ontolog-forum
[Top] [All Lists]

Re: [ontolog-forum] Child architecture

To: "ontolog-forum@xxxxxxxxxxxxxxxx" <ontolog-forum@xxxxxxxxxxxxxxxx>
From: Philip Jackson <philipcjacksonjr@xxxxxxxxxxx>
Date: Fri, 19 Dec 2014 12:40:25 -0500
Message-id: <SNT147-W37D76FF4BB538F93989D59C16B0@xxxxxxx>
Hi Rich,
 
You (RC) wrote:
> ...my comments are below, interspersed with yours,
>
I (PJ) wrote:
> > Perhaps I am missing something, but I don't see the
> > creation of new names or new words as being a major
> > problem, of central importance to human-level AI. What is
> > important is the creation of new concepts / conceptual
> > structures that are useful in understanding and interacting
> > with the world. A rose by any other name would smell as
> > sweet...
>
> RC:
> Agreed, name generation is not a *major* problem, but it is
> a *necessary* condition for naming those *concepts*, if
> only in a default setting.
 
OK, at least we agree on this much.
 
> RC:
> The default can then be reduced
> in size (remember the goal was a *succinct* name that the
> concept can be designated and/or invoked with.  The
> original default name, if too long, needs a memorable short
> name.  If there is an algorithm generating short names for
> concepts, how should it proceed to eliminate parts of the
> phrase?
 
Why must a name generation algorithm start with long, multi-word names and then eliminate parts to create shorter names? Why not pick two words from a Tala conceptual structure (e.g. a noun and verb) and use these to create an initial name? If that isn't unique enough, then add another word from the conceptual structure, etc. Or alternatively, add a new sense to a word used in the conceptual structure, and refer to the conceptual structure metonymically -- this is how people frequently create names for  concepts, in natural language, and human-level AI will need to be able to support this, for communication with humans. (viz. p.88)
 
>
> PJ:
> > Creation of new conceptual structures ...
> > should not happen via processes that
> > "concatenate every word of that phrase with every word of
> > some other phrase to create most of the words in a new
> > combined concept, with a few added for whatever reason in
> > the Horne clause form". -- Or perhaps you were describing
> > this as a process for somehow inventing new names for
> > concepts...?
>
> RC:
> That is my intent.
 
OK, thanks for clarifying this was your intent. Again, this would be an inefficient approach for creating names - you indicated as much by writing it could "recurse until overflow"...
 
> RC:
> If there is a Child architecture, and
> if two or more agents must converse to implement the Child,
> then the concept has to be communicable among the agents.
> If the concepts can be generated in unlimited numbers, as
> we would like a Child to do, the algorithm for shortening
> the default name, as the new concepts are developed as
> combinations of the old ones with known names, producing a
> reliably shortened and unique name that can reference the
> concept, its definition(s), and historical examples
> experienced or empathized vividly.
 
For two agents within the society of mind of a single human-level AI Child system, naming concepts can be simpler. In the TalaMind approach, a single human-level AI system is called a Tala agent, and for clarity the agents in its society of mind are "subagents". (p.43) The subagents can share access to a lexicon, and share access to associations between names and conceptual structures. So given a name for a concept, any subagent can access the associated conceptual structure. If the concepts are only used within the Tala agent, then creating new names can be accomplished by creating a unique symbol or pointer used within the Tala agent.
 
Two Tala agents (two different Child AI's) could still share a common lexicon, e.g. available on the web, and communicate using Tala conceptual structures specifying word senses.
 
Disambiguation of word referents is another issue related to sharing concept names and word senses. Within a Tala agent, subagents could share access to referents that exist as conceptual structures within the Tala agent. So, if the subagents are considering a conceptual structure for "The cat is on the mat", they could each access the internal referent to a specific cat being observed by the Tala agent.
 
To some extent this might also be possible for communication between two Tala agents. If the cat has an RFID tag and is registered in the "Internet of Things", then two Tala agents could exchange a conceptual structure for "the cat is on the mat" and share an external reference to the same cat.
 
However, in general for communication between two Tala agents, or between a Tala agent and a human being, the disambiguation problems would be the same as for communication between two human beings. (Viz. section 3.6)
 
Perhaps you are limiting your focus to communication between two AIXItl systems, which can communicate only via a set of signals for reward and punishment (?) -- If so, the problem seems to be much more difficult, due to issues of opacity and efficiency. If I understand correctly, the two systems would need to induce a common language syntax, as well as common meanings for words, working only from signals for reward and punishment exchanged between them (?)... One might argue this is mathematically possible, but it seems likely to be mathematically intractable, and realistically impractical...
 
> PJ:
> > When an agent creates a conceptual structure that needs a
> > new name, not corresponding to some word already known, a
> > new word can be developed, as John Bottoms described. If it
> > is too difficult to find a variation or concatenation of
> > existing words that is appropriate, then as a last resort,
> > a random, nonsense word can be used.  There's a website
> > that strings syllables together to create random new words,
> > at:
> >
> > http://randomwordmachine.com/
> >
> > Phil
>
>
> RC:
> But the problem still remains, in a very complex knowledge
> soup, how do agents communicate, and find a way to name
> their newly learned experiences acquired from other agents
> through that communication?
>
> The problem is easier to state this way.  How do a
> plurality of agents intercommunicate using words and
> phrases to succinctly name any new concepts which the
> agents create through the learning process?
>
> Randomly generated words are easy to do, but that doesn’t
> make them *meaningful* to the other agents.  New concept
> names would have to be a construction of symbols known to
> many of the other agents, so the agents can use these
> experiences as rules of generating messages back and forth,
> including support for the new concept names as they are
> generated.
 
I think I've addressed these questions, above. If there is still a question, I'll try to answer. As I wrote on 12/13, the TalaMind approach is focused on creating concepts meaningfully. It supports creation of words with meanings, and communication of meaningful sentences.
 
Incidentally, when I wrote:
 
> > Creation of new conceptual structures should happen via
> > conceptual processes for reasoning and learning that are
> > efficient and rational, such as those described in my
> > thesis. It should not happen via processes that
> > "concatenate every word of that phrase with every word of
> > some other phrase to create most of the words in a new
> > combined concept, with a few added for whatever reason in
> > the Horne clause form".
 
My intent was just to say the TalaMind conceptual processes should be possible to implement with adequate efficiency for human-level AI. The thesis has disclaimers noting the prototype does not claim to be efficient. Per page 25, developing scalable, efficient conceptual processes is a topic for future research.
 
Phil
 
 
 
 

_________________________________________________________________
Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/  
Config Subscr: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/  
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/ 
To join: http://ontolog.cim3.net/cgi-bin/wiki.pl?WikiHomePage#nid1J    (01)

<Prev in Thread] Current Thread [Next in Thread>