Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
On Tue, Mar 20, 2007 at 09:20:06PM -0700, Chuck Esterbrook wrote: > I generally agree, but wanted to ask this: Shouldn't AGIs be visual in > focus because we are? We want AGIs to help us with various tasks many The best way to extract information is by observing the environment. The only way to build effective robotics platform is use vision, or an equivalent high-bandwidth channel. This is the acid test for any useful general AI. If it can't learn to deal with extracting knowledge from a high-bandwidth, noisy feed then it's not much of a general AI. > of which will require looking at diagrams, illustrations and pictures. So you're still relying on human-input material, only at a higher level. Instead of doing it explicitly, you're feeding it with symbols and diagrams, and expect it to have just enough symbol recognition power. > And that's just the static material. > > I guess I worry about the applicability... Would a blind AGI really be > able to find more effective treatments for heart disease, cancer and > aging? A blind AI without a body does really have very little intersection with the human world. It could as well be not there. > Regarding Numenta, they tout "irrespective of scale, distortion and > noise" and they chose a visual demonstration, so it seems that at > least their AGI work is deserving of Kevin's criticism. Still, I give > them credit for breaking ground and pushing forward. It may turn into > something yet. And we need more AGI efforts, not less. Not just quantity, also diversity. -- Eugen* Leitl http://leitl.org";>leitl http://leitl.org __ ICBM: 48.07100, 11.36820http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE - 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/?list_id=303
Re: [agi] My proposal for an AGI agenda
On 3/21/07, Chuck Esterbrook <[EMAIL PROTECTED]> wrote: Sometimes the slowness of a program is not contained in a small portion of a program. Sure. For us however this isn't the case. Cobra looks nice, very clean to read, even more so than Python. However the fact that it's in beta and .NET sort of kills it for us. As we will be going into high performance computing, we have no choice but to do the core work in plain C running on Linux. Shane - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
On 3/21/07, Eugen Leitl <[EMAIL PROTECTED]> wrote: The best way to extract information is by observing the environment. The only way to build effective robotics platform is use vision, or an equivalent high-bandwidth channel. This is the acid test for any useful general AI. If it can't learn to deal with extracting knowledge from a high-bandwidth, noisy feed then it's not much of a general AI. Hi Eugen, This opinion is *biased* by placing too much emphasis on sensory / vision. I tried to build such a vision-centric AGI a couple of years back, and found that it has severe deficiencies when it comes to *symbolic* and logical aspects of cognition. If you spend some time thinking about the latter domains, you'd likely change your mind. But the current status of neuroscience is such that vision is the most understood aspect of the brain, so the vision-centric view of AGI is prevalent among people with a strong neuroscience background. YKY - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
Having had a long standing interest in machine vision I find this demo deeply unimpressive. It's a classic example of the kind of "toy" problem which may work well within its very restricted domain but when supplied with more realistic data (i.e. photos of real objects) fails catastrophically. Scalability from theoretical principle to practical perception system is a key requirement. Many folks in the machine vision world become seduced by two dimensions. Because photos or TV screens are two dimensional they reckon so must be the processes which interpret them. It should be remembered that two dimensional objects in a photo are just shadows of a higher dimensional realm. I believe that we can only recognise 2D objects because during the early stages of our lives we have learned to be able to abstract out certain commonly observed shadows from the full range of spatio-temporal experience, effectively substituting a higher dimensional representation with a quick lookup. Fundamentally though we are creatures designed to interpret three or four dimensional events (or more if you include multi-modal integration). I believe that the invariant object recognition problem has already been solved, but that most people just don't yet realise it. If you imagine that the fundamental mode of representation of the world consists of three or four dimensional objects, with quick lookups for common 2D shadows, then its quite easy to see how its possible to view a photo, perform a mental transformation and then draw the object shown in the photo as it might appear from a different angle. This skill of course relies upon early embodied experiences of the world. In an AGI context such experiences could be gained either using a physical robot, or within a simulated world such as Ben's. The ability to perform mental transformations isn't only useful for vision, and I think it's no coincidence that skilled scientists, mathematicians or other creative people are often described to possess good "mental imagery" skills. On 21/03/07, Chuck Esterbrook <[EMAIL PROTECTED]> wrote: "recognition ... irrespective of scale, distortion and noise" sounds pretty interesting. Are these capabilities outside of current NNs? I'm familiar with NNs ignoring noise, but not scale. But my NN investigations are several years old... I wonder if distortion includes any degree of rotation. I don't have time for the demo this night. - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
