[agi] Doubling-time watcher - March 2003.
I didn't intend this to become a monthly advertisement for Dell, but if someone comes up with more bang-for-the-buck (BFTB) from someone else I would be very interested. The February 2003 most BFTB system ran $399, this month you have to spend a little more to get the best deal. $499 including Free shipping. Dell Dimension 2350 Series: Intel Celeron Processor at 1.80GHz Memory: 256MB DDR SDRAM Keyboard: Dell Quietkey Keyboard Monitor: New 17 in (16.0 in v.i.s., .27dp) E772 Monitor Video Card: Integrated Intel Extreme 3D Graphics Hard Drive: 30GB Value Hard Drive Floppy Drive and Additional Storage Devices: 3.5 in Floppy Drive Operating System: Microsoft Windows XP Home Edition Mouse: Dell 2-button scroll mouse Network Interface: Integrated 10/100 Ethernet Modem: 56K PCI Data/Fax Modem CD or DVD Drive: 48x Max CD-ROM Drive Sound Card: Integrated Audio Speakers: New Harman Kardon HK-206 Speakers Bundled Software: WordPerfect Productivity Pack with Quicken New User Edition Digital Music: Dell Jukebox powered by MUSICMATCH Digital Photography: Dell Picture Studio Image Expert Standard Limited Warranty, Services and Support Options: 1Yr Ltd Warr plus 1Yr At-Home Service + 90Days Dell SecurityCenter (McAfee) Internet Access Services: 6 Months of EarthLink Internet Access FREE! Lexmark X75 Inkjet Printer After we have a few more data points we can discuss how best to graph the power/price function as it applies specifically to the AGI application. Mike Deering, Director www.SingularityActionGroup.com --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
[agi] doubling time watcher.
Unless Ben thinks it would not be appropriate for this list, I would like to start a "doubling time" watcher monthly posting of retail computer pricesfor purposes of establishing a historical record so that questions of doubling time can be grounded in current data. My choice of category is "most bang for the buck" complete system from a major retailer or manufacturer. Usually this will be their lowest priced system, as upgrades generally cost more than the differential computational value they add. Anyone that would like to post a different category, well, you can never have too much data. My selection for "most bang for the buck" category for 2/18/03 is: Dell Dimension 2350 Series Processor: Celeron 1.7 GHz Memory: 128 MB Hard Drive: 60 GB Monitor: 15 inch CD: 48 speed Floppy drive: Y Keyboard: Y Mouse: Y GraphicsCard: Extreme 3D Graphics OS: Windows XP (HOME) Speakers: Y Sound card: Y Ethernet: Y Modem: Y Software: WordPerfect, Quicken. Price: $399 I might get one of these for my wife so she will stay off mine. We are a poor one computer family. Mike Deering. www.SingularityActionGroup.com ---new website.
RE: [agi] doubling time watcher.
It's not totally on-focus for the list, but, a monthly post on the topic certainly won't hurt. It will be interesting to see just how cheap computers do become over the next couple years! That $399 computer has a faster processor than any of my 8 machines, i believe !! -- Ben -Original Message-From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]]On Behalf Of Mike DeeringSent: Tuesday, February 18, 2003 12:00 PMTo: [EMAIL PROTECTED]Subject: [agi] "doubling time" watcher. Unless Ben thinks it would not be appropriate for this list, I would like to start a "doubling time" watcher monthly posting of retail computer pricesfor purposes of establishing a historical record so that questions of doubling time can be grounded in current data. My choice of category is "most bang for the buck" complete system from a major retailer or manufacturer. Usually this will be their lowest priced system, as upgrades generally cost more than the differential computational value they add. Anyone that would like to post a different category, well, you can never have too much data. My selection for "most bang for the buck" category for 2/18/03 is: Dell Dimension 2350 Series Processor: Celeron 1.7 GHz Memory: 128 MB Hard Drive: 60 GB Monitor: 15 inch CD: 48 speed Floppy drive: Y Keyboard: Y Mouse: Y GraphicsCard: Extreme 3D Graphics OS: Windows XP (HOME) Speakers: Y Sound card: Y Ethernet: Y Modem: Y Software: WordPerfect, Quicken. Price: $399 I might get one of these for my wife so she will stay off mine. We are a poor one computer family. Mike Deering. www.SingularityActionGroup.com ---new website.
Re: [agi] doubling time watcher.
I would like to contribute new SPEC CINT 2000 results as they are posted to the SPEC benchmark list by semiconductor manufacturers. I expect to post perhaps 10 times per year with this news. This is the source data for my Human Equivalent Computing spreadsheet and regression line. If Kurzweil and Mike Deering are right, then the new processor benchmarks should mostly appear above the existing regression line. [I hope there is time to make Cyc -or some other AGI software - safely smart before the danger of spontaneous emergence arrives.] -Steve On Tue, 18 Feb 2003, Mike Deering wrote: Unless Ben thinks it would not be appropriate for this list, I would like to start a doubling time watcher monthly posting of retail computer prices for purposes of establishing a historical record so that questions of doubling time can be grounded in current data. My choice of category is most bang for the buck complete system from a major retailer or manufacturer. Usually this will be their lowest priced system, as upgrades generally cost more than the differential computational value they add. Anyone that would like to post a different category, well, you can never have too much data. My selection for most bang for the buck category for 2/18/03 is: Dell Dimension 2350 Series Processor: Celeron 1.7 GHz Memory: 128 MB Hard Drive: 60 GB Monitor: 15 inch CD: 48 speed Floppy drive: Y Keyboard: Y Mouse: Y Graphics Card: Extreme 3D Graphics OS: Windows XP (HOME) Speakers: Y Sound card: Y Ethernet: Y Modem: Y Software: WordPerfect, Quicken. Price: $399 I might get one of these for my wife so she will stay off mine. We are a poor one computer family. Mike Deering. www.SingularityActionGroup.com---new website. --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED] -- === Stephen L. Reed phone: 512.342.4036 Cycorp, Suite 100 fax: 512.342.4040 3721 Executive Center Drive email: [EMAIL PROTECTED] Austin, TX 78731 web: http://www.cyc.com download OpenCyc at http://www.opencyc.org === --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: [agi] doubling time watcher.
I would like to contribute new SPEC CINT 2000 results as they are posted to the SPEC benchmark list by semiconductor manufacturers. I expect to post perhaps 10 times per year with this news. This is the source data for my Human Equivalent Computing spreadsheet and regression line. I'm uncomfortable with the phrase Human Equivalent because I think we are very far from understanding what that phrase even means. We don't yet know the relevant computational units of brain function. It's not just spikes, it's not just EEG rhythms. I understand we'll never know for certain, but at the moment, the possibility of guesstimating within even an order of magnitude seems premature. This isn't to say that the regression isn't a bad idea, or irrelevant to AGI design. I just don't like the title. -Brad --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: [agi] doubling time watcher.
Brad writes, Might it not be a more accurate measure to chart mobo+CPU com= bo prices? Maybe. If you wanted to research and post this data I'm sure it would be = helpful to have. Check out www.pricewatch.com. They have a search engine which ranks products by vendors. Using this, you could get lots and lots of data from one source. By averaging mean prices from the top 10 cheapest vendors, you'd wash out wierd one-time price break deals that would pollute your data if you only considered the cheapest. They also have data for complete systems. It's also probable that pricewatch keeps archived data of prices. You might consider emailing them. Finding a smart techie in their NOC who thinks AI is cool and you might get your hands on 5+ years of perfect data on every index of computing power. CPU, hard drives, tape storage, RAM, everything. -Brad --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
RE: [agi] doubling time watcher.
