In response to the below post from Mike Tintner of 10/4/2007 12:33 PM: You talk about the Cohen article I quoted as perhaps leading to a major paradigm shift, but actually much of its central thrust is similar to ideas that have been around for decades. Cohens gists are surprisingly similar to the scripts Schank was talking about circa 1980.
Again, I think the major paradigm shift needed for AGI is not so much some new idea that blows everything away, but rather a realization of how most of the basic problems in AI have actually been solved at a conceptual level, an appreciation of the power of the concepts we already have, and understanding of what they could do if put together and run on brain level hardware that has human-level world knowledge, and a focus on learning how to pick and chose from all the ideas the right components and getting them to all work together well automatically on such really powerful hardware. As Goertzel points out in his articles on Novamente -- and as anyone who has thought about the problem understand -- even with brain level hardware you have to come up with good context appropriate schemes for distributing the computational power you have to where it is most effective. This is because no mater how great your computational power is, it will always be infinitesimal compared to the massively combinatorial space of possible inferences and computations. There are lots of possible schemes for how to do this, including sophisticated probabilistic inference and context specific importance weighting. But until I see results from actual world-knowledge-size systems running with various such algorithms, I cant begin to understand how big a problem it is to get things to work well. Regarding your disappointment that Cohens schema operated at something close to a predicate logic level -- far removed from the actual sensation of from which one would think they would be derived -- I expressed a similar sentiment in my response to you are now responding. A good human-level system should have much more visually grounding, and a much more sophisticated on at that. But that is not meant as a criticism of Cohen's work, because he is trying to get stuff done on relatively small hardware. At the risk of repeating myself, check out the visual grounding in Thomas Serres great article about a visual recognition system (http://cbcl.mit.edu/projects/cbcl/publications/ps/MIT-CSAIL-TR-2006-028.p df). I have cited this article multiple times in the last week, but it blows me away. This article does not pretend to explain the whole story in visual recognition, but it explains a lot of it. It gives a very good idea of both the hierarchical memory that Jeff Hawkings and many others are heralding as the solution to much of the non-literal match problem (previously one of the major problems in AI), it gives a pretty good feel for the types of grounding our brains actually use, it demonstrates the importance of computer power, thru simulations, in brain understanding, and it is a damn powerful little system. To the extent that there are new paradigms, this article captures a few of them. You will note that the type of hierarchical representation used in Serres paper would not normally be comparing views of similar objects at a pixel level, but at levels higher up in its hierarchical memory scheme that are derived from pixel level mappings against the different views separately. So schemas of the type Cohen talks, if operating at a semantic level on top of a hierarchical representation like that used in Serre would not be operating at anything close to the pixel level, but they could be quickly mapped to, or from, the pixel level and intermedial levels in between. Implications from such intermediate representations could be combined with those from the semantic level to improve semantic implication from visual information. They could also be used by imagination from such intermediary representations, semantically relevant information such as generalizations of how, or whether or not, a context appropriate view of an objecte would fit in a given context. (I don't think Serre focuses much on top down processessing, except mainly for inhibition of less relevant upward flow, but there has been much work on top down information flows, so it is not hard to imagine how it could be mapped into his system.) As the Serre article shows, grounding of semantic representations in the pixel level and more importantly in the many levels between the semantic and the pixel level is possible with todays hardware in limited domains. It should be fully possible across all sensory domains with the much more powerful hardware that the Serres of the future will be working on. Edward W. Porter Porter & Associates 24 String Bridge S12 Exeter, NH 03833 (617) 494-1722 Fax (617) 494-1822 [EMAIL PROTECTED] -----Original Message----- From: Mike Tintner [mailto:[EMAIL PROTECTED] Sent: Thursday, October 04, 2007 12:33 PM To: [email protected] Subject: Re: [agi] breaking the small hardware mindset Edward P: II skimmed LGIST: Learning Generalized Image Schemas for Transfer Thrust D Architecture Report, by Carole Beal and Paul Cohen at the USC Information Sciences Institute. It was one of the PDFs listed on the web link you sent me (at http://eksl.isi.edu/files/papers/cohen_2006_1160084799.pdf). It was interesting and valuable. I found its initial few pages a good statement of some solid AI ideas. Its idea of splitting states based on entropy is a good one, one that I have myself have considered as a guide for when and where in semantic space to split models and how to segment temporal representations. Thanks for pointing this out. My v. quick impression is that it is a step, at least, to a major paradigm shift (although all your criticisms may be valid). [JWJohnston - I only saw this site literally today after I had posted] However - correct me - their "image schemas" are symbolically represented. They are not true image schemas in my or Lakoff/Mark Johnson's terms. I believe one of the central sources of the brain's adaptive power is the ability to represent, manipulate and compare visual and other kinds of graphics/ "image schemas" directly. IOW it can represent "an agent goes to a place" as (speaking very roughly): outline graphics of - a circle or similar (for "agent") - an arrow (for "goes to") - another circle or square (for "place"). (AFAICT this is consistent with Lakoff & Johnson's thinking). If I ask you or any human to tell me a story about "an agent going to a place", you will of course, be able to tell me a virtually infinite number of stories - a prime example of the brain's adaptive power and ability to draw analogies. That ability, I believe, derives from being able to directly, visually transform a circle or similar into almost any object or creature . Thus you will be able to tell me a story about a human/man/woman/rabbit/snake etc for your "agent." That ability can also visually transform an arrow or similar into any form of object movement - into say a human running/walking/driving/riding a bus etc. for "goes to" - and can transform a square into a skyscraper/ town/ shop/ forest etc. for "place." (One obvious piece of evidence for this is the brain's ability to see any objects whatsoever their shape as balls on an abacus - it's the foundation of our ability to count objects and maths). But all this - as I understand - is beyond digital computers. They can't handle visual shapes directly - no "imagination." And that is just one of the absolute brick walls AGI faces, which no amount of tweaking will overcome. P.S. I have to say I wasn't that impressed by the other 2 papers of Cohen linked by JWJohnston. But thanks also for pointing them out ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?&; ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=50026498-5f2264
