That's the base assumption for my current project. I'm starting from human language and attempting to derive an internal representation that corresponds roughly to that used by humans. It is my hope that once an appropriate model of how humans internally represent knowledge is available, the actual mental computations we perform to handle higher-level rational thought should become much more amenable to understanding and analysis. This seems self evident to me, since clearly human beings have solved the GI problem, which means we probably have some sort of internal representation that makes the sorts of representational gymnastics that are necessary for GI much simpler to perform. This approach has the advantage, as well, that its accuracy as a model of human internal knowledge representation should be directly verifiable in experiments with human subjects.
The idea is that I can run a parser on a piece of natural language, extract out the relationships between the words as a semantic net, convert that format into another semantic net that represents the meaning of the language sample, and then reverse the flow back to natural language which is identical in meaning but may be stated differently. When the meaning has been extracted and represented using the internal format, it can be linked up with other semantic nets that represent the meanings of other statements/questions. This combined net in turn can be analyzed directly as a collection of logical predicates and queries in which the binding of two symbols (word/phrase occurrences) to a common referent are directly represented as links from those symbols to that referent's node. New statements/queries can be generated via inference rules and other daemons, and then converted to natural language using the parser, etc. in reverse. I have already built a small system as a proof of concept with Boolean links -- either a semantic link exists or doesn't, rather than allowing links to have real-valued strengths -- and was able to resolve anaphora moderately well, given its toy nature. Since this initial implementation left me unsatisfied with how uncertainty was handled, I'm working now on rebuilding the system using real-valued links that represent probability/uncertainty, similarly to the truth values used by the term logic-based inference system of NARS ( https://sites.google.com/site/narswang/). Adding in the ability to represent uncertainty will allow the system to compare alternatives and choose the most salient anaphoric referent *that makes sense*, taking advantage of knowledge the system has already acquired to determine what makes sense, rather than just taking the most salient/obvious choice in terms of raw language structure independent of conceptual context. On Mon, Aug 27, 2012 at 4:55 PM, Anastasios Tsiolakidis < [email protected]> wrote: > As I started reading I thought to myself "I told you 1000 times, it > depends on the criteria". Reading on, I saw that it is precisely the > criteria you use as a parameter. Well, I'd like to find out the > programming language that makes the most money while giving > immortality :) A little more seriously, if the criteria are cognitive, > as they often are in the real world, you'd be digging yourself a hole > too deep to get out of. On the other hand, if the criteria are > domain-specific, relating to well-behaved domains, I am afraid we are > heading towards tautologies and trivialities. Something like > Mathematica would be optimal for algebra, analysis, gravity, mechanics > etc (though what about instead of calculating a parachute drop > actually measure a real parachute drop), for economics, psychology, > necromancy most things would do equally badly, and for AGI all options > have so far being worse than bad. Mind you, I am in the process of > defining an AGI architecture not as a compression problem but as a > distributed computation problem, and I would challenge you to answer > the question: > > Which programming language/mechanism would be ideal for calculating X > as quickly as possible. > > where X, for the sake of argument, is just a/any "heavy calculation" > without necessarily any of the anomalies of chaotic behavior, pi's > infinite series etc. It is not that I expect intelligence to arise out > of PDEs and integrals, rather I am asking which is the "perfect" > distributed system for calculus, as I am expecting your answer to take > the form of multipliers and other exotic units all converging in an > addition pipeline. I still can't help thinking that the fastest way > for parallel computations is the actual experiment, after all we have > the 3/n body problem and a ton of mathematics OR just an experiment > with n bodies in a field. > > With regards to a possible language for AGI, I don't see how you can > do much better than a human language. Never mind Turing completeness, > we have GI completeness here (except for that part of human language, > perhaps 100% of it, that gets its meaning from its grounding, its > grounding from its embodiment, and its embodiment from - god?) > > AT > > On Mon, Aug 27, 2012 at 10:44 PM, Russell Wallace > <[email protected]> wrote: > > On Mon, Aug 27, 2012 at 9:12 PM, Ben Goertzel <[email protected]> wrote: > >> For domains in which one is concerned with recognizing large ensembles > >> of weak patterns, the language one uses to represent patterns can make > >> a big difference... > >> > >> Image analysis, genetic data analysis and financial prediction are > >> contexts in which I've found this to be the case > >> > >> In these settings, if one does pattern recognition via automated > >> program learning with an Occam bias, > >> the underlying language relative to which the Occam bias is expressed > >> makes a big difference... > > > > Absolutely, but these overheads are not constants - the computational > > cost of a poor choice of representation language is typically > > exponential. > > > >> From a different direction, consider Hutter's proof that AIXI-tl is as > >> good as any other reinforcement learning system ... up to an arbitrary > >> constant. > > > > Well, much violence is being done to the word 'constant' in this case. > > Sure, f(N) is constant for a given N, but... :) > > > > > > ------------------------------------------- > > AGI > > Archives: https://www.listbox.com/member/archive/303/=now > > RSS Feed: > https://www.listbox.com/member/archive/rss/303/14050631-7d925eb1 > > Modify Your Subscription: https://www.listbox.com/member/?&; > > Powered by Listbox: http://www.listbox.com > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/23050605-bcb45fb4 > Modify Your Subscription: > https://www.listbox.com/member/?&; > Powered by Listbox: http://www.listbox.com > ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