On Tue, 2003-02-18 at 10:48, Ben Goertzel wrote: A completely unknown genius at the University of Outer Kirgizia could band together with his grad students and create an AGI in 5 years, then release it on the shocked world. Ack! I thought this was a secret! Curses, foiled again... -James Rogers [EMAIL PROTECTED] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: [agi] doubling time watcher.
I used the assumptions of Hans Moravec to arrive at Human Equivalent Computer processing power: http://www.frc.ri.cmu.edu/~hpm/ Of course as we get closer to AGI then the error delta becomes smaller. I am comfortable with the name for now and will adjust the metric as more info becomes available. The error delta depends more on neuroscience research than AGI progress. I'm not comfortable with Moravec's calculations, but his approach of estimating based on retinal processing power is better than anything else I've read on it. Retinal neurons aren't quite the same beasts as the enormous pyramidal's that make up much of the brain though. This isn't to say that the regression isn't a bad idea, or irrelevant to AGI design. I just don't like the title. -Brad Oops, I meant to that This isn't to say that the regression *is* a bad idea. --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Billy, I agree that AGI is a complicated architecture of hundreds of separarate software solutions. But all of these solutions have utility in other software environments and progress is being made by tens of thousands of programmers each working on improving some little software function for some other purpose that they have no idea will someday be used in AGI. There is nothing truly unique about the functional building blocks of AGI, just the overall architecture. Having gone way out on a limb here, all you AGI experts can now start sawing. Mike Deering. www.SingularityActionGroup.com ---new website.
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Yeah, I don't think your statement is true... And I'm a huge advocate of the "integrative" approach. My feeling is that maybe half of the ingredients of an AGI are things that were created for other (usually narrow AI) purposes and can be used, not "off the shelf", but with only moderate rather than severe modifications. The other half are things that are certainly *related* to known science, but are unique to AGI, without asignificant use as"standalone" systems. For example, overall regulation of system attention (focus) is a big part of any AGI system, and I don't think any non-AGI algorithms are ever going to be helpful for solving the problem in an AGI context. (Novamente has its own way of doing this, which does not draw on any narrow-AI or general-Cs methods except loosely). I note that at least one famous AI guy -- Danny Hillis -- agrees with you though. That's part of the reason he gave up on AI work. He figures AGI is "just a lot of little things" and he's working on some of the little things now, not on the big picture. Of course, this philosophy is probably why his company Thinking Machines didn't have a coordinated AI research program, instead it worked on a lot of different things using a common hardware architecture... -- Ben G -Original Message-From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]]On Behalf Of Mike DeeringSent: Tuesday, February 18, 2003 3:20 PMTo: [EMAIL PROTECTED]Subject: Re: AGI Complexity (WAS: RE: [agi] "doubling time" watcher.) Billy, I agree that AGI is a complicated architecture of hundreds of separarate software solutions. But all of these solutions have utility in other software environments and progress is being made by tens of thousands of programmers each working on improving some little software function for some other purpose that they have no idea will someday be used in AGI. There is nothing truly unique about the functional building blocks of AGI, just the overall architecture. Having gone way out on a limb here, all you AGI experts can now start sawing. Mike Deering. www.SingularityActionGroup.com ---new website.
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
From recent comments here I can see there are still a lot of people out there who think that building an AGI is a relatively modest-size project, and the key to success is simply uncovering some new insight or technique that has been overlooked thus far. I would agree with that though the key is not so much insight as the term is commonly used but rather a willingness to accept the ugly truths of human intelligence... IMHO this is partly a matter of necessary optimism (i.e. we can only afford a 4-man-year project, so let's hope that will be enough), There was a fair ammount of that, especially when hardware was even tinier than it is today. =P and partly a sort of bleedover from the view of human minds that dominated the social sciences for most of the 20th century (i.e. infants are a blank slate, and blank slates sound pretty simple, so a newly-written AGI must be a relatively simple program). In some ways that is the best perspective (in contrast with Cyc which attempts to engrave everything first...) But you are right, that a topologicly flat blank slate won't work either... complex adaptive behavior requires a complex, specialized implementation. Always. No exceptions, no free lunches, no magic connectoplasmic shortcuts. The brain is actually fantasticly simple... It is nothing compared with the core of a linux operating system (kernel+glibc+gcc). Heck, even the underlying PC hardware is more complex in a number of ways than the brain, it seems... The brain is very RISCy... using a relatively simple processing pattern and then repeating it millions of times. So while an adult brain has a few billion neurons, the program which produces it is only a few megabytes in size... (The entire genome is about 750 mb, most of which is beleived to be either inactive or there purely for structural reasons). We know from the biology folks that the human mind contains at least dozens, and probably hundreds of specialized subsystems. In the cortex, I would propose the number is 28 for the left hemisphere, and maybe another 10 or so in the right hemisphere which don't directly overlap with the ones on the left. The sense of smell is strange, but the vision and motor reflexes only constitute maybe two dozen instances of maybe 5 or so distinct design patterns. (We only need to worry about the design patterns). I do agree that a early AI praject should try to replicate as much of the functionality found in the brain as possible. I will be proposing an architecture along these lines in a few months... The ones that computer scientists have tried to replicate, like vision and hearing, have turned out to contain massive amounts of complexity - computer vision alone is apparently the kind of problem that takes a good, well-funded team several decades to solve. Consider the chess problem. The present computer Chess solutions are widely acknowleged to be much less efficient than the ones in the brain. So the complexity that you are trying to argue is necessary for AGI is merely reflective of our currently poor programming methodologies. What this means for AI research is that any serious attempt to create an AGI by duplicating the way human minds work would be a massive effort, at least one and probably two orders of magnitude larger than any software development effort ever attempted. I would say that it would require maybe a dozen highly gifted devels with maybe 20 code-grunts for the support framework. That makes it much too big for current software engineering methods, so the effort would almost certainly fail. Don't implement the mind, implement the brain! =P -- I WANT A DEC ALPHA!!! =) 21364: THE UNDISPUTED GOD OF ALL CPUS. http://users.rcn.com/alangrimes/ [if rcn.com doesn't work, try erols.com ] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
