Re: [agi] knowledge representation, Cyc
Hi Ben,
I understand the current situation with Novamente. It seems that one
fundamental difference between Cyc and Novamente is that Cyc is focused on the
linguistic / symbolic level whereas Novamente is focused on sensory /
experiential learning.
My current intuition is that Cyc's route may achieve "a certain level of
intelligence" *sooner*. (Although the work done with sensory-based AGI would
probably still be useful.) This may sound kind of vague, but my intuition is
that if we invest on a Cyc-like AGI for 5 years, it may be able to converse
with humans in a natural language and answer some commonsense queries (which,
the current Cyc actually is somewhat capable of). But if you invest 5 years in
a sensory-based AGI, the resulting AGI baby may be still at the level of a 3-5
years old human. It seems that much of your work may be wasted on dealing with
sensory processing and experiential learning, the latter is particularly
inefficient.
The Cyc route actually bypasses experiential learning because it allows us to
directly enter commonsense knowledge into its KB. That is perhaps the most
significant difference between these 2 approaches.
YKY
My first thought and gut-reaction was "Yeah" record all knowledge about all
objects and things, and then you can build an AI on top of that. The problem
is you need a perfectly compiled, ready to use DB of information.
How do you tell its perfect? Use it, test it, in all situations. This is
where Cyc fails most horribly unfortunatly :{
There is no real way to test all information in all situations, so any amount
of Cyc's information is incorrect for some usage, so when the AI tries it it
fails and must correct the information in some way.
Now, I think we need a small core of knowledge, and abilities, wherin we can
get a AI up on its feet and able to interact as soon as possible. Does it know
everything, can can it do everything, no, but if you have a hosue robot, it
really doesnt need to know about Paris, and cities in France, until it has
reason to know it.
If it has a rich ability to interact with its users, to ask questions, to get
data from a simple source like Wikipedia or direct experience, then it has the
ability of humans to learn.
James Ratcliff
___
James Ratcliff - http://falazar.com
Looking for something...
-
Ahhh...imagining that irresistible "new car" smell?
Check outnew cars at Yahoo! Autos.
-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
Re: [agi] knowledge representation, Cyc
YKY, Cyc has been around a long time with a large amount of financial, computational and human resources invested into it. Why do you think it will succeed in the next 5 years when it hasn't for the last 20+? What novel ideas do you intend to introduce into Cyc that will make it suddenly begin to think and understand? You say The Cyc route actually bypasses experiential learning because it allows us to directly enter commonsense knowledge into its KB. That is perhaps the most significant difference between these 2 approaches. In fact, you can directly enter knowledge into Novamente in logic form, just like Cyc. We could load the Cyc KB into Novamente next week if we wanted to. The problem is that neither Cyc, nor NM, nor any other system is going to be able to do any interesting learning and thinking based solely on this kind of formal, abstracted "quasi common sense" knowledge. I think this is rather amply demonstrated by the long and profoundly uninspiring history of Cyc and related systems. I conjecture that for an AI system to make good use of a Cyc type KB, it must have a reasonable level of experiential grounding for many of the concepts in the KB. -- Ben On 4/12/07, YKY (Yan King Yin) <[EMAIL PROTECTED]> wrote: On 4/6/07, Benjamin Goertzel <[EMAIL PROTECTED]> wrote: > > Ben: Are you interested in translating LRRH into Novamente's KR, as a demo? > > Not really... > > Here's the thing: Novamente's KR is very flexible... > > So, one could translate LRRH into Novamente-ese in a way that would sorta resemble "Cyc plus probabilities, with some higher-order functions and pattern-intensities too" > > But, that wouldn't be likely to closely resemble the way LRRH would wind up being represented in the mind of a Novamente instance that really understood the story. > > So the exercise of explicitly writing LRRH in Novamente's KR would likely wind up being not only pointless, but actively misleading ;-) > > While I do think that a probabilistic logic based KR (as NM uses) is a good choice, I don't think that the compact logical representation a human would use to explicitly represent a story like LRRH, is really the right kind of representation for deep internal use by an AGI system. An AGI's internal representation of a story like this may be logical in form, but is going to consist of a very large number of uncertain, contextual relationships, along with some of the crisper and more encapsulated ones like those a human would formulate if carrying out the exercise of encoding LRRH in logic. > > It is for this reason, among others, that I find Cyc-type AI systems a bit misguided > (another main reason is their lack of effective learning algorithms; and then there's the fact that the absence of perceptual-motor grounding makes it difficult for a useful self-model to emerge; etc. etc.) Hi Ben, I understand the current situation with Novamente. It seems that one fundamental difference between Cyc and Novamente is that Cyc is focused on the linguistic / symbolic level whereas Novamente is focused on sensory / experiential learning. My current intuition is that Cyc's route may achieve "a certain level of intelligence" *sooner*. (Although the work done with sensory-based AGI would probably still be useful.) This may sound kind of vague, but my intuition is that if we invest on a Cyc-like AGI for 5 years, it may be able to converse with humans in a natural language and answer some commonsense queries (which, the current Cyc actually is somewhat capable of). But if you invest 5 years in a sensory-based AGI, the resulting AGI baby may be still at the level of a 3-5 years old human. It seems that much of your work may be wasted on dealing with sensory processing and experiential learning, the latter is particularly inefficient. The Cyc route actually bypasses experiential learning because it allows us to directly enter commonsense knowledge into its KB. That is perhaps the most significant difference between these 2 approaches. 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/?&; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
