common architecture to many different AI tasks. I do think a lot of experimentation in terms of parameter tuning, refinements, etc, will be required to get a Novamente-like system to work, so to a certain extent it will be an experimental system > </ED PORTER>==== > Since computation irreducibility is something you make a big point > about in one of your own papers you frequently cite --- you --- of all > people --- should not deny the right to believe in approaches to AGI > before they are supported by concrete proof. Quite the reverse: all of these people deny that the complexity issue is at all relevant to their systems. You cannot say, on their behalf, that complexity is an excuse for not being able to predict why the systems should work, while at the same time they protest that my complex systems analysis is wrong. ;-) ====<ED PORTER= Richard, If you will remember, I actually wrote a post, admitting to having to eat my words, at least in part, saying there was something to your complex systems analysis viewpoint. > </ED PORTER>==== > But I do have, what I consider, rational reasons for believing a > Novamente-like systems will work. > > One of them is that you have failed to describe to me any major > conceptual problem for the AGI community does not have what appear to > be valid approaches. Me? What did my question have to do with me? I asked about your optimism. But since you mention me, I have (if I undertand your statement, which was a little confusingly worded) described a major, crippling reason why the AGI community does not have valid approaches. > > For more reasons why I believe a Novamente-like approach might work I > copy the following portion of a former post of from about 5 months ago > --- describing my understanding of how Novamente itself would probably > work, based on reading material from Novamente and my own other > reading and thinking. But this general description below does not really say why everything is on track to succeed. ====<ED PORTER= I agree your complexity arguments have to be kept in mind, just as does combinatorial explosion. Keeping a highly dynamic system such as a complex Novamente System from being blown away from productively performing its intended function is an important design concern --- but like combinatorial explosion, there are reasons to believe we can deal with it. One of the best of which, is that even if the system has its dynamism really damped down you should be able to still get useful work out of it. But it will probably take a lot of experimental tuning and refinement to learn how to run the system with the most efficient and productive form of dynamic control. The availiability of more cheap massively parallel hardware that will increasingly arrive over then next decade should make it possible to perform such tuning experiments in parallel, which should speed them up considerable. So net-net, Richard, I don't currently consider this a major conceptual problem --- although it might be. > </ED PORTER>==== ========================================================== ==== William Pearson Sun 4/20/2008 4:45 PM ========================================================== I'm not quite sure how to describe it, but this brief sketch will have to do until I get some more time. These may be in some new AI material, but I haven't had the chance to read up much recently. Linguistic information and other non-inductive information integrated into learning/modelling strategies, including the learning of linguistic rules. Consider an AI learning chess, it is told in plain english that "Knights move two hops in one direction and one hop 90 degrees to that". Now our AI has learnt english so how do we hook this knowledge into our modelling system, so that it can predict when it might lose or take a piece because of the position of a knight? Consider also the sentence, "There are words such as verbs, that are doing words, you need to put a pronoun or noun before the verb". People are given this sort of information when learning languages, it seems to help them. How and why does it help them? ====<ED PORTER= William, I assume you are asking how a system designed to automatically learn from experience would know how to handle knowledge handed to it in a natural language declarative form. I do not see this as a problem. A Novamente-like approach records, generalizes over, creates compositions out of, and creates a multi-level hierarchical memory of such generalization and compositions, along with episodic experiences represented as networks within such memory. As Jeff Haskins and others point out hierarchical memory provides invariant representation. This means it can not only recognize multiple different sets of inputs, such as different views of an object, as corresponding to a given concept, such as a given type of object, but also can take a given concept and map an appropriate version of it into a current context. This includes both context appropriate imagining of sensual information and the generating of context specific behaviors. Natural language involves such perception and behavior, involving both perception of words, but also their experienced connections in the hierarchical memory to other sensory, emotional, and higher level patterns and episodes. So when people are given a sentence such as the one you quoted about verbs, pronouns, and nouns, presuming they have some knowledge of most of the words in the sentence, they will understand the concept that verbs "are doing words." This is because of the groupings of words that tend to occur in certain syntactical linguistic contexts, the ones that would be most associated with the types of experiences the mind would associates with "doing" would be largely word senses that are verbs and that the mind's experience and learned patterns most often proceeds by nouns or pronouns. So all this stuff falls out of the magic of spreading activation in a Novamente-like hierarchical experiential memories (with the help of a considerable control structure such as that envisioned for Novamente). Declarative information learned by NL gets projected into the same type of activations in the hierarchical memory as would actual experiences that teaches the same thing, but at least as episodes, and in some patterns generalized from episodes, such declarative information would remain linked to the experience of having been learned from reading or hearing from other humans. So in summary, a Novamete-like system should be able to handle this alleged problem, and at the moment it does not appear to provide an major unanswered conceptual problem. > </ED PORTER>==== ========================================================== ==== Derek Zahn Sun 4/20/2008 6:29 PM ========================================================== William Pearson writes: > Consider an AI learning chess, it is told in plain english that... I think the points you are striving for (assuming I understand