I agree with your qualitative point that a computationally efficient intelligence has got to consist of a combination of specialized systems (operating tightly coupled togetherin a common framework, and with many commonalities and overlaps). However, I don't agree with your quantitative estimate that an AGI has to be orders of magnitude bigger than any software project ever attempted. I agree that many people underestimate the problem, but I think you overestimate the problem. And mis-estimate it. I think you overestimate the bulk of the problem and underestimate the subtlety of finding the right framework and the right algorithms. The brain is a hugely complex tangled mess of structures and processes, but that doesn't mean that an AGI has to be. AGI does not mean brain emulation. Legs are vastly more complex than wheels, yet wheels are good at moving around too. (And wheels can't help you invent artificial legs, whereas a nonhuman AGI can potentially help you figure out how to make a more human AGI if you want to). You mention the vast amount of work that's gone into computer vision and audition. That is true, but I think that those disciplines would be a lot more tractable if they were carried out together with AGI cognition, rather than separately. Pursuing them standalone may make them harder in many ways, rather than easier. My guess, not surprisingly, is that the Novamente design is close to the minimal level of complexity needed ;) Dozens of node and link types, a few dozen mental processes, and a couple dozen functionally-specialized units combining node and link types and processes in appropriate ways. This is a lot more complexity than the typical AI program but a lot less complexity than you seem to be alluding to. But of course, none of us *really know*. Eliezer Yudkowsky in the past has partially agreed with you, in that he's proposed the Novamente design is significantly too simple. -- Ben -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]]On Behalf Of Billy Brown Sent: Tuesday, February 18, 2003 2:54 PM To: [EMAIL PROTECTED] Subject: AGI Complexity (WAS: RE: [agi] doubling time watcher.) From recent comments here I can see there are still a lot of people out there who think that building an AGI is a relatively modest-size project, and the key to success is simply uncovering some new insight or technique that has been overlooked thus far. IMHO this is partly a matter of necessary optimism (i.e. we can only afford a 4-man-year project, so let's hope that will be enough), and partly a sort of bleedover from the view of human minds that dominated the social sciences for most of the 20th century (i.e. infants are a blank slate, and blank slates sound pretty simple, so a newly-written AGI must be a relatively simple program). Unfortunately for AI optimists, all the evidence points in the opposite direction. If we have learned nothing else about the nature of Mind in the last 50 years, we should at least have learned this: complex adaptive behavior requires a complex, specialized implementation. Always. No exceptions, no free lunches, no magic connectoplasmic shortcuts. We know from the biology folks that the human mind contains at least dozens, and probably hundreds of specialized subsystems. The ones that computer scientists have tried to replicate, like vision and hearing, have turned out to contain massive amounts of complexity - computer vision alone is apparently the kind of problem that takes a good, well-funded team several decades to solve. Now, it may be that some particular subsystems can be omitted from an AGI that isn't intended to be very humanlike. An AGI with no body may not need a kinesthetic sense or motor skills, an AGI without cameras may not need vision, and so on. But anyone who thinks there is some tiny kernel of pure thought in there waiting to be duplicated, and all the rest can be safely ignored, is just kidding themselves. Every part of the mind that we have any understanding of at all has turned out to be a tangle of complex algorithms interacting in very complex ways. There is no reason to believe the parts we don't understand are any different. What this means for AI research is that any serious attempt to create an AGI by duplicating the way human minds work would be a massive effort, at least one and probably two orders of magnitude larger than any software development effort ever attempted. That makes it much too big for current software engineering methods, so the effort would almost certainly fail. For projects that intend to implement a completely novel design, the implication is that you can't realistically expect anything like human-equivalent performance on unrestricted tasks. Evolution wouldn't have given us the equivalent of hundreds of millions of lines of specialized software if there were some easy shortcut waiting to be found. So, if you're
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Ben Goertzel wrote: But of course, none of us *really know*. Technically, I believe you mean that you *think* none of us really know, but you don't *know* that none of us really know. To *know* that none of us really know, you would have to really know. -- Eliezer S. Yudkowsky http://singinst.org/ Research Fellow, Singularity Institute for Artificial Intelligence --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Ben Goertzel wrote: And I'm a huge advocate of the integrative approach. My feeling is that maybe half of the ingredients of an AGI are things that were created for other (usually narrow AI) purposes and can be used, not off the shelf, but with only moderate rather than severe modifications. The other half are things that are certainly *related* to known science, but are unique to AGI, without a significant use as standalone systems. For the most part I agree with you (at least, until narrow-AI technology becomes more common in commercial apps). I do, however, think that a lot of people take this as an excuse to just write everything themselves. In particular, I've noticed people tend to invent their own network protocols, database systems, and other basic building blocks despite the fact that in most cases they would be better off just buying a commercial product. IMHO this is a complete waste of effort - an AI team should spend as much of its time as possible solving AI problems, not trying to optimize their file IO. Billy Brown --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
The brain is actually fantasticly simple... It is nothing compared with the core of a linux operating system (kernel+glibc+gcc). Heck, even the underlying PC hardware is more complex in a number of ways than the brain, it seems... The brain is very RISCy... using a relatively simple processing pattern and then repeating it millions of times. Alan, I strongly suggest you increase your familiarity with neuroscience before making such claims in the future. I'm not sure what simplified model of the neuron you are using, but be assured that there are many layers of complexity of function within even a simple neuron, let alone in networks. The coupled resistor/capacitor model is only given as a simplified version in textbooks to make the topic of neural networks digestible to the entry-level student. Dendrites are not simple summators, they have a variety of nonlinear processes including recursive, catalytic chemical reactions and complex second-messenger systems. That's just the tip of the iceberg once you get into pharmacological subsystems, the complexity becomes a bit staggering. If it were fanastically simple, more so than a Linux box, do you think that thousands of scientists working over more than one hundred years would still understand it so poorly, yet it takes a group of 5 people 2 years to crank out a new Linux OS? We know from the biology folks that the human mind contains at least dozens, and probably hundreds of specialized subsystems. In the cortex, I would propose the number is 28 for the left hemisphere, and maybe another 10 or so in the right hemisphere which don't directly overlap with the ones on the left. You realize that the blobs drawn on images of the brain in college level textbooks are simply areas of cell responsivity, and not diagrams of the systems themselves? The cortex is highly differentiated containing probably dozens if not hundreds of systems, not to mention the enormous variety of specialized systems at the subcortical level. The complex soup of the reticular formation is sufficient to turn a sane anatomist into a sobbing wreck with its dozens of specific nerve clusters. Consider the chess problem. The present computer Chess solutions are widely acknowleged to be much less efficient than the ones in the brain. So the complexity that you are trying to argue is necessary for AGI is merely reflective of our currently poor programming methodologies. Chess is a game designed by the mind, so it is no surprise that it is something the mind is good at. It is trivial to design games that computers are vastly superior at, but that does not mean the mind has poor programming methodologies. _Brad --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: [agi] doubling time watcher.