On Wed, Mar 21, 2007 at 06:12:45PM +0800, YKY (Yan King Yin) wrote: >Hi Eugen, This opinion is *biased* by placing too much emphasis on >sensory / vision. I tried to build such a vision-centric AGI a couple We know that logic is easy. People only had to learn to deal with it evolutionary only recently, and computers can do serial symbol string transformations quite rapidly. Already computer-assisted proofs have transformed a branch of mathematics into an empirical science. Building world/self models in realtime from noisy, incomplete and inconsistent data takes a lot of processing, and parallel processing. For some reason traditionally AI considered the logic/mathematics/formal domain for hard, and vision as easy. It has turned out exactly the other way round. Minsky thought porting SHRDLU to the real world was a minor task. Navigation and realtime control, especially cooperatively is hard. We've desintegrated into discussion minutiae (which programming language, etc.) but the implicit plan is to build a minimal seed that can bootstrap by extracting knowledge from its environment. The seed must be open-ended, as in adapting itself to the problem domain. I think vision is a reasonable first problem domain, because insects can do it quite well. You can presume that a machine which has bootstrapped to master vision will find logic a piece of cake. Not necessarily the other way round. I understand some consider self-modification a specific problem domain, so a system capable of targeted self-inspection and self-modification can self-modify itself adaptively to a given task, any given task. I think there is absolutely no evidence this is doable, and in fact there is some evidence this is a Damn Hard problem. Do you think this is arbitrary and unreasonable? >of years back, and found that it has severe deficiencies when it comes >to *symbolic* and logical aspects of cognition. If you spend some >time thinking about the latter domains, you'd likely change your >mind. But the current status of neuroscience is such that vision is >the most understood aspect of the brain, so the vision-centric view of >AGI is prevalent among people with a strong neuroscience background. I think there's merit in recapitulating the capabilities as they arose evolutionary. We're arguably below insect level now, both in capabilities, and from the computational potential of the current hardware. It's best to learn to walk before trying to win the sprinter Olympics, no? -- Eugen* Leitl http://leitl.org";>leitl http://leitl.org __ ICBM: 48.07100, 11.36820http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE - 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/?list_id=303
Re: [agi] My proposal for an AGI agenda
On 3/21/07, Shane Legg <[EMAIL PROTECTED]> wrote: On 3/21/07, Chuck Esterbrook <[EMAIL PROTECTED]> wrote: > Sometimes the slowness of a program is not contained in a small > portion of a program. Sure. For us however this isn't the case. Cobra looks nice, very clean to read, even more so than Python. However the fact that it's in beta and .NET sort of kills it for us. As we will be going into high performance computing, we have no choice but to do the core work in plain C running on Linux. .NET and Mono. The latter is developed on Linux and Cobra releases are tested on Mono. But I can see how the beta status excludes it from efforts that are already underway. -Chuck - 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/?list_id=303
Re: [agi] My proposal for an AGI agenda
--- "YKY (Yan King Yin)" <[EMAIL PROTECTED]> wrote: > On 3/20/07, rooftop8000 <[EMAIL PROTECTED]> wrote: > > Hi, I've been thinking for a bit about how a big collaboration AI project > > could work. I browsed the archives and i see you guys have similar > > ideas > > > > I'd love to see someone build a system that is capable of adding any > > kind of AI algorithm/idea to. It should unite the power of all existing > > different flavors: neural nets, logical systems, etc > > > > [...] > > > > But you could always allow people to make their own critics (modules) > > in any way, as long as they are executable. So people can still use there > > preferred way of doing things and contribute. > > > > Singh described a central hierarchy of critics, but it will be probably > > be better to distribute it somehow, so different people can run their > > own hierarchy and communicate results over the web to other > > people (connecting different hierarchies into a big one) > > They can also fine-tune things (spend cpu-time on what they find > > interesting) and maybe locally test out their own critics. > > > > This can scale very well. People can offer their own modules > > for people to run and reflective critics will probably be designed that > favor > > the most useful modules. > > > > Singh uses 1 type of representation for communication between all > > the modules. This isn't really necessary, you can use any data structure > for > > critic to critic communication. But you'll need to make flexible ones for > > the critics you code, or people won't collaborate. > > > > Do you guys think all this makes sense? Are there any previous > collaboration > > attempts like this that i should be aware of? > > Thanks > > Just one big question: are you way *overestimating* the number of people > who could contribute such modules? Is your idea like, people contributing > modules for face-recognition, fruit-recognition, license-plate recognition, > etc, etc, and you join these modules together? What mechanism do you have > to ensure that you would end up with a system that can recognize *all* > things? > > It seems that you're taking the idea of online collaboration too far. > (Please correct me if I've mistaken your position...) > > YKY > I think splitting things up in small manageable modules will allow this. Not all the work is developing some killer learning algorithm, you also need a lot of modules in between that do simpler tasks For example simple translation modules or modules that handle connections to peers over the web... Modules that manage a knowledge database... I'm sure a lot of people can help. I am not sure this ends up being able to recognize all things. I will be happy to see a system where people can add their ideas ( new and existing) to with minimal effort, but thats a huge step in the good direction, i think I still have no idea how it could work though, ideas are welcome :) Food fight? Enjoy some healthy debate in the Yahoo! Answers Food & Drink Q&A. http://answers.yahoo.com/dir/?link=list&sid=396545367 - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