what you mean) are very important and interesting. Even the first simplest steps toward this clear and (seemingly) simple task baffle me. How does the concept of 'knight' poof into existence during the conversation? How does a system learn how to learn to play a game in the first place? I like this task as a tool for considering how a potential AGI approach is truly general -- by asking over and over again "how and why could that happen" for any imagining of how each sentence could be processed. Now, Edward, I hope you are right about Novamente but I don't quite follow the reasoning behind your confidence. I'm imagining that in a previous life you'd pointed me toward a drawing of a DaVinci flying machine, excitedly projecting 3-8 years until we'd be flying around. Now DaVinci's a bright guy (smarter than me) and it's a nice concept, and I can't prove it won't work -- I'd have to invent a pretty effective aerodynamic science to do so. I still might not be convinced. Absence of disproof is not necessarily strong evidence. I'm looking forward to getting more info about Novamente soon and hopefully understand the nuts and bolts of how it could do tasks like the ones William wrote about. I have some concerns about things like whether propagating truth values around is really a very effective modeling substrate for the world of objects and ideas we live in -- but since I don't understand Novamente well enough, there's lititle I can say pro or con beyond those vague intuitions (and the last thing I'd want to do is bug Ben with questions like "how would Novamente do X? How about Y?" He has plenty of real work to do.) ====<ED PORTER COMMENT= Derek, How the concept of "knight" poofs into existence during a conversation about chess is no great mystery for a Novamente-like system. If a Novamente has former experience which chess they have, within their hierarchical memory recorded patterns and experiences with chess knights, and links between them and the representation for "knight." When the context suggests the sound "night" or "knight" refers to chess, those chess "knight" patterns and experiences get activated sufficiently to be brought to the conciousness of the system. Regarding projections of a DaVinci flying machine operating in 3-8 years, if we had no helicopters, but had the rest of today's technology, machines based on the concept of DaVinci's helicopter-like design would be flying within 3-8 years. By analogy to AGI, the hardware for interesting AGI test systems is already here in the form of powerful PCs. Today for 40K you can buy systems with 16 2ghz cores and 256 GBytes of DRAM which allows even more powerful test systems. And in five or six years for even less you should be able to buy systems that should be able to prove, in the form of smaller proto-types, most of the key design elements of a human level AGI. Regarding the sufficiencly of truth values, Novamente also uses importance values, which are just as important as truth values. >From my reading and remembering of Novamente I don't remember how it dealt with the type of representations that have traditionally be represented in narrow AI as vectors, other than to say Novamente nodes can represent vectors and matrices upon which standard mathematical techniques can be used. But I don't see this as a major problem. Throw in GNG (Growing Neural Gas), and tune it to fit with the Novamente architecture and you have a way to learn vector representations that operate much as the other components in the Novamente hypegraph, and which could be used in hierarchical memory systems, like that shown to be very powerful in Thomas Serre's "Learning a Dictionary of Shape-Components in Visual Cortex: Comparison with Neurons, Humans and Machines" (In fact, one could argue the main Novamente architecture already supports something very similar to GNG.) This should allow Novamente to learn sensory patterns and behaviors within them as well as semantic ones. Of course, a Novamente-like system should be allowed to take advantage of all the wonderful front ends that have been developed for sensory and other types of information over the years by narrow AI research. So again I do not see this issue of appropriate understanding as a major missing conceptual piece of the AGI problems ></ED PORTER>==== ========================================================== ==== Linas Vepstas Mon 4/21/2008 9:12 AM ========================================================== On 20/04/2008, Derek Zahn <[EMAIL PROTECTED]> wrote: William Pearson writes: > Consider an AI learning chess, it is told in plain english that... I think the points you are striving for (assuming I understand what you mean) are very important and interesting. Even the first simplest steps toward this clear and (seemingly) simple task baffle me. How does the concept of 'knight' poof into existence during the conversation? One has to have grounding: prior experience with checkers, or parchesi or other board games, and thus the concept of moving a gamepiece on a board. And, prior to that, the concept of 3D space,e.g. that of shoving a toy around on the floor. Also, the concept of having something taken away from them: something one wants to have but can't. In humans, desires seem grounded in biology, but then, like colorful tropical birds with bzarre mating rituals, grow to be their own (biology-unmotivated) thing. Only after one has mastered these concepts is one ready to understand a knight. How does a system learn how to learn to play a game in the first place? Well, one has to be psychologically motivated to participate. There has to be some motivator to make one want to learn. My experience with children shows that they lack motivators for most things, bar one: if they can get Dad's attention, they're willing to try anything. I like this task as a tool for considering how a potential AGI approach is truly general -- by asking over and over again "how and why could that happen" for any imagining of how each I think most researchers are in general agreement with you on this. Thus the current focus on integrating knowledge-bases, and 3D spatial knowledge, with a 3D (virtual) body, and, to a lesser degree, psychological/motivational desires/needs. ====<ED PORTER= I agree with this comment. It supports what I have said above in response to comments by William Pearson and Derek Zahn. > </ED PORTER>==== ========================================================== IN SUMMARY --- FROM THE POSTS SO FAR --- I HAVE YET TO SEE ANYONE SHOW ME A MAJOR CONCEPTUAL PROBLEMS BETWEEN US AND AGI --- JUST LOTS OF WORK TO BE DONE TO BUILD, TUNE, AND REFINE OUR CURRENT DESIGNS. BUT I DON'T DENY THAT --- AS WE START TO GET PROTOTYPES OF SUCH SYSTEMS UP AND, AT LEAST PARTIALLY, RUNNING --- WE MAY LEARN OF MAJOR CONCEPTUAL PROBLEMS OF WHICH WE ARE CURRENTLY EITHER NOT AWARE OR SUFFICIENTLY APPRECIATIVE. ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
<<attachment: winmail.dat>>