Brad Wyble wrote: I'm uncomfortable with the phrase Human Equivalent because I think we are very far from understanding what that phrase even means. We don't yet know the relevant computational units of brain function. It's not just spikes, it's not just EEG rhythms. I understand we'll never know for certain, but at the moment, the possibility of guesstimating within even an order of magnitude seems premature. See also Human-level software crossover date from the human crossover metathread on SL4: http://sl4.org/archive/0104/1057.html -- Eliezer S. Yudkowsky http://singinst.org/ Research Fellow, Singularity Institute for Artificial Intelligence --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Well, we invented our own specialized database system (in effect) but not our own network protocol. In each case, it's a tough decision whether to reuse or reimplement. The right choice always comes down to the nasty little details... The biggest Ai waste of time has probably been implementing new programming languages, thinking that if you just had the right language, coding the AI would be SO much easier. Ummm... Ben -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]]On Behalf Of Billy Brown Sent: Tuesday, February 18, 2003 4:13 PM To: [EMAIL PROTECTED] Subject: RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.) Ben Goertzel wrote: And I'm a huge advocate of the integrative approach. My feeling is that maybe half of the ingredients of an AGI are things that were created for other (usually narrow AI) purposes and can be used, not off the shelf, but with only moderate rather than severe modifications. The other half are things that are certainly *related* to known science, but are unique to AGI, without a significant use as standalone systems. For the most part I agree with you (at least, until narrow-AI technology becomes more common in commercial apps). I do, however, think that a lot of people take this as an excuse to just write everything themselves. In particular, I've noticed people tend to invent their own network protocols, database systems, and other basic building blocks despite the fact that in most cases they would be better off just buying a commercial product. IMHO this is a complete waste of effort - an AI team should spend as much of its time as possible solving AI problems, not trying to optimize their file IO. Billy Brown --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Well, we invented our own specialized database system (in effect) but not our own network protocol. In each case, it's a tough decision whether to reuse or reimplement. The right choice always comes down to the nasty little details... The biggest Ai waste of time has probably been implementing new programming languages, thinking that if you just had the right language, coding the AI would be SO much easier. Ummm... The thing that gives me the most confidence in you Ben is that you made it to round 2 and you're still swinging. You've personally learned the hard lessons of AGI design and its pitfalls that most of the rest of us can only imagine by analogy. -Brad --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Brad Wyble wrote: Heck, even the underlying PC hardware is more complex in a number of ways than the brain, it seems... The brain is very RISCy... using a relatively simple processing pattern and then repeating it millions of times. Alan, I strongly suggest you increase your familiarity with neuroscience before making such claims in the future. I'm not sure what simplified model of the neuron you are using, but be assured that there are many layers of complexity of function within even a simple neuron, let alone in networks. I havn't looked at the neuron in quite a while. =P But I don't consider myself [completely] insane in this context either. Dendrites are not simple summators, they have a variety of nonlinear processes including recursive, catalytic chemical reactions and complex econd-messenger systems. That's just the tip of the iceberg once you get into pharmacological subsystems, the complexity becomes a bit staggering. Yeah, the dendrite _trees_ are quite complex. My interest, however, lies in the *forest*. ;) So the question is: what program is necessary to generate a system with the same computational charactoristics as the brain? (completely ignoring the implementation details, most of which are irrelevant or artifacts of the general implementation strategy). My current understanding draws heavily on the Cerebral Code by William H. Calvin (assuming I don't have to go all the way over to the shelf to check the name). Calvin proposes what ammounts to a sophisticated, optomized Celular Automata. I'll go ahead and sketch it out here: Start with Conway's game of life... Notice that it is rather slow because of its topology, if it were more strongly connected signals could travel faster and more efficiently... To solve this we add a second layer of topology in the form of shortcuts between the varrious regions and hence we have the subcortical pathways... Now that our system is roughly brain-shaped we consider the cells individually. Conway proposed a computationally universal model which possessed only one bit of state. This system would require large numbers of cells to express concepts such as degree of magnitude and other similarly important facets. It also has no inherant distinction between situational awareness and long-term skill and memory systems making it vulnerable to computer viruses and generally too dynamic to support stable long-term behavior patterns. We solve the first problem by increasing the ammount of state the thing can carry... From a single bit we now have a vector of some unknown length (probably 10-12 8-bit words or less) that expresses the current pattern under study. This actually reduces the total complexity of the system drasticly. This system is still too dynamic, we want to ground it in a more stable system. We create two classes of state, a persistant structural state and a dynamic state that expresses the present activation of the persistant state. In almost all higher animals, a sleep period is required to clear the chaotic dynamic state of the matrix and re-initialize it from the persistant state. The reset process occours during delta wave sleep and the re-init process occours during beta wave sleep. Also during this time, the almost totally unbiased computational matrix which is the cortex is programmed through a program running on a small subset of the cortex loaded from what is essentially a ROM being the Amigdalya and hypothalamus as well as certain structures in the reticular formation. The neocortex, as far as I know, is fairly uniform in general algorithm. We only need to wire it up slightly differently for each region. I don't know wheather this applies to the older cortical regions such as the hypocampus as well. I do know that the latter structures use a different and moderately less complex algorithm... If it were fanastically simple, more so than a Linux box, do you think that thousands of scientists working over more than one hundred years would still understand it so poorly, yet it takes a group of 5 people 2 years to crank out a new Linux OS? That's not a proof at all. The evident fact that nobody has yet tried the right approach has no relationship to the nature of that correct approach. In the cortex, I would propose the number is 28 for the left hemisphere, and maybe another 10 or so in the right hemisphere which don't directly overlap with the ones on the left. You realize that the blobs drawn on images of the brain in college level textbooks are simply areas of cell responsivity, and not diagrams of the systems themselves? [/me feels a sudden intense wave of frustration.] MY LACK OF KNOWLEGE OF ANY SUCH SYSTEM IS A DIRECT RESULT OF THE DEFICIENCIES OF SAID COLEGE TEXTBOOKS. = I'm 100% self taught at this point. =\ The cortex is highly differentiated containing probably dozens if not hundreds of systems, not to mention the enormous variety of specialized
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Ben Goertzel wrote: However, I don't agree with your quantitative estimate that an AGI has to be orders of magnitude bigger than any software project ever attempted. I agree that many people underestimate the problem, but I think you overestimate the problem. And mis-estimate it. I think you overestimate the bulk of the problem and underestimate the subtlety of finding the right framework and the right algorithms. The brain is a hugely complex tangled mess of structures and processes, but that doesn't mean that an AGI has to be. AGI does not mean brain emulation. Legs are vastly more complex than wheels, yet wheels are good at moving around too. (And wheels can't help you invent artificial legs, whereas a nonhuman AGI can potentially help you figure out how to make a more human AGI if you want to). That isn't as close an analogy as it seems. A leg must do many things that wheels don't - grow, heal, resist microorganisms, raise and lower the body, cross a wide variety of rough terrain, etc. If we tried to build a machine with all of the same capabilities, it is not at all clear that it would be simpler. The brain does have a few tasks an AGI doesn't have to worry about, like metabolism and immune response. But these complexities are mostly down at the cellular level, and I wasn't arguing that an AGI has to duplicate such things. The biggest simplification I see that is relevant here is the fact that the brain must self-organize to a large extent, while an AGI could be coded in its final configuration. But AI projects usually expect most of the complexity of the final system to emerge through some kind of training process, which means you're tackling exactly the same problem. That leaves two popular options that I don't think will work out: 1) You can leave out huge chunks of functionality in the hope that they aren't needed for intelligence. This might work, but it isn't nearly as safe as it might seem. Our human version of general intelligence seems to rely heavily on drafting big specialized systems (like visualization and language) for use in new domains whose problems happen to have analogous regularities. Without a lot more knowledge than anyone currently has about how intelligence works, it seems likely that you'll omit something you can't get by without. 