Chuck Esterbrook wrote: On 3/20/07, Ben Goertzel <[EMAIL PROTECTED]> wrote: I would certainly expect that a mature Novamente system would be able to easily solve this kind of invariant recognition problem. However, just because a human toddler can solve this sort of problem easily, doesn't mean a "toddler" level AGI should be able to solve it equally easily. Different specific modalities will come more naturally to different intelligences, and humans are particularly visual in focus... I generally agree, but wanted to ask this: Shouldn't AGIs be visual in focus because we are? We want AGIs to help us with various tasks many of which will require looking at diagrams, illustrations and pictures. And that's just the static material. Eventually, yeah, a useful AGI should be able to process visual info, just like it should be able to understand human language. But I feel that the strong focus on vision that characterizes much AI work today (especially AI work with a neuroscience foundation) generally tends to lead in the wrong direction, because vision processing in humans is carried out largely by fairly specialized structures and processes (albeit in combination with more general- purpose structures and processes). So, one can easily progress incrementally toward better and better vision processing systems, via better and better emulating the specialized component of human vision processing, without touching the general-understanding-based component... Of course, the same dynamic happens across all areas of AI (creating specialized rather than general methods being a better way to get impressive, demonstrable incremental progress), but it happens particularly acutely with vision Gary Lynch, in the late 80's, made some strong arguments as to why olfaction might in some ways be a better avenue to cognition than vision. Walter Freeman's work on the neuroscience of olfaction is inspired by this same idea. One point is that vision processing has an atypically hierarchical structure in the human brain. Olfaction OTOH seems to work more based on attractors and nonlinear dynamics (cf Freeman's work), sorta like a fancier Hopfield net (w/asymmetric weights thus leading to non fixed point attractors). The focus on vision has led many researchers to overly focus on the hierarchal aspect rather than the attractor aspect, whereas both aspects obviously play a bit role in cognition. I guess I worry about the applicability... Would a blind AGI really be able to find more effective treatments for heart disease, cancer and aging? IMO vision is basically irrelevant to these biomedical research tasks. Direct "sensory" connections to biomedical lab equipment would be more useful ;-) Regarding Numenta, they tout "irrespective of scale, distortion and noise" and they chose a visual demonstration, so it seems that at least their AGI work is deserving of Kevin's criticism. I agree. Poggio's recent work on vision processing using brain models currently seems more impressive than Numenta's, in terms of combining -- far greater fidelity as a brain simulation -- far better performance as an image processing system But, the Numenta architecture is more general and may be used very interestingly in future, who knows... -- Ben - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
FYI...After reading Hawkins book I actually believe that his ideas may indeed underlie a future AGI system...but they need to be fleshed out in much greater detail... Cheers, K Their current concept implementation did not change substantially since their first "proof-of-concept" implementation. That was two years ago. Their "solution" of invariance is extracting groups spatial and temporal patterns, but still only works for very very trivial problems. Instead of pushing through this needless platform, they should return to theory. Currently, its simply unimpressive. - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
On Wed, Mar 21, 2007 at 08:21:57AM -0400, Ben Goertzel wrote: > Eventually, yeah, a useful AGI should be able to process visual info, > just like it should be able to understand human language. Being able to learn to see and to learn to hear, yes? How much of it do you expect to be hardwired? E.g. part of what cochlea does directly in hardware is a Fourier transform. Do you expect to start with that as a prepositioned building block, or let the system figure out the appropriate transformation on its own? > But I feel that the strong focus on vision that characterizes much > AI work today (especially AI work with a neuroscience foundation) > generally tends to lead in the wrong direction, because vision > processing in humans is carried out largely by fairly specialized > structures and processes (albeit in combination with more general- There's a reason for that, it's in the law of optics and the kind of structures a camera sees in the world. Vision (or an equivalent high-bandwidth channel, direct depth perception by TOF, or LIDAR, whatever) is a basic instrument for knowledge extraction. You can skip on that by making the agent directly sense the simulation grid (externalising the representation), but you'd have to abandon that if your system does its first steps in the real world. > purpose structures and processes). So, one can easily progress > incrementally > toward better and better vision processing systems, via better and > better emulating the specialized component of human vision processing, > without touching the general-understanding-based component... Only parts of the visual processing pathway are "hardwired" (not really, and it's not a linear thing at all), and of course the upper stages use every trick the neocortex can muster. So, no, I don't think you can trivialize vision, or postpone it. > Of course, the same dynamic happens across all areas of AI > (creating specialized rather than general