John, I am sure there is interesting stuff that can be done. It would be interesting just to see what sort of an agi could be made on a PC.
I would be interested in you Ideas for how to make a powerful AGI without a vast amount of interconnect. The major schemes I know about for reducting interconnect involve allocating what interconnect you have to the links with the highest probability or importance, varying those measures of probability and importance in a contest specific way, and being guided by prior similar experiences. Ed Porter -----Original Message----- From: John G. Rose [mailto:[EMAIL PROTECTED] Sent: Tuesday, December 04, 2007 1:42 AM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] Ed, Well it'd be nice having a supercomputer but P2P is a poor man's supercomputer and beggars can't be choosy. Honestly the type of AGI that I have been formulating in my mind has not been at all closely related to simulating neural activity through orchestrating partial and mass activations at low frequencies and I had been avoiding those contagious cog sci memes on purpose. But your expose on the subject is quite interesting and I wasn't that aware that that is how things have been being done. But getting more than a few thousand P2P nodes is difficult. Going from 10K to 20K nodes and up, getting more difficult to the point of being prohibitively expensive to being impossible or extremely lucky. There are ways to do it but according to your calculations the supercomputer mayt be more of a wise choice as going out and scrounging up funding for that would be easier. Still though (besides working on my group theory heavy design) exploring the crafting and chiseling of an activation model you are talking about to the P2P network could be fruitful. I feel that through a number of up front and unfortunately complicated design changes/adaptations that the activation orchestrations could be improved thus bringing down the message rate requirements, reducing activation requirements, depths and frequencies, through a sort of computational resource topology consumption, self-organizational design molding. You do indicate some dynamic resource adaption and things like "intelligent inference guiding schemes" in your description but it doesn't seem like it melts enough into the resource space. But having a design be less static risks excessive complications... A major problem though with P2P and the activation methodology is that there are so many variances in the latencies and availability that serious synchronicity/simultaneity issues would exist that even more messaging might be required. Since there are so many variables in public P2P, empirical data also would be necessary to get a gander on feasibility. I still feel strongly that the way to do AGI P2P (with public P2P as core not augmental) is to understand the grid, and build the AGI design based on that and what it will be in a few years, instead of taking a design and morphing it to the resource space. That said, there are finite designs that will work so the number of choices is few. John _____________________________________________ From: Ed Porter [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 6:17 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] John, You raised some good points. The problem is that the total number of messages/sec that can be received is relatively small. It is not as if you are dealing with a multidimensional grid or toroidal net in which spreading tree activation can take advantage of the fact that the total parallel bandwidth for regional messaging can be much greater than the x-sectional bandwidth. In a system where each node is a server class node with multiple processors and 32 or 64Gbytes of ram, much of which is allocable to representation, sending messages to local indices on each machine could fairly efficiently activate all occurrences of something in a 32 to 64 TByte knowledge base with a max of 1K internode messages, if there was only 1K nodes. But in a PC based P2P system the ratio of nodes to representation space is high and the total number of 128 byte messages/sec than can be received is limited to about 100, so neither methods of trying to increase number of patterns than can be activated with the given interconnect of the network buy you as much. Human level context sensitivity arises because a large number of things that can depend on a large number of things in the current context are made aware of those dependencies. This takes a lot of messaging, and I don't see how a P2P system where each node can only receive about 100 relatively short messages a second is going to make this possible unless you had a huge number of nodes. As Richard Loosemore said in his Mon 12/3/2007 12:57 PM post. "It turns out that within an extremely short time of the forst word being seen, a very large numbmer of other words have their activations raised significantly. Now, whichever way you interpret these (so called "priming") results, one thing is not in doubt: there is massively parallel activation of lexical units going on during language processing." With special software, a $10M dollar supercomputer cluster with 1K nodes, 32TBytes of Ram, and a dual ported 20Mb infiniband interconnect send about 1 to 5 billion 128byte messages/sec. Since there are only 1K nodes that means a global activation would require a max of 1K internode messages. If you had 10K P2P nodes each with 2G of ram dedicated to representation, each of which could receive only about 100 128byte message/sec, you would have a total message limit of 1M 128 msg/sec addressing 20TBytes of Ram. A global activation would require 10K messages meaning only 100 could be done a second. Of course with such fine grained nodes many activations of all occurrences of a given pattern would not activate all 10K nodes, but many would probably activate at least a third of them, meaning only 100 to 300 system wide activations of the occurrences of a pattern could be done a second, which is about 10K times slower than the above mentioned super computer. If you had 1M P2P nodes each with 2G of ram dedicated to representation, each of which could receive only about 100 128byte message/sec, you would have a total message limit of 100M 128 msg/sec addressing 2000TBytes of Ram. This is 1/10 to 1/50 the number of similarly sized messages on the super computer. But because of the much larger number of separate machines, a larger number of message is required for a spreading activation from a pattern to allow of the other patterns in which it occurs. A global activation talking to all compute nodes would require 1M messages meaning only 100 could be done a second. But due to the fact that much of world knowledge is sparsely connected a complete activation of all of a pattern's occurrences might activate anywhere roughly between 1 and 10M patterns. We don't know the average level of interconnection appropriate for world knowledge, but let us assume an average of occurrences for the average pattern. (It should be remembered however that activations are often made not only from a given pattern or concept