Re: [agi] knowledge representation, Cyc
On 4/6/07, Benjamin Goertzel <[EMAIL PROTECTED]> wrote: > Ben: Are you interested in translating LRRH into Novamente's KR, as a demo? Not really... Here's the thing: Novamente's KR is very flexible... So, one could translate LRRH into Novamente-ese in a way that would sorta resemble "Cyc plus probabilities, with some higher-order functions and pattern-intensities too" But, that wouldn't be likely to closely resemble the way LRRH would wind up being represented in the mind of a Novamente instance that really understood the story. So the exercise of explicitly writing LRRH in Novamente's KR would likely wind up being not only pointless, but actively misleading ;-) While I do think that a probabilistic logic based KR (as NM uses) is a good choice, I don't think that the compact logical representation a human would use to explicitly represent a story like LRRH, is really the right kind of representation for deep internal use by an AGI system. An AGI's internal representation of a story like this may be logical in form, but is going to consist of a very large number of uncertain, contextual relationships, along with some of the crisper and more encapsulated ones like those a human would formulate if carrying out the exercise of encoding LRRH in logic. It is for this reason, among others, that I find Cyc-type AI systems a bit misguided (another main reason is their lack of effective learning algorithms; and then there's the fact that the absence of perceptual-motor grounding makes it difficult for a useful self-model to emerge; etc. etc.) Hi Ben, I understand the current situation with Novamente. It seems that one fundamental difference between Cyc and Novamente is that Cyc is focused on the linguistic / symbolic level whereas Novamente is focused on sensory / experiential learning. My current intuition is that Cyc's route may achieve "a certain level of intelligence" *sooner*. (Although the work done with sensory-based AGI would probably still be useful.) This may sound kind of vague, but my intuition is that if we invest on a Cyc-like AGI for 5 years, it may be able to converse with humans in a natural language and answer some commonsense queries (which, the current Cyc actually is somewhat capable of). But if you invest 5 years in a sensory-based AGI, the resulting AGI baby may be still at the level of a 3-5 years old human. It seems that much of your work may be wasted on dealing with sensory processing and experiential learning, the latter is particularly inefficient. The Cyc route actually bypasses experiential learning because it allows us to directly enter commonsense knowledge into its KB. That is perhaps the most significant difference between these 2 approaches. 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/?member_id=231415&user_secret=fabd7936
Re: [agi] knowledge representation, Cyc
Ben: Are you interested in translating LRRH into Novamente's KR, as a demo? Not really... Here's the thing: Novamente's KR is very flexible... So, one could translate LRRH into Novamente-ese in a way that would sorta resemble "Cyc plus probabilities, with some higher-order functions and pattern-intensities too" But, that wouldn't be likely to closely resemble the way LRRH would wind up being represented in the mind of a Novamente instance that really understood the story. So the exercise of explicitly writing LRRH in Novamente's KR would likely wind up being not only pointless, but actively misleading ;-) While I do think that a probabilistic logic based KR (as NM uses) is a good choice, I don't think that the compact logical representation a human would use to explicitly represent a story like LRRH, is really the right kind of representation for deep internal use by an AGI system. An AGI's internal representation of a story like this may be logical in form, but is going to consist of a very large number of uncertain, contextual relationships, along with some of the crisper and more encapsulated ones like those a human would formulate if carrying out the exercise of encoding LRRH in logic. It is for this reason, among others, that I find Cyc-type AI systems a bit misguided (another main reason is their lack of effective learning algorithms; and then there's the fact that the absence of perceptual-motor grounding makes it difficult for a useful self-model to emerge; etc. etc.) -- Ben G -- 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] knowledge representation, Cyc