2) You can ignore all the messy stuff devoted to dealing with the physical world, like sensory processing and motor control, and concentrate solely on implementing abstract thought. That sounds promising, except that its exactly what most AI project have been doing for 50 years and the progress to date has been underwhelming. Besides which, that only cuts out something like 40% - 80% of the brain (depending on where you draw the line), which would still leave you with a gigantic project implementing the features you decided to keep. Do you see another option for simplification? You mention the vast amount of work that's gone into computer vision and audition. That is true, but I think that those disciplines would be a lot more tractable if they were carried out together with AGI cognition, rather than separately. Pursuing them standalone may make them harder in many ways, rather than easier. Maybe. Maybe not. To be honest, I think most people in this field have a bad habit of using general intelligence as a magic wand to gloss over hard problems that are going to require specialized mechanisms no matter how smart the overall system is. For example, in the case of computer vision, just getting from a 2D array of pixels to a possible set of object geometries takes a heck of a lot of work, and it has to be done by fast, dumb code for performance reasons. After that you have to recognize objects (a narrow problem), build a useful world-model (another narrow problem), detect and fix visual illusions and other data corruptions (yet another narrow problem), and so on. Once you have all these mechanisms you might be able to improve the results a bit by having the AI think about the output (Hmm, no, I'm sure that can't really be Santa Clause on that rooftop. It must be a Christmas display.). But you can't avoid building the specialized mechanisms in the first place. My guess, not surprisingly, is that the Novamente design is close to the minimal level of complexity needed ;) Well, of course. Otherwise you wouldn't be building it. :) But I do think there would be a lot more progress in AI if more people were building systems designed merely to solve the next obvious obstacle on the path to AGI, or to provide a platform for future work. What we have now is like a football team where the quarterback won't throw a pass unless the receiver is standing next to the goal post. Lots of long shots, little progress. OTOH, at least Novamente has enough internal complexity to reach territory that hasn't already been explored by classical AI research. I don't expect it to wake up, but I expect it will be a lot more productive than
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Alan, I strongly suggest you increase your familiarity with neuroscience before making such claims in the future. I'm not sure what simplified model of the neuron you are using, but be assured that there are many layers of complexity of function within even a simple neuron, let alone in networks. The coupled resistor/capacitor model is only given as a simplified version in textbooks to make the topic of neural networks digestible to the entry-level student. Dendrites are not simple summators, they have a variety of nonlinear processes including recursive, catalytic chemical reactions and complex second-messenger systems. That's just the tip of the iceberg once you get into pharmacological subsystems, the complexity becomes a bit staggering. agreed that the brain is enormously complex; however I think the point Alan was making hinges on a slightly different interpretation of the word complexity. His interpretation seems to be similar to that which Hofstadter elucidates in GEB; namely the idea of 'sealing off' of levels. You can look at the mind through different perspectives and at varying scales because of it's high complexity. Yet this very trait, arising from the brain's mind-boggling complexity, allows one to model it at a system-scale level. At a high enough level, you can start treating various major components as black boxes, and dealing only with their high functionality. Of course you lose a certain amount of accuracy in doing this, but it is nonetheless a valid approach. We view and deal with other people as unified personalities who we cannot 'read their mind'. Rather we observe their actions and draw conclusions about internal states that cannot be directly observed in the absence of sophisticated brain-scanning technology. Despite this limitation, we are able to interact with others and predict their future behavior and mental states to a reasonable degree. Say I'm designing an AGI architecture (which I am btw, but it is irrelevant to this discussion :) and I want to preprocess audio data so that speech is already parsed by the time it enters the AI's cognitive modules. All I need to do is obtain a preexisting natural language parser program and then tailor the AI cognitive module(s) to work w/ it's output instead of raw audio data. I don't need to even look at the parsers' code if I don't want to. (Although it may ease the use of it if I do examine it, it;s not necessary) I suppose I'm saying you can approach the mind (or any complex system that has at least vaguely recognizable functional subsystems) in a manner analogous to that of Object Oriented Programming Jonathan Standley --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
OTOH, at least Novamente has enough internal complexity to reach territory that hasn't already been explored by classical AI research. I don't expect it to wake up, but I expect it will be a lot more productive than those One True Simple Formula For Intelligence-type projects. Yes and no. You are right that an AI entity will be a complex system with a great deal of pre-programmed structure. You are mostly wrong, however, in denouncing the one true formula approach because there really is only one obvious problem standing in our way to general intelligence and that's the friggin translation problem... Its friggin cuz Hofstadter came up with it some 20 years ago and the solution has not yet been discovered... After the translation problem is solved, creating AI entities should be so simple that even an un-modified human 8-year old could do it if provided with a toolbox of drag-and-drop components... (Actually creating the components would require at least a HS level compsci course)... -- I WANT A DEC ALPHA!!! =) 21364: THE UNDISPUTED GOD OF ALL CPUS. http://users.rcn.com/alangrimes/ [if rcn.com doesn't work, try erols.com ] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Do you see another option for simplification? I am not starting from a foundational concept of brain emulation, so I'm not really faced with the problem of simplifying the brain. Maybe. Maybe not. To be honest, I think most people in this field have a bad habit of using general intelligence as a magic wand to gloss over hard problems that are going to require specialized mechanisms no matter how smart the overall system is. I like to distinguish two kinds of specialized mechanisms: 1) those that are autonomous 2) those that build specialized functionality on a foundation of general-intelligence-oriented structures and dynamics The AI field, so far, has focused mainly on Type 1. But I think Type 2 is more important. For example, in the case of computer vision, just getting from a 2D array of pixels to a possible set of object geometries takes a heck of a lot of work, and it has to be done by fast, dumb code for performance reasons. After that you have to recognize objects (a narrow problem), build a useful world-model (another narrow problem), detect and fix visual illusions and other data corruptions (yet another narrow problem), and so on. Once you have all these mechanisms you might be able to improve the results a bit by having the AI think about the output (Hmm, no, I'm sure that can't really be Santa Clause on that rooftop. It must be a Christmas display.). But you can't avoid building the specialized mechanisms in the first place. I think the general intelligence mechanisms for vision occurs at a much lower level than your example suggests. I think that object recognition and world-model-building, for example, use Type 2 specialization, not Type 1 I agree that edge detection, for example, is pure Type 1 specialization But I do think there would be a lot more progress in AI if more people were building systems designed merely to solve the next obvious obstacle on the path to AGI, or to provide a platform for future work. I think that is what the bulk of academic AI researchers are doing. The folks on this list who are actively working on AI tend to be exceptions, with more ambitious goals. What we have now is like a football team where the quarterback won't throw a pass unless the receiver is standing next to the goal post. Lots of long shots, little progress. Again, the contemporary mainstream AI field is really very conservative, concerned entirely with taking small steps in a risk-averse way. OTOH, at least Novamente has enough internal complexity to reach territory that hasn't already been explored by classical AI research. I don't expect it to wake up, but I expect it will be a lot more productive than those One True Simple Formula For Intelligence-type projects. Well I certainly hope Novamente will be more productive than that type of projec ;) However, the type of project you cite is more characteristic of AI of the 60's and 70's than of modern mainstream AI. Nearly all contemporary AI researchers are not actively seeking AGI at all; by and large, they think it's hundreds of years off, and are working on highly specialized algorithms attacking subproblems of intelligence. Which seems to be exactly what you think they should be doing! -- Ben G --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