methods being a better > way to get impressive, demonstrable incremental progress), but it happens > particularly acutely with vision > > Gary Lynch, in the late 80's, made some strong arguments as to why > olfaction might in some ways be a better avenue to cognition than vision. > Walter Freeman's work on the neuroscience of olfaction is inspired by > this same idea. The bit rate you get from olfaction is really low. Yes, you can sense gradients, and if you code everything with volatile carriers you can recognize about everything. What I don't like about olfaction is that it's evolutionary even older than vision, and directly wired to attention allocation (emotion) processes. It's like vision, only without the advantages, and even more hardwired. > One point is that vision processing has an atypically hierarchical > structure in the > human brain. Olfaction OTOH seems to work more based on attractors > and nonlinear dynamics (cf Freeman's work), sorta like a fancier Hopfield > net (w/asymmetric weights thus leading to non fixed point attractors). The > focus on vision has led many researchers to overly focus on the hierarchal > aspect rather than the attractor aspect, whereas both aspects obviously > play > a bit role in cognition. Makes sense. > Direct "sensory" connections to biomedical lab equipment would > be more useful ;-) It would be interesting to see which sensory modalities are optimal e.g. for medical voxelsets. One could wire a 3d retina on the voxelset, or let the system look at the space/time domains directly, and let it build its own processing and representation. -- Eugen* Leitl http://leitl.org";>leitl http://leitl.org __ ICBM: 48.07100, 11.36820http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
Eventually, yeah, a useful AGI should be able to process visual info, just like it should be able to understand human language. Being able to learn to see and to learn to hear, yes? How much of it do you expect to be hardwired? E.g. part of what cochlea does directly in hardware is a Fourier transform. Do you expect to start with that as a prepositioned building block, or let the system figure out the appropriate transformation on its own? If we were given a bunch of $$ to embody Novamente in a robot today, and I had to design a vision system for Novamente, I would not try to imitate human visual cortex. Rather, I would * use a combination of lidar and camera input * write code that took this combined input to make a 3D contour map of the perceived surfaces in the world * use standard math transforms to triangulate this contour map * use some AI heuristics (with feedback from the more general AI routines) to approximate sets of these little triangles by larger polygons * finally, feed these larger polygons into the "polygon vision" module we have designed for NM in a sim-world context This is not what the human brain does, but I don't see why it wouldn't work OK -- though I'm sure a whole lot of "small problems" would emerge requiring ample tuning Novamente is not supposed to be an emulated brain. The interesting cognitive problems of vision processing arise at the level of "polygon vision" and don't have that much to do with the lower levels. We just don't have staff bandwidth to play with this stuff right now, but that doesn't mean I find it uninteresting or unimportant. -- Ben - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
I think making direct comparisons between computational power and the animal kingdom has always been a questionable exercise, but I generally agree with trying to tackle problems in a similar order that evolution did, because evolution needed to find incremental solutions. I've long wanted to build an intelligent robot, and by that I mean *really* intelligent, not just faking it for a few minutes with some Eliza-like program. I would like to develop systems for abstract reasoning and concept formation, but unless the robot can actually see and experience its world these reasoning skills are totally useless (unless of course it were a super-arrogant robot - "I'm so smart I don't even need to be able to see, because in my state of singularitarian enlightenment I know the complete state of the universe!"). So I got stuck on the vision problem, and there I have remained to this day. On 21/03/07, Eugen Leitl <[EMAIL PROTECTED]> wrote: I think there's merit in recapitulating the capabilities as they arose evolutionary. We're arguably below insect level now, both in capabilities, and from the computational potential of the current hardware. - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
On 21/03/07, Ben Goertzel <[EMAIL PROTECTED]> wrote: * use a combination of lidar and camera input * write code that took this combined input to make a 3D contour map of the perceived surfaces in the world * use standard math transforms to triangulate this contour map * use some AI heuristics (with feedback from the more general AI routines) to approximate sets of these little triangles by larger polygons * finally, feed these larger polygons into the "polygon vision" module we have designed for NM in a sim-world context This is very much the traditional machine vision approach, described by Moravec and others and used with some success recently in the DARPA Grand Challenge. I'm also following the same approach which is a very straightforward application of standard engineering techniques. The logistics of doing this are quite complicated, involving camera calibration, correspondence matching and probabilistic spatial modelling and I think the sheer complexity (and drudgery) of the programming task is the reason why few people have ever attempted to do this so far. Being able to create large scale voxel models which can be maintained in a computationally efficient manner suitable for real time use also involves some fancy algorithms. I would agree that where things start to become interesting are at the polygon level, but you still need to maintain an underlying voxel model of space because you can't calculate probability distributions accurately using polygons alone. - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