node, but also from a set of similar pattern nodes. Let us assume on average each message to a P2P node activates 2 related pattern nodes in its RAM, so only 10K messages on aveage are needed for each activation, that would allow 10K nodes to spread full activation a second. This figure is just a guessed average of the total number of occurrences for the average pattern in the world knowledge. So the 1M node P2P network could only due full-occurrence activations for 10K pattern/sec, which is 1/100th the amount that would be allowed on the super computer mentioned above. Fully activating 10K nodes a second might sound like a lot, until you realize that spreading activation is typically done by relaying such spreading activation through multiple levels of implication, with the number of activated nodes growing somewhat exponentially at each level unless one greatly prunes down the percent of links that get activated at each successive radial level out in the search. Say you have 10 pattern nodes a second that are added to your context, and for each we send out full activations from it and 10 similar nodes. That would use 100 of the 10K average full activations (i.e., 20Kpattern nodes, and 10K messages each), so now we have 2M second level partial activations. Lets wildly guess the number of activated second level pattern nodes should be reduced to 200K because there probably is a lot of overlap in these activations. These 200k 2nd level patterns have to share an amount of messaging equal to 9900 full average activations (assumed above to be 10K messages each). This only allows an average of 1/20 a full activation or 500 messages from each of the 200k second level activations. It would not leave any messages for from the millions of third level activations that received such messages. And this doesn't account for spreading activation from similar nodes at each successive generation of an activation. It also doesn't allow for the fact that in a system with multiple changing constraints on a node's activation level, a node's activation level can change many times a second, and if the change is large enough, news of such changes should be relayed via multi-level spreading activation multiple times a second, not only for the new concepts activated each second, but also for many of the concept nodes in the context that have been previously activated in STM. In Shurti's system he assumes activations take place at roughly a gama wave frequencies (30-40Hz), and that most slot filler concepts are repeated about 4 times a second, and that relationship concepts are sending messages to each multiple times a second to update the values of relationships they are connected to. The 100M inter node messages in the 1M node P2P example, should be large enough to at least do interesting exploration in large AGI. If one could greatly limit the combinatorial rate of growth of spreading activation messaging, one might actually be able to do some impressive multilevel. Implication. At this time we just don't know how well implication can be controlled and how much in the way of spreading activation messaging will be required. The 1M node P2P example, which would allow 100M 128 byte messages/sec, would allow roughly 10 activations a second each activating 2k other patterns, each activing 200 other patterns, each activating 20 other patterns (The actual number would be higher because of the overlap of activations). You should be able to do some sort of inferences with this, particularly if one takes into account some of features mentioned in my recent prior email to Richard Loosemore. If by indicies of indices, John, you mean a probabilistic hiearachy of activation patterns used to activate each other, you might actually be able to do something impressive in improving the efficiency of activation. In the Shruiti example I copied in the picture I sent out earlier today (and am copying again) the implication required activation across 8 nodes from "fall" to "cleans". This is a radial search so it requires an average of four hops from each of those two nodes. You can see the amount of inferencing provided by the 1M node P2P network would not allow such 4 level deep activation across any significant part of many large pattern spaces at a rapid speed. Even the above mentioned supercomputer with several million activations a second would not, but with intelligent inference guiding schemes it would at least have a much better shot. Of course, it is my belief that in this "John Fell in the Hallway. Tom had cleaned it. He was hurt." Shastri example shown in the attached figure, we humans are actually likely to have at least one individual pattern which include slipping and falling on a wet floor that had been cleaned all in one pattern, requiring many fewer hops. The fact that we humans have very large experiential knowledge base in one of the things that makes search relatively efficient, because our large number of relatively complex patterns greatly reduced the amount of search that is required for solution to many problems. We do not know the number and width of the spreading activation that is necessary for human level reasoning over world knowledge. Thus, we really don't know how much interconnect is needed and thus how large of a P2P net would be needed for impressive AGI. But I think it would have to be larger than say 10K nodes. Ed Porter << File: SHRUTI IMPLICATION-2.jpg >> -----Original Message----- From: John G. Rose [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 12:37 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] > From: Ed Porter [mailto:[EMAIL PROTECTED] > Once you build up good models for parsing and word sense, then you read > large amounts of text and start building up model of the realities > described > and generalizations from them. > > Assuming this is a continuation of the discussion of an AGI-at-home P2P > system, you are going to be very limited by the lack of bandwidth, > particularly for attacking the high dimensional problem of seeking to > understand the meaning of text, which often involve multiple levels of > implication, which would normally be accomplished by some sort of search > of > a large semantic space, which is going to be difficult with limited > bandwidth. > > But a large amount of text with appropriate parsing and word sense > labeling > would still provide a valuable aid for web and text search and for many > forms of automatic learning. And the level of understanding that such a > P2P > system could derive from reading huge amounts of text could be a > valuable > initial source of one component of world knowledge for use by AGI. I kind of see the small bandwidth between (most) individual nodes as not a limiting factor as sets of nodes act as temporary single group entities. IOW the BW between one set of 50 nodes and another set of 50 nodes is quite large actually and individual nodes' data access would depend on - indexes of indexes to minimize their individual BW requirements. Does this not apply to your model? John ----- 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=71808695-d7fd06