The thing that gives me the most confidence in you Ben is that you made it to round 2 and you're still swinging. You've personally learned the hard lessons of AGI design Well, some of them ;) I'm sure there are plenty of hard lessons ahead!! -- ben and its pitfalls that most of the rest of us can only imagine by analogy. -Brad --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Say I'm designing an AGI architecture (which I am btw, but it is irrelevant to this discussion :) and I want to preprocess audio data so that speech is already parsed by the time it enters the AI's cognitive modules. All I need to do is obtain a preexisting natural language parser program and then tailor the AI cognitive module(s) to work w/ it's output instead of raw audio data. I don't need to even look at the parsers' code if I don't want to. (Although it may ease the use of it if I do examine it, it;s not necessary) From the MS Speech Development Kit genre, I believe some of the early SAPI versions, i.e. = 4.0, did some limited amount of syntactical natural language parsing along with the speech recognition. It's been some time since I looked at this, but I believe my conclusion was that it wasn't all that reliable, i.e. low % accuracy for correct POS identification?, etc. I don't know if this gets you where you want to go, but it might be worth looking at. BTB, it seems a better, more forward looking approach to your architecture might be to implement audio parsing (AP - or speech recognition SR?), natural language parsing (NLP) and cognitive processing (CP) or cognition as a coherent whole, not the other way around with separate and distinct audio parsing (AP), natural language parsing (NLP), and cognitive processing (CP) modules...as you suggest with your comments about an OO approach. In addition to the tremendous benefits of architecting something closer to real AGI, i.e. an obvious increase in the 'Goertzelian Real-AGI' level ;-), you would have the benefits of computational optimization, specifically, reduced # of ops to cognition, reduced object I/O, reduced latency, reduced processing redundancy, etc. assuming, of course, your implementation of the cognitive processing (CP) doesn't incur a tremendous overhead from the synthesis with the other two modules. I suppose I'm saying you can approach the mind (or any complex system that has at least vaguely recognizable functional subsystems) in a manner analogous to that of Object Oriented Programming Ibid. Just my $0.02 worth. EGHeflin Jonathan Standley --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
On Tue, 18 Feb 2003, Brad Wyble wrote: . . . Incorrect. The cortex has genetically pre-programmed systems. It cannot be said that is a matrix loaded with software from subcortical structures.. . . . Yes, but there is a very interesting experiment with rewiring brains of young ferrets so that visual signals from their retinas do not connect to their visual cortex area V1, but instead connect to their auditory cortex. The same banded structure of cells specialized for different line direction orientations that normally develop in the visual cortex develop in the auditory cortex. This suggests that these structures are not encoded in ferret genes, but rather are learned in response to the structure of visual stimuli. The reference is: Sharma, J., Angelucci, A., Sur. M. Induction of visual orientation modules in auditory cortex. Nature 404, 841-847. 2000. I don't pretend to know anywhere near what you do about neuroscience, but thought you might find this interesting. Cheers, Bill --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
[META: please turn line-wrap on, for each of these responses my own standards for outgoing mail necessitate that I go through each line and ensure all quotations are properly formatted...] Brad Wyble wrote: The situation for understanding a single neuron is somewhat disastrous. ... I'm just trying to give you a taste of the sophistications that are relevant to brain function and cannot be glossed over. Iff the brain is not unique in its capability to support intelligence then all of this can be replaced by some abstract model with the same basic computational charactaristics but in a very different way. So the question is: what program is necessary to generate a system with the same computational charactoristics as the brain? (completely ignoring the implementation details, most of which are irrelevant or artifacts of the general implementation strategy). The implementation details are what tells you how the brain functions. I don't care _HOW_ it functions, I care about _WHAT_ a given section accomplishes through its functioning. Given that, it should be relatively streight forward to find a work-alike Failing that, it is still possible to set up a system akin to Creatures but with a much more powerful engine and wait untill a good'nuff algorithm evolves on its own... This is, infact, my basic plan at this juncture. =) We don't know the computational characteristics yet because they are so extraordinarily complex. We don't yet completely understand how a *single synapse* functions. You're squinting too hard. My current understanding draws heavily on the Cerebral Code by William H. Calvin (assuming I don't have to go all the way over to the shelf to check the name). Calvin proposes what ammounts to a sophisticated, optomized Celular Automata. He's a fine author of pop neuroscience, but in order to be accessible he necessarily glosses over many layers of complexity. It is a mistake to take his simplified representations at face value. He needs to simplify to get his good ideas across. Use the ideas, but don't extrapolate brain functions from his simplistic depictions. rant mode engaged I HATE IVORYTOWERISM!!! IF A BOOK DOESN'T TELL IT LIKE IT IS, IT SHOULD NEVER BE PUBLISHED, EVEN TO LITTLE CHILDREN!! (Especially not to little children.) Actually, I was looking at the book for the first time in years, trying to use it as a refferance text. I gave up because the damn thing had so much fluff as to be a waste of time... (Is there a paper on the theory?) IvoryTowerism: You have to sneak onto a university campus to get anywhere near a reasonably complete library/bookstore and then pay black-market prices to cart one off... =((( Rant disengaged This system is still too dynamic, we want to ground it in a more stable system. We create two classes of state, a persistant structural state and a dynamic state that expresses the present activation of the persistant state. In almost all higher animals, a sleep period is required to clear the chaotic dynamic state of the matrix and re-initialize it from the persistant state. The reset process occours during delta wave sleep and the re-init process occours during beta wave sleep. Also during this time, the almost totally unbiased computational matrix which is the cortex is programmed through a program running on a small subset of the cortex loaded from what is essentially a ROM being the Amigdalya and hypothalamus as well as certain structures in the reticular formation. Incorrect. The cortex has genetically pre-programmed systems. It cannot be said that is a matrix loaded with software from subcortical structures.. Your are actually agreeing with me. =P The brain does have an innate structure in the form of the topology I mentioned earlier. This topology naturally leads to the development of functional systems. HOWEVER, there is no law in the *cortex* which governs what behaviors it will produce (likes, dislikes etc...) these must be inputed either from the environment or from the subcortical structures. The neocortex, as far as I know, is fairly uniform in general algorithm. We only need to wire it up slightly differently for each region. I don't know wheather this applies to the older cortical regions such as the hypocampus as well. I do know that the latter structures use a different and moderately less complex algorithm... It is not, in fact, fairly uniform. It varies in architecture (the type percentage of various cell types as well as layer thickness) as well as by connectivity with other structures. The variations are on the scale of millimeters, so there will be quite alot of them. Yes, and I don't think those varriations in layers or even connectivity are at all significant. Ofcourse you want to know which layer is for input and which layer is for feedback but you don't really worry yourself about the measurements which are probably a
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Ben Goertzel wrote: I like to distinguish two kinds of specialized mechanisms: 1) those that are autonomous 2) those that build specialized functionality on a foundation of general-intelligence-oriented structures and dynamics The AI field, so far, has focused mainly on Type 1. But I think Type 2 is more important. Hmm. Well, using your terminology, I would say that: 1) Type 2 mechanisms are only possible once you have the proper set of type 1 mechanisms (i.e. the ones that implement thought in the first place). 2) Type 2 mechanisms that are not supported by the proper type 1 mechanisms for a particular problem domain tend to be astronomically inefficient. 