I've read the whitepaper on HTM's and NuPIC. It seems more of a marketing strategy to attract laypeople, since I can't see anything it can solve that a NN (a recurrent and well designed/evolved one with a little preprocessing of input) can't. On 3/21/07, Chuck Esterbrook <[EMAIL PROTECTED]> wrote: "recognition ... irrespective of scale, distortion and noise" sounds pretty interesting. Are these capabilities outside of current NNs? I'm familiar with NNs ignoring noise, but not scale. But my NN investigations are several years old... I wonder if distortion includes any degree of rotation. I don't have time for the demo this night. -- Forwarded message -- From: Numenta, Inc. <[EMAIL PROTECTED]> Date: Mar 20, 2007 6:30 PM Subject: Numenta Newsletter: March 20, 2007 To: [EMAIL PROTECTED] Numenta, Inc. Newsletter March 20, 2007 Dear newsletter subscriber: I am pleased to make several announcements regarding the Numenta Platform for Intelligent Computing (NuPIC). First, we now have available on our web site a demonstration program using the Pictures example which enables recognition of simple line drawings irrespective of scale, distortion and noise. This demo is a Windows(r) client application, and is very easy to use and install. The Windows program accesses NuPIC running on the Numenta Linux cluster in California to perform the recognition task, so you do not need to have Linux installed on your local machine. However, you must be connected live to the Internet in order to use the demo. You can find the demo here: http://www.numenta.com/about-numenta/technology/pictures-demo.php The purpose of this demo is to give you a sense of the capabilities of the platform without having to install the platform itself. And, the demo is a lot of fun to use! (Note that the complete platform only is available on Linux and MacOS today - see below for more information about NuPIC and Windows.) The demo is particularly well-suited for non-engineers, as well as being a good starting point for those engineers who want to get a feel for the platform before going through the full installation process. Second, several developers have commented in our forums expressing interest in a Windows development environment. I'd like to let you know that we are working on this request, and you will see us respond in two phases. We shortly will have information on better packaging of a virtual Linux machine running under Windows. In addition, we plan to add the ability to run NuPIC on native Windows, but that will take longer. Those of you who have subscribed to the developer newsletter will receive further details shortly. Finally, I wanted to let you know that we are pleased with the response to our launch. We've had nearly 2000 people download the platform, and have been excited to watch the forum and wiki sections become active. We're just at the beginning of an exciting new era, and want to thank you for your early and enthusiastic participation. Donna Dubinsky CEO, Numenta This email was sent to [EMAIL PROTECTED], by [EMAIL PROTECTED] Update Profile/Email Address http://ui.constantcontact.com/d.jsp?p=oo&m=1101575916857&ea=chuckesterbrook%40yahoo.com&se=336&t=1101584457428&lang=en&reason=F Instant removal with SafeUnsubscribe(TM) http://ui.constantcontact.com/d.jsp?p=un&m=1101575916857&ea=chuckesterbrook%40yahoo.com&se=336&t=1101584457428&lang=en&reason=F Privacy Policy: http://ui.constantcontact.com/roving/CCPrivacyPolicy.jsp Powered by Constant Contact(R) www.constantcontact.com Numenta, Inc. | 1010 El Camino Real | Suite 380 | Menlo Park | CA | 94052 - 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/?list_id=303 -- = Rafael C.P. = - 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/?list_id=303
Re: [agi] My proposal for an AGI agenda
I totally agree that maximum speed is required for AGI. I agree with Ben that flexabilty is also needed (he might have said liked, not needed) like seen in Ruby. My language IDE has a small simple editor built-in that automatically reformats the code and compiles as it saves making compiling totally invisible. Edit/compile/run turnaround is as fast as one click of the mouse. My system, although unintentionally somewhat like Ruby (my system was totally designed without any knowledge of Ruby) has almost all of it's flexability while retaining almost the speed of C++. My worst case internal looping is 15x slower than C++ but most functions and overall program execution are much closer to C++ speed. The looping overhead on my machine is about 10M/sec for a store operation. My looping speed is 2-3 times faster than Visual FoxPro which is very fast for a compiler/byte code interpreter. I haven't tested Ruby on my machine but I would suggest it's looping overhead to be at least 100-200 times slower than C++. My indexing times are similiar to Visual FoxPro or better. I have never seen a faster indexing/data access system than Visual FoxPro (any bench mark between MySQL and Visual FoxPro would be no contest at all). I can import 100,000 names/addresses in 1.5 seconds and search every record in this file by first name, last name and prov in 1/10 second (data searched is 17.8M and all records are stored on the disk). I have tested full text search on 16K emails representing 100M of data in 1 second where the data is stored on the disk (it would be faster if only in memory). I have had 2.5K programs running at the same time and created tables of over 2M records. I have created arrays up to 10M elements. (size is limited to memory space) -- David Clark - Original Message - From: "Chuck Esterbrook" <[EMAIL PROTECTED]> To: Sent: Tuesday, March 20, 2007 7:43 PM Subject: Re: [agi] My proposal for an AGI agenda > Sometimes the slowness of a program is not contained in a small > portion of a program. Especially, after a few rounds of profiling and > improving performance render the profile graph fairly flat. > > I experienced this on a financial program that I ported from Python to > C# for an 18 X speed up (and I had already improved the Python program > itself 2 - 4 X (memory getting fuzzy...) before running out of > reasonable ways to make it faster). So then instead of having to wait > 20 minutes per run to study the results and continue development, I > only needed to wait a little over a minute. That made a *huge* > improvement in speed of development. > > Unfortunately, C# is not a rapid coding language like Python and Ruby > and consequently some of the speed up (but not all) was lost in > development. > > There are some applications that don't suffer much from this either > because they are overwhelmed by external factors (such as network > latency), or they are a small part of the system (providing glue), or > they have a specific piece of code that overwhelms performance which > can be pushed out to C or C++ (tight inner loop). > > But there's plenty of programs that are down right fascinating *and* > compute intensive including AGI, GP/GA/EC, NNs, simulation, finances, > compression, search and more. Many of these require speed in order to > work with non-trivial data sets while still enabling a fast "run, > study, edit" loop. And, of course, you want quick, high level coding > to keep the "run, study, edit" loop tight as well. > > I think we can have both with the right language design. I'll finish > this off in a response to Ben in this same thread... > > -Chuck > > - > 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/?list_id=303 - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
- Original Message - From: "Eugen Leitl" <[EMAIL PROTECTED]> To: Sent: Wednesday, March 21, 2007 4:32 AM Subject: Re: [agi] Fwd: Numenta Newsletter: March 20, 2007 > On Wed, Mar 21, 2007 at 06:12:45PM +0800, YKY (Yan King Yin) wrote: > > We know that logic is easy. People only had to learn to deal with > it evolutionary only recently, and computers can do serial symbol > string transformations quite rapidly. Already computer-assisted > proofs have transformed a branch of mathematics into an empirical > science. Your first 3 points aren't without debate. I have seen no general/good logic based system yet, have you? "computer-assisted proofs" is a specialized domain and doesn't make a dent in AGI IMO. Another example might be chess programs but few would say that is AGI. > Building world/self models in realtime from noisy, incomplete and inconsistent > data takes a lot of processing, and parallel processing. For some reason > traditionally AI considered the logic/mathematics/formal domain for hard, > and vision as easy. It has turned out exactly the other way round. > Minsky thought porting SHRDLU to the real world was a minor task. > Navigation and realtime control, especially cooperatively is hard. Regardless of how easy Minsky or others thought porting from small artificial domains to the real thing would be, that is hardly the point. I agree 100% that "Navigation and realtime control" is a hard problem but so are higher level systems like language and semantic learning. AGI needs both and needs them to communicate intelligently. > We've desintegrated into discussion minutiae (which programming language, etc.) > but the implicit plan is to build a minimal seed that can bootstrap by > extracting knowledge from its environment. The seed must be open-ended, > as in adapting itself to the problem domain. I think vision is a reasonable > first problem domain, because insects can do it quite well. You can presume > that a machine which has bootstrapped to master vision will find logic a > piece of cake. Not necessarily the other way round. I understand some > consider self-modification a specific problem domain, so a system capable > of targeted self-inspection and self-modification can self-modify itself > adaptively to a given task, any given task. I think there is absolutely > no evidence this is doable, and in fact there is some evidence this is > a Damn Hard problem. It's funny how many times you have mentioned parallel hardware and other esoteric minutiae of a very low level but I for one, have enjoyed that info very much. I too, have worked with robotics, sonar range finders, micro controllers etc. I appreciate how hard it is to deal with real world data, but that is hardly all that an AGI should be capable of. What evidence or experiments do you have to substantiate that "a machine which has bootstrapped to master vision will find logic a piece of cake"? To dismiss whole quadrants of AI development with just a conclusion, doesn't seem very tolerant or warranted from the current data. The ability of "self-inspection and self-modification" is a prerequisite IMO to creating a system that doesn't need to be hand coded in it's entirety by human programmers. It is possible that some relatively small amount of code could be used and all the smarts could be into the data but history has shown IMO that as the complexity of the data goes up, sub-languages are created to make sense of the data. As the level of interpreter goes up, the speed falls precipitously. Data and programs have been shown in many ways to be interchangeable with some problems going one way and others going the other. Do you believe that some tiny (relative to the size of the AGI) algorithm can be found that will magically create all the systems required for AGI intelligence? I am not sure what you mean exactly by "no evidence". I have made programs that create other programs (on the fly) for almost 20 years. The chances any of these generated programs would be the same is highly unlikely. I don't presume to stretch that into implying I have created a program to produce programs for "any given task" but I don't think your (or my) brain can do that either. > Do you think this is arbitrary and unreasonable? I would say your conclusions are "arbitrary and unreasonable" even if I am happy that you are thinking hard about areas of the AGI problem that desperately need attention. > I think there's merit in recapitulating the capabilities as they arose > evolutionary. We're arguably below insect level now, both in capabilities, > and from the computational potential of the current hardware. Why does our silicon based hardware always have to be compared with "carbon based units"? Computers don't have to have the requirement that they contain all the information to reproduce themselves as humans do. Our AGIs are not limited to a single building block, like human DNA and the brain synapse. A single AGI could conta
