I'm writing as one who was involved in what are now considered the early days of AI but who has been away from the field for quite some time. So I can't claim to be up to date on all the research and commercial activities in field; nevertheless I follow it to some extent and hope to offer some constructive comments.
I am encouraged by your work and by other attempts to open up the science and try to find common goals and roadmaps and, we all hope, common vehicles for carrying out the research required to build a generally intelligent machine.
The need to define the behaviors which we call intelligent is clear, as well as behaviors which characterize intermediate steps. To have those definitions be accepted by a broad range of researchers is also necessary. Lack of this agreement has been seriously hampering the field from its very beginning.
The Turing test, having undergone much criticism over the years, nevertheless displays an important property: it defines a measure of success based on ideas completely divorced from those ideas' implementation. In that sense, if a machine could be constructed which contained a canned set of all possible intelligent responses to stimuli, then despite the idea that we as engineers and scientists would probably not consider it successful, the independent test must nevertheless judge it so. As an example, consider if we narrowed the field down to only chess. I doubt any AI researcher would agree that the current programs which beat the best humans embody true AI. Yet, viewed behaviorally, they do just that -- and much of the world believes it, of course.
I believe that this purely behavioral aspect of an AGI test must remain a central defining property. But clearly, it's not enough. We in addition need to account for aspects of the implementation, and in general terms this can only be described by efficiency considerations, possibly by evoking ideas from algorithmic complexity and similar subfields. If those ideas can in turn be described by externally measurable quantities, then that's a major step toward agreeing on testable AGI. For example, if a learning model takes linear time and logarithmic space, then perhaps that could be considered efficient, and an actual measure of that would pass the test. However, if a learning model took linear time and space (as perhaps in a "pure memorization" model), then that would fail. Similar measures could be devised to weed our "brute force" models from ones considered tractable.
This aspect of the definitions requires great care and will be the most controversial part of the work: The "Enhanced Turing Test", or whatever is devised, will be easy by comparison.
What is hoped by your efforts is that researchers will not only find common ways to test their work, but will gravitate toward common ways to work as well, resulting in a necessary pooling of resources to promote the success of the field. Many years ago I considered these ideas (in unpublished writings), asking whether AI should become a "big science" as had physics, and now biology and astronomy. It seems clear that AI -- especially AGI -- needs a large-scale approach, the problems are simply so vast. This is true even if single machines a few years from now have the putative power to be intelligent on their own. It will simply take a lot of resources and large-scale experiments to boil the algorithms and data down to something which can then be confined to a "small" physical device.
Today, it is encouraging to see efforts along these lines: "standard" KB models and ontologies, the Semantic Web, Open Cog, Open Mind (MIT), and so on (I'm sure I don't know them all). These efforts are lacking three things: common goals, common language, and funding.
Physics, Biology, Chemistry, and other scientific fields enjoy funding considerably above what AI gets. Ironically, they get more computers as well. The nature of the computational facilities devoted to these fields is astonishing, and is moving the whole field of computer, network, and storage engineering forward (not that your average home user will see a dedicated WDM link any time soon). In addition, researchers in these fields can globally share much of this power and the resulting data. These fields have a remarkable coherence when it comes to organizing and standardizing their work, despite what one reads about fragmentation in their disciplines [1]. You want fragmentation? Come to AI. That doesn't seem to have changed since I was involved.
For AGI to proceed, it needs shared facilities proportionally on par with these other scientific endeavors. Also, unlike these fields, even your work will not establish a standard basis, like the Standard Model drives physics or genome and protein work drive biology. I expect your work will develop several promising areas -- and they may well be subfields, like speech or vision -- which need to be nurtured and synthesized into a whole. If you can get the majority of AI researchers to agree on, say, ten primary approaches, you will have succeeded well.
Consider, for example, two approaches, one evolutionary, where we might try to devise general learning algorithms using genetic algorithms, and another, as exemplified by Cyc, where we encode explicit knowledge in logical form over the long term. Both of these will take considerable computational resources. In fact, one might expect the evolutionary algorithm easily to consume hundreds of machines and terabytes of space, to develop populations of meaningful size. The Cyc-like approach could just require a central machine with many processors and considerable fast storage -- but lots of parallelism could move things along more quickly if a connectionist-logic inference style were used.
Both of these approaches will benefit from representational standards, in that each should be able to share results with the other. But each will also benefit from pooled resources and the economy of scale which those resources will provide, and from the ability to measure the degree of success based on common definitions of what success is. This is exactly analogous to the state of physics today: Yes, there are controversies and debates, but general agreement that, for instance, seeing a Higgs Boson emerge from the LHC will be a great advance -- and just how to interpret the measurements which will confirm that emergence.
And I mentioned just two approaches. There are many others, of course, which need to be sorted through. These will need common measurements, models, and computational facilities. Even just getting a good pool of systems to work with will be an advance, as several promising areas will need to continue separately before being brought into the "standard model". You could make a good start by having a common "AI Cloud" or "AI Grid" to share.
In summary, in addition to encouraging your attempt to bring more unification to the field, I believe the field is ready for -- and needs -- a Big Science approach to AI. It is long overdue.
--Lstabile 19:14, 15 March 2009 (UTC) (Larry Stabile l.stabile@comcast.net)
[1] See, for example, "The Trouble with Physics", by Lee Smolin. It's a very readable work on the current debates in the field with respect to String Theory.
One suggestion:
Having made the decision to use the physical world (or a simulation thereof) as the basis for a road to AGI, I think it is very important to stick to that in a principled way, despite inconveniences. Specifically, interaction with a simulated world should only be done via sensory and motor systems that are "as close as possible" to human or robotic systems -- as much simulation detail as possible and inputs that consist of simulated retinas or cameras, simulated haptic and proprioceptive sensors, and so on.
Sensory modalities that skip steps -- such as direct access to X,Y,Z world location or "visual" input pre-chewed into polygons -- incorporate gravely risky assumptions about mental function and allow dangerous illusions about system capability.
The roots of core philosophical issues like symbolization, reference, and uncertainty come into play at the very first stages of the vast rich interface between mind and universe, and must not be abstracted away.
Doubtless somebody will say something like: "If I'm just going to parse a scene into the same polygonal representation as was used in the simulation model in the first place, why bother with intervening graphics and low-level vision processes?" I can only hope that the assumptions embedded in such an argument are plain and self-evidently unhealthy.
THat concern aside, this is a great effort! If standards like this can be developed in some detail (especially if they are not too obviously designed to highlight the capabilities of pre-existing systems) it could provide a very helpful shared context for many research projects.
-- DerekZahn@msn.com