On 3/29/07, Matt Mahoney <[EMAIL PROTECTED]> wrote: Of course, that is a good way to learn Cyc. My concern is that Cyc has been trying for years without success to build a natural language interface. There is a huge mismatch between structured knowledge representations like Cyc and the way that children actually learn language. Huger than huge. Cyc, like a lot of AI, is designed following the idea that you can design a knowledge representation for your data, and then you can design an architecture that says how to process that representation. In the human brain, the representation IS the architecture. The representation of a concept like "hood" is a pattern of activation in neurons in different regions of the brain. The set of neurons that are activated to represent "hood" is a function of how these regions are connected together, which ALSO determines the order of events when neurons are activated and ideas are processed. Architecture and representation are identical. - Phil Goetz - 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] knowledge representation, Cyc
--- "YKY (Yan King Yin)" <[EMAIL PROTECTED]> wrote: > On 3/30/07, Matt Mahoney <[EMAIL PROTECTED]> wrote: > > Wouldn't it save time in the long run to build a system that could > translate > > English into your KR? > > Yes, that's the goal. I'm just doing a human translation of the first > paragraph or so, to get the feel of CycL. Of course, that is a good way to learn Cyc. My concern is that Cyc has been trying for years without success to build a natural language interface. There is a huge mismatch between structured knowledge representations like Cyc and the way that children actually learn language. > It can also be compared with Novamente's version. I think world-wide there > are only about 3-5 KR schemes capable of representing such stories > adequately. Would one of these representations be the original English text? -- Matt Mahoney, [EMAIL PROTECTED] - 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] knowledge representation, Cyc
>> It can also be compared with Novamente's version. In one respect, Novamente's version is analogous to assembly language. It's at a very detailed level and can represent anything but you're going to spend a lot of time and verbosity doing it. In another respect, Novamente's version can't represent anything at all without either a) a standardized dictionary that is shared between KRs or b) a ton of rules like CycL has that are actually definitions in disguise (i.e. a dictionary). A fully loaded Novamente with a dictionary and Cyc's rules would be awesome and you could probably do close to an auto-translation -- but the question is how you *consistently* load Novamente to that point. My personal opinion and approach is that Novamente gives you too much flexibility and not enough structure (although, of course, the flexibility allows you to build the structure if you are so inclined). Your mileage may vary since the proper point on the flexibility vs. structure trade-off spectrum is a serious guess on anyone's part and I'd value Ben's guess over anyone else's (except my own :-). Anyways, my point is that you're not going to be (effectively) translating Little Red Riding Hood into Novamente anytime in the near future. >> I think world-wide there are only about 3-5 KR schemes capable of >> representing such stories adequately. Could you name which ones you believe ARE capable? (and I would love to see your translation of LRR into CycL -- or any other KR) - Original Message - From: YKY (Yan King Yin) To: agi@v2.listbox.com Sent: Thursday, March 29, 2007 4:35 PM Subject: Re: [agi] knowledge representation, Cyc On 3/30/07, Matt Mahoney <[EMAIL PROTECTED]> wrote: > Wouldn't it save time in the long run to build a system that could translate > English into your KR? Yes, that's the goal. I'm just doing a human translation of the first paragraph or so, to get the feel of CycL. It can also be compared with Novamente's version. I think world-wide there are only about 3-5 KR schemes capable of representing such stories adequately. 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 - 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] knowledge representation, Cyc
On 3/30/07, Matt Mahoney <[EMAIL PROTECTED]> wrote: Wouldn't it save time in the long run to build a system that could translate English into your KR? Yes, that's the goal. I'm just doing a human translation of the first paragraph or so, to get the feel of CycL. It can also be compared with Novamente's version. I think world-wide there are only about 3-5 KR schemes capable of representing such stories adequately. 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] knowledge representation, Cyc
--- "YKY (Yan King Yin)" <[EMAIL PROTECTED]> wrote: > I just talked to some Cyc folks, and they assured me that CycL is adequate > to represent entire stories like Little Red Riding Hood. > > The AGI framework has to operate on a knowledge representation language, and > building that language is not a programming task, rather a ontology > engineering task, which I'm not very familiar with. I guess we should not > underestimate the amount of work required for the KR scheme. If we use CycL > we may save a lot of time. > > I may try to translate LRRH into CycL to see if it is too cumbersome or > what. Wouldn't it save time in the long run to build a system that could translate English into your KR? -- Matt Mahoney, [EMAIL PROTECTED] - 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] knowledge representation, Cyc
I just talked to some Cyc folks, and they assured me that CycL is adequate to represent entire stories like Little Red Riding Hood. The AGI framework has to operate on a knowledge representation language, and building that language is not a programming task, rather a ontology engineering task, which I'm not very familiar with. I guess we should not underestimate the amount of work required for the KR scheme. If we use CycL we may save a lot of time. I may try to translate LRRH into CycL to see if it is too cumbersome or what. Ben: Are you interested in translating LRRH into Novamente's KR, as a demo? 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