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
In my previous email, I mistakenly edited out the part from Yan King Yin and it looks like the "We know that logic is easy" was attributed to him when it was actually a quote of Eugen Leitl. Sorry for my mistake. -- David Clark - Original Message - From: "David Clark" <[EMAIL PROTECTED]> To: Sent: Wednesday, March 21, 2007 10:33 AM Subject: Re: [agi] Fwd: Numenta Newsletter: March 20, 2007 > - Original Message - > From: "Eugen Leitl" <[EMAIL PROTECTED]> > To: > Sent: Wednesday, March 21, 2007 4:32 AM > Subject: Re: [agi] Fwd: Numenta Newsletter: March 20, 2007 > > > > On Wed, Mar 21, 2007 at 06:12:45PM +0800, YKY (Yan King Yin) wrote: > > > > We know that logic is easy. People only had to learn to deal with > > it evolutionary only recently, and computers can do serial symbol > > string transformations quite rapidly. Already computer-assisted > > proofs have transformed a branch of mathematics into an empirical > > science. - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
> > We've desintegrated into discussion minutiae (which programming language, > etc.) > but the implicit plan is to build a minimal seed that can bootstrap by > extracting knowledge from its environment. The seed must be open-ended, > as in adapting itself to the problem domain. I think vision is a reasonable > first problem domain, because insects can do it quite well. You can presume > that a machine which has bootstrapped to master vision will find logic a > piece of cake. Trying to make a seed AI is the same as hoping to win the lottery. You're just hoping you only have to do one thing so you can forget about all the other stuff that is required. And if i could pick things that wouldn't be needed in a seed AI, it would be real-world vision and motor skills. I agree that understanding movement and diagrams and figures is essential to thought, but why would a computer program need to do recognize a picture of a chair or a picture of a horse or be able to track a flying bird in the sky? I don't think that's required for most problems. I also don't see how you get to all other thoughts from there? (Not that it can't be useful to have in your system..) >Not necessarily the other way round. I understand some > consider self-modification a specific problem domain, so a system capable > of targeted self-inspection and self-modification can self-modify itself > adaptively to a given task, any given task. I think there is absolutely > no evidence this is doable, and in fact there is some evidence this is > a Damn Hard problem. I agree. you can only do some minor self-modification if you don't fully understand your inner workings/code. I still think most of this AGI will have to coded by hand, and it will be a lot of software engineering and not the romantic seed AI or minimal subset of 10 perfect algorithms... Seems like people don't seem to want to put in all the energy and keep looking for a quick solution Get your own web address. Have a HUGE year through Yahoo! Small Business. http://smallbusiness.yahoo.com/domains/?p=BESTDEAL - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
On Wed, Mar 21, 2007 at 12:00:09PM -0700, rooftop8000 wrote: > Trying to make a seed AI is the same as hoping to win the lottery. Winning the lottery is an unbiased stochastical process. Darwinian co-evolution is a highly biased stochastical process. Seeds are one-way hashes (morphogenetic code expands the small seed into a structure which is appropriately positioned (environment-shaped) to extract knowledge from the environment (aka doting parents). Such seeds can be relatively tiny, see fertilized human eggs (the womb does not seem to contribute noticeable amounts of complexity). Hence they contain far less complexity than an adult, winning which by stochastical process, unbiased or otherwise, takes terrible odds. > You're just hoping you only have to do one thing so you can forget about > all the other stuff that is required. No. I don't think that other stuff required can be done. This is the same reason I don't subscribe to SENS. I thought this was unlikely when I was a 15 year old, and I still think it's unlikely as a 40 year old. > And if i could pick things that wouldn't be needed in a seed AI, it would be > real-world vision and motor skills. I agree that understanding movement and Learning from the environment takes navigation in and active manipulation of the environment. The G in AGI doesn't stand for domain-specific. > diagrams and figures is essential to thought, but why would a computer > program > need to do recognize > a picture of a chair or a picture of a horse or be able > to track a flying bird in the sky? I don't think that's required > for most problems. I also don't see how you get to all other thoughts from > there? > (Not that it can't be useful to have in your system..) > > >Not necessarily the other way round. I understand some > > consider self-modification a specific problem domain, so a system capable > > of targeted self-inspection and self-modification can self-modify itself > > adaptively to a given task, any given task. I think there is absolutely > > no evidence this is doable, and in fact there is some evidence this is > > a Damn Hard problem. > > I agree. you can only do some minor self-modification if you don't fully > understand your inner workings/code. I have reasons to suspect that a system can't understand its inner workings well enough to do radical tweaks. Well, we can (in theory) mushroom our cortex by a minimal genetic tweak. That's a trivial modification, which doesn't reengineer the microarchitecture. Live brain surgery on self or a single copy doesn't strike me as a particularly robust approach. Add a population of copies, and