3) Achieving a human-like generality of intelligence is likely to require a human-like assortment of Type 1 mechanisms, except in areas where you can afford astronomical inefficiency. An obvious example of 2 is the world-model problem in robotics. If a dumb AI doesn't have a specialized mechanism for dealing with physical objects interacting in 3-D space, it just gets stuck. A smart AGI might be able to fake it by reasoning about the same data in a more abstract fashion, but this is like a human trying to aim a tennis serve with a physics book and a calculator - slow and error prone. One interesting prediction of this view is that it should be very easy to build an AI that seems promising in a domain much broader than those addressed by expert systems (like data analysis or even logical reasoning), and yet fails miserably when you try to introduce it to some other challenge humans consider routine (like predicting where a tennis ball will go after it gets hit). In other words, the brittleness problem may be intractable. I think the general intelligence mechanisms for vision occurs at a much lower level than your example suggests. I think that object recognition and world-model-building, for example, use Type 2 specialization, not Type 1 In the case of object recognition, that would be possible but amazingly inefficient compared to a type 1 approach. For a world model I don't see how it is possible at all, unless you artificially limit what kinds of facts about the world you need to work with. I think that is what the bulk of academic AI researchers are doing. The folks on this list who are actively working on AI tend to be exceptions, with more ambitious goals. Again, the contemporary mainstream AI field is really very conservative, concerned entirely with taking small steps in a risk-averse way. Nearly all contemporary AI researchers are not actively seeking AGI at all; by and large, they think it's hundreds of years off, and are working on highly specialized algorithms attacking subproblems of intelligence. Which seems to be exactly what you think they should be doing! Not exactly. It isn't that I think we should give up on AGI, but rather that we should be consciously planning for it to take several decades to get there. We should still tackle the problems in front of us, instead of giving up on real AI work altogether. But we need to get past the idea that every AI project should start from scratch and end up delivering a human-equivalent AGI, because that isn't going to happen. We just aren't that close yet. The way the software industry has solved big challenges in the past is to break them up into sub-problems, figure out which sub-problems can be solved right now, solve them as thoroughly as possible, and offer the resulting solutions as black boxes that can then become inputs into the next round of problem solving. That's what happened with operating systems, and development environments, and database systems. If we want to see real progress in AI, the same thing needs to happen to problems like NLP, computer vision, memory, attention, etc. Too bad there isn't much of a market for most of those partial solutions... Billy Brown --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
[META: please turn line-wrap on, for each of these responses my own standards for outgoing mail necessitate that I go through each line and ensure all quotations are properly formatted...] I think we're suffering from emacs issues, I'm using elm. Iff the brain is not unique in its capability to support intelligence then all of this can be replaced by some abstract model with the same basic computational charactaristics but in a very different way. I totally agree. But the genesis of this debate was whether the brain is complicated in a non-trivial way. The fact that it is complicated does not mean it cannot be replicated in a different substrate (and like Ben, I think it would be a misapplication of effort to try). The implementation details are what tells you how the brain functions. I don't care _HOW_ it functions, I care about _WHAT_ a given section accomplishes through its functioning. The nature of neuroscience research doesn't really differentiate between the two at present. In order to understand WHAT a brain part does, we have to understand HOW it, and all structures connected to it function. We need to understand the inputs and the outputs, and that's all HOW. There are people who approach the problem from a purely black-box perspective of course, by giving people memory tests and looking at the pattern of failures. This is extremely interesting work, particularly as regards the types of errors people make while speaking. (http://www.wjh.harvard.edu/~caram/pubs.htm) I don't think it's sufficient, on its own, to figure out the brain without simulteanously looking at the neural data. Given that, it should be relatively streight forward to find a work-alike well, it just isn't. Brains are hard to reverse engineer, and that's basically what you're talking about. Failing that, it is still possible to set up a system akin to Creatures but with a much more powerful engine and wait untill a good'nuff algorithm evolves on its own... It took evolution billions of years with an enormous search space. Obviously we can speed the process. But in the end, you'd end up an equally inscrutable mass of neural tissue. You'd be better off getting yourself a real kid :) rant mode engaged I HATE IVORYTOWERISM!!! IF A BOOK DOESN'T TELL IT LIKE IT IS, IT SHOULD NEVER BE PUBLISHED, EVEN TO LITTLE CHILDREN!! (Especially not to little children.) My comment was in the context of you saying that the brain is fantastically simple and then citing Calvin as a source for your conclusion. I'm saying that books by pop authors are insufficient to draw conclusions from, not that they are useless. His ideas are great, I love his work. The brain does have an innate structure in the form of the topology I mentioned earlier. This topology naturally leads to the development of functional systems. HOWEVER, there is no law in the *cortex* which governs what behaviors it will produce (likes, dislikes etc...) these must be inputed either from the environment or from the subcortical structures. I disagree with this, but I see where you are coming from. We don't know enough about the cortex to say things like this. The reason that subcortical structures seem more concrete to us, is that they are simpler in design and therefore easier to understand than cortical structures. Yes, and I don't think those varriations in layers or even connectivity are at all significant. Ofcourse you want to know which layer is for input and which layer is for feedback but you don't really worry yourself about the measurements which are probably a biproduct of having more neurons in those regions that are heavily connected and not, in themselves, interesting... The extray layers in the occipital lobe are probably nothing more than the equivalent of a math coprocessor in a computer... The addition or deletion of layers is going to drastically change the nature of computations a given bit of cortex performs. I've spent 8 years studying hippocampal anatomy. It is fascinating and highly structured in a way the cortex isn't (or its simplicity allows us to perceive the structure). Vast volumes of data about its anatomy are available and I have read most of it. GIMME GIMME GIMME!!! =P I said I read it, I didn't say I could remember all of it :) I( and the rest of the hippocampal community) am at a loss to tell you how it functions. Do we know what it does? (how its outputs relate to its inputs) Nope. We think it might have to do with spatial navigation in rodents (rats tend to think in terms of 2-D space) and more complex types of memory in higher order critters. Anatomy and neurophysiology seem to suggest it should relate memory to motor actions and behavioral states, but lesion it and animals seem relatively unimpaired in that respect(lesions are a troublesome way to reverse engineer the brain). *throws up
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
I believe that the precision with which digital computers can do things, will allow intelligence to be implemented more simply on them than in the brain. This precision allows entirely different structures and dynamics to be utilized, in digital AGI systems as opposed to brains. For example, it allows correct probabilistic inference calculations (which humans, at least on the conscious level) are miserable at making; it allows compact expression of complex procedures as higher-order functions (a representation that is really profoundly unbrainlike); etc. I'd be curious to hear more about what you mean by this last statement. You are referring to the nature of nesting complex function calls within one another? Brad No, higher-order functions is a technical term from the theory of functional programming. It refers to the use of functions that have functions as arguments. For instance, the derivative operator in calculus is a higher-order function: it maps functions into functions. So, the type of the real function x^2 is R--R, but the type of the derivative operator is [R--R]--[R--R] so the derivative is a second-order function... Programming languages like Haskell (www.haskell.org) use higher-order functions to achieve remarkably compact programs doing very complex things. These programs are not terribly intuitive to most humans, mainly because our limited stack size runs into trouble when dealing with functions deeper than maybe third-order... Combinatory Logic, invented by Haskell Curry in the 50's, is a foundation for mathematics based on higher-order functions, see e.g. http://www.cwi.nl/~tromp/cl/cl.html The Novamente design involves using higher-order functions to represent complex procedures and patterns. There are a lot of technical advantages to this. For one thing, it allows one to express extremely complex mathematical patterns without using any variables Having complex patterns expressed with no variables is good for Novamente's reasoning algorithms; variables as used in ordinary non-combinator math would complicate things TERRIBLY (as we discovered in Webmind). Anyway, this is a very deep and technical topic; I introduced it as an example of the kind of direction you can get led in when you think NOT about the human brain but rather about the FUNCTIONS carried out by the brain and how to most effectively carry them out in a digital computer context. Higher-order function representations are not robust in the sense that neural representations probably are: they aren't redundant at all, one error will totally change the meaning. They're not brainlike in any sense. But maybe (if my hypothesis is right) they provide a great foundation for complex procedure learning and pattern recognition in a digital computer context. They seem to integrate very nicely with the other parts of Novamente, anyhow. -- Ben G --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Not exactly. It isn't that I think we should give up on AGI, but rather that we should be consciously planning for it to take several decades to get there. We should still tackle the problems in front of us, instead of giving up on real AI work altogether. But we need to get past the idea that every AI project should start from scratch and end up delivering a human-equivalent AGI, because that isn't going to happen. We just aren't that close yet. The way the software industry has solved big challenges in the past is to break them up into sub-problems, figure out which sub-problems can be solved right now, solve them as thoroughly as possible, and offer the resulting solutions as black boxes that can then become inputs into the next round of problem solving. That's what happened with operating systems, and development environments, and database systems. If we want to see real progress in AI, the same thing needs to happen to problems like NLP, computer vision, memory, attention, etc. In as much as I'm a neurophile, I disagree that this is the best approach. AI research has been having a hard time making progress by working on little black boxes and then hooking them together. I think without the context of the whole entity (the top level AGI), it's harder to think about and implement solutions to the black box problems. Evolution certainly didn't work with black boxes. It made functionally complete organisms at each step of the way, and I think AI design can work in the same manner. The progress of bottom-up, whole organism roboticism, ala Rod Brooks, is an impressive example of what can happen when you attack the whole organism simultaneously. The top level thinking is grounded in the structure of the representations used by the lower level stuff that actually interacts with the world. Now this agrees with most of what you are saying, namely that we can't implement a cloud in the sky AGI that thinks in a vacuum. But it disagrees with you in saying that we can't afford to work on these sub-problems without the context of the entire organism. -Brad --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
RE: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Ben Goertzel wrote: I like to distinguish two kinds of specialized mechanisms: 1) those that are autonomous 2) those that build specialized functionality on a foundation of general-intelligence-oriented structures and dynamics The AI field, so far, has focused mainly on Type 1. But I think Type 2 is more important. Hmm. Well, using your terminology, I would say that: 1) Type 2 mechanisms are only possible once you have the proper set of type 1 mechanisms (i.e. the ones that implement thought in the first place). Well my Type 1 and Type 2 are both specialized-intelligence mechanisms. I also posit general-intelligence mechanisms, which are separate from Type 1 and Type 2 specialized intelligence mechanisms. In the Novamente design, we have three generalized intelligence mechanisms: * higher-order probabilistic inference * evolutionary learning * reinforcement learning each with its own strengths and weaknesses. We also have some complementary specialized cognitive mechanisms, like first-order inference, neural-net-like association-finding, cluster formation, etc. Specialized intelligence components may be built on top of these. For instance, language processing uses aspects of all these (e.g. parsing is largely unification, an aspect of higher-order inference) Or, for something like edge detection, we would use a type 1 specialized mechanism, and general intelligence wouldn't enter into it at all. But we need to get past the idea that every AI project should start from scratch and end up delivering a human-equivalent AGI, because that isn't going to happen. We just aren't that close yet. I don't think all of us are trying to start from scratch. I'm certainly not, I'm using a lot of ideas developed by others over the past few decades. The way the software industry has solved big challenges in the past is to break them up into sub-problems, figure out which sub-problems can be solved right now, solve them as thoroughly as possible, and offer the resulting solutions as black boxes that can then become inputs into the next round of problem solving. That's what happened with operating systems, and development environments, and database systems. If we want to see real progress in AI, the same thing needs to happen to problems like NLP, computer vision, memory, attention, etc. I completely disagree. Building a complex self-organizing system is not like building an ordinary engineered software system. You can't design the parts in isolation. You have to design each part with explicitly consciousness of the whole. Which means it has to be a unified project, not a collection of disparate subprojects aimed at producing black boxes to later be hooked together. This is a profound difference between minds on the one hand, and OS's, DB's and IDE's on the other. And I still say, this is pretty much exactly the approach that conventional academic AI is taking. There is a conventional breakdown of the AI problem into subproblems (of which you've listed several), and people tend to work on each one separately. I don't understand how what you suggest is different from what nearly everyone in the field is doing. -- Ben G --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
Higher-order function representations are not robust in the sense that neural representations probably are: they aren't redundant at all, one error will totally change the meaning. They're not brainlike in any sense. But maybe (if my hypothesis is right) they provide a great foundation for complex procedure learning and pattern recognition in a digital computer context. They seem to integrate very nicely with the other parts of Novamente, anyhow. -- Ben G /me dispenses a You are on the right track, dude. medal. -- I WANT A DEC ALPHA!!! =) 21364: THE UNDISPUTED GOD OF ALL CPUS. http://users.rcn.com/alangrimes/ [if rcn.com doesn't work, try erols.com ] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
Re: AGI Complexity (WAS: RE: [agi] doubling time watcher.)
The nature of neuroscience research doesn't really differentiate between the two at present. In order to understand WHAT a brain part does, we have to understand HOW it, and all structures connected to it function. We need to understand the inputs and the outputs, and that's all HOW. I wouldn't say even that much... The exact format of the IO is not necessary either but only the general Information X Y and Z is carried to here from here. We don't even know what the information is, honestly. Cells fire spikes. Sometimes there are clear behavioral correlates which makes it easy to figure out (place cells), usually not. The spike firing code depends on the function of the underlying structures. We have to know how they represent information to know what information is being transmitted. Understand, by the way, that there are plenty of computational and mathematical specialists working on this, applying plenty of information theoretic approaches. I've seen a very interesting report on the reverse engineering of the hearing system though I am still months away from finishing my first reading of Principles of neuroscience. The primary modalities are the easiest systems to decode, because you can control precisely what the inputs are. Those are the first systems to be decoded. Yes, that is because they don't constitute a computer. I suppose you need a really deep understanding of what computation is to see how the cortex is a computer (and hence has all the same properties of nonpredictability and such...) Well it computes. So... it's a computer, sure. Feel free to tell me more. Does it really? ;) I would suggest that the individual cortical columns represent a fairly consistient set of adaptive logic gates (of considerable complexity). I would further suggest that as the ferrit example showed the computation the cortical region performs depends mostly on where in the logic network the inputs are sent and the outputs taken. In this way you can take just about any cortical region and get it to do just about anything any other region does (except for the extra layers of the occipital lobe) just by hooking it up differently... I don't really have any strong data for or against that hypothesis. We're not sure how brittle columns are, functionally. Simple neural net models tell us though that it's very easy to drastically alter the functional character of a network by changing one parameter. I'll read the ferret example, but I'm guessing that all they found was evidence of striation, which doesn't mean the system is working correctly. However, given the resilience of the brain to changes performed at a young age, it is likely there was some visual perception. Where is the evidence for celular differentiation beyond the 20 or so classes of neurons? I'm not talking just about neuron types, but also about connection patterns of neurons between and within areas. Subregions CA3 and CA1 of the hippocampus are identical from a cellular composition perspective, but their connectivity patterns are so different that noone who studies the system would expect them to do the same thing. Neurophysiological evidence demonstrates that they do in fact differ in their functional characteristics. Absent this evidence, how can you say that a certain structure of cells X, Y, and Z which are arranged in layers 1-6 in cortical region A do something significantly different from those in region B? For starters, an autoassociative network performs differently than a heteroassociative one. Or add noradrenergic modulation(or one of 10+ other neuromodulators), or delete a subclass of GABA cells, or triple the percentage of stellate cells. It is easy to make a neural network behave differently. This is easily demonstrated with models. --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]