a voting selection or an external unbiased evaluator, and you're already in Darwin/Lamarck country. > I still think most of this AGI will have to coded by hand, and it will I don't think this is doable by mere humans. This is a few orders of magnitude below of what the maximum complexity ceiling is (tools only take you that far). If AI is numerics, Fortran+MPI would be enough. C would be arguably less painful. If AI is not numerics, you're equally screwed, whether this is Lisp, Erlang, Ruby or Fortran. > be a lot of software engineering and not the romantic seed AI or minimal To clarify, I'm only interested in ~human equivalent general AI, and only in co-evolution from a reasonable seed pool in a superrealtime virtual environment heavily skewed towards problem-solution as fitness function as a design principle. The only reason for this is that it looks as if all other approaches are sterile. You're of course quite welcome to prove me wrong by delivering a working product. > subset of 10 perfect algorithms... Seems like people don't seem to > want to put in all the energy and keep looking for a quick solution My estimate is several % of yearly GNP for several decades for a likely success by above design mechanism. If you call that a quick solution, many will disagree. -- Eugen* Leitl http://leitl.org";>leitl http://leitl.org __ ICBM: 48.07100, 11.36820http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
On Wed, Mar 21, 2007 at 10:47:35AM -0700, David Clark wrote: > In my previous email, I mistakenly edited out the part from Yan King Yin and > it looks like the "We know that logic is easy" was attributed to him when it > was actually a quote of Eugen Leitl. > > Sorry for my mistake. It's not your mistake. It's the mistake of those who choose to ignore http://www.netmeister.org/news/learn2quote.html It is really a great idea to use plaintext posting and set standard quoting in your MUA. For those with braindamaged MUAs there are workarounds like http://home.in.tum.de/~jain/software/outlook-quotefix/ -- Eugen* Leitl http://leitl.org";>leitl http://leitl.org __ ICBM: 48.07100, 11.36820http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE - 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/?list_id=303
Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
I put up with 1 person out of all the thousands of emails I get who insisted on sending standard text messages as a attachment. Because of virus infections, I had normally set all emails with attachments to automatically get put in the garbage can. I had to stop that so I could read your emails for the past 2 years. You have a lot of nerve, indeed. I made a number of arguments in my email about your conclusions (supported I might add by no arguments) and you respond by pointing me to "how to post email" URL's. Your arrogance surely exceeds your intelligence. -- David Clark - Original Message - From: "Eugen Leitl" <[EMAIL PROTECTED]> To: Sent: Wednesday, March 21, 2007 2:04 PM Subject: Re: [agi] Fwd: Numenta Newsletter: March 20, 2007 > On Wed, Mar 21, 2007 at 10:47:35AM -0700, David Clark wrote: > > > In my previous email, I mistakenly edited out the part from Yan King Yin and > > it looks like the "We know that logic is easy" was attributed to him when it > > was actually a quote of Eugen Leitl. > > > > Sorry for my mistake. > > It's not your mistake. It's the mistake of those who choose to ignore > > http://www.netmeister.org/news/learn2quote.html > > It is really a great idea to use plaintext posting and set standard > quoting in your MUA. For those with braindamaged MUAs there are > workarounds like > > http://home.in.tum.de/~jain/software/outlook-quotefix/ > > -- > Eugen* Leitl http://leitl.org";>leitl http://leitl.org > __ > ICBM: 48.07100, 11.36820http://www.ativel.com > 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE > > - > 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/?list_id=303 - 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/?list_id=303
[agi] META: People ... be nice, please! [list moderation action]
Hmmm. I am rarely inspired to an official "moderation" action [though recently some people moderated **me** due to my overly sarcastic humor regarding AIXI] But, comments like " Your arrogance surely exceeds your intelligence. " should be avoided on this list Let's keep things civil! Thanks Ben Goertzel (list owner/ moderator) David Clark wrote: I put up with 1 person out of all the thousands of emails I get who insisted on sending standard text messages as a attachment. Because of virus infections, I had normally set all emails with attachments to automatically get put in the garbage can. I had to stop that so I could read your emails for the past 2 years. You have a lot of nerve, indeed. I made a number of arguments in my email about your conclusions (supported I might add by no arguments) and you respond by pointing me to "how to post email" URL's. Your arrogance surely exceeds your intelligence. -- David Clark - Original Message - From: "Eugen Leitl" <[EMAIL PROTECTED]> To: Sent: Wednesday, March 21, 2007 2:04 PM Subject: Re: [agi] Fwd: Numenta Newsletter: March 20, 2007 On Wed, Mar 21, 2007 at 10:47:35AM -0700, David Clark wrote: In my previous email, I mistakenly edited out the part from Yan King Yin and it looks like the "We know that logic is easy" was attributed to him when it was actually a quote of Eugen Leitl. Sorry for my mistake. It's not your mistake. It's the mistake of those who choose to ignore http://www.netmeister.org/news/learn2quote.html It is really a great idea to use plaintext posting and set standard quoting in your MUA. For those with braindamaged MUAs there are workarounds like http://home.in.tum.de/~jain/software/outlook-quotefix/ -- Eugen* Leitl http://leitl.org";>leitl http://leitl.org __ ICBM: 48.07100, 11.36820http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE - 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/?list_id=303 - 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/?list_id=303 - 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/?list_id=303