RE: [agi] Intelligence by definition
Hi, You make a lot of different points... I'll just grab one for the moment *** I don't believe in combining different methods because cognition deals with the unknown, - we can't a priori split it into different areas, except to the extent that they're sensor/hardware specific, or levels, except that syntactic complexity of inputs should be sequentially increased. *** I like to distinguish between *functional specialization* and *integrated cognition* Novamente (my own AI system) has a mix of cognitive algorithms, which work together to provide overall cognitive functionality. The exact mixture of algorithms is determined by a bunch of parameters. This is one example of "integrated cognition". Functional specialization has to do with there being modules of an intelligent system devoted to particular areas like language processing, vision processing, social interaction, etc. In the Novamente design, each functionally specialized lobe has its own parameter values which determine the specific mix of cognitive algorithms operating within it. (We haven't gotten to experimenting with this yet, now we're just experimenting with mixing cognitive algorithms.) Generally, a mixture of cognitive algorithms is just as capable of dealing with the unknown as a single cognitive algorithm. Sometimes more so On the other hand, functional specialization biases one's system to deal with some parts of the space of the unknown better than others. This is a plus and a minus, obviously. Human cognition deals with the truly unknown very slowly and awkwardly. The human brain is specialized not only based on its sensors and actuators, but also for linguistic processing, social interaction, temporal event processing, etc. etc. etc. This means that it would not work as well taken outside of its ordinary social and physical situations. But it means that its limited resources are generally well deployed within its usual environments. -- Ben Goertzel.
Re: [agi] Intelligence by definition
Thanks for your comments! I like to distinguish between *functional specialization* and *integrated cognition* Novamente (my own AI system) has a mix of cognitive algorithms, which work together to provide overall cognitive functionality. The exact mixture of algorithms is determined by a bunch of parameters. This is one example of "integrated cognition". Functional specialization has to do with there being modules of an intelligent system devoted to particular areas like language processing, vision processing, social interaction, etc. It seems to me that conceptual difference between vision & language is in the level of generalization, aside from different sensor/actuator orientation. Social Interaction? Once you start coding things that are learnable, where do you stop before ending up with just another expert system? Isn't this all about scalable learning, which should develop environmentally specific functional specialization on it's own? In the Novamente design, each functionally specialized lobe has its own parameter values which determine the specific mix of cognitive algorithms operating within it. (We haven't gotten to experimenting with this yet, now we're just experimenting with mixing cognitive algorithms.) Generally, a mixture of cognitive algorithms is just as capable of dealing with the unknown as a single cognitive algorithm. Sometimes more so What single algorithm? How do you evaluate 'dealing'? How do you derive/select you algorithms for unknown inputs without first quantitatively defining your objectives? You definition of intelligence doesn't seem to be functional to me, goals can't be defined solely by their complexity. Without deductive derivation we are stuck with trial & error, which can take millenia. On the other hand, functional specialization biases one's system to deal with some parts of the space of the unknown better than others. This is a plus and a minus, obviously. Human cognition deals with the truly unknown very slowly and awkwardly. I mean 'unknown' not to the cognitive system but to it's designer. Also, the reason human learning is so slow is 'hardware' - specific: it takes a lot longer to build new connections than to access them. That's not the case for computer hardware. The human brain is specialized not only based on its sensors and actuators, but also for linguistic processing, social interaction, temporal event processing, etc. etc. etc. This means that it would not work as well taken outside of its ordinary social and physical situations. But it means that its limited resources are generally well deployed within its usual environments. That's true, but human brain is an accident of incremental & obviously unfinished evolution, not some grand design. Besides, I think to some extent these different areas are specialized not so much by genetic design but by the impact of the input types they recieve. In any case, you must admit, this stone age 'design' doesn't perform very well now & it will get worse as the changes accelerate. Regards! Boris.
RE: [agi] Intelligence by definition
hi, *** It seems to me that conceptual difference between vision & language is in the level of generalization, aside from different sensor/actuator orientation. Social Interaction? Once you start coding things that are learnable, where do you stop before ending up with just another expert system?Isn't this all about scalable learning, which should develop environmentally specific functional specialization on it's own? *** Well, in Novamente we are not coding *specific knowledge* that is learnable... but we are coding implicit knowledge as to what sorts of learning processes are most useful in which specialized subdomains... *** What single algorithm? How do you evaluate 'dealing'? How do you derive/select you algorithms for unknown inputs without first quantitatively defining your objectives? You definition of intelligence doesn't seem to be functional to me, goals can't be defined solely by their complexity. Without deductive derivation we are stuck with trial & error, which can take millenia. *** The Novamente design is mathematically formulated, but not mathematically derived. That is, individual formulas used in the system are mathematically derived, but the system as a whole has been designed by intuition (based on integrating a lot of different ideas from a lot of different domains) rather than by formal derivation. In my view, we are nowhere near possessing the right kind of math to derive a realistic AI design from definitions in a rigorous way. Juergen Schmidhuber's OOPS system is an attempt in this direction, but though I like Juergen's work, I think this design is too simplistic to be a functional AGI. http://www.idsia.ch/~juergen/oops.html Maybe further work in the OOPS direction will yield something like what you're suggesting... *** Also, the reason human learning is so slow is 'hardware' - specific: it takes a lot longer to build new connections than to access them. That's not the case for computer hardware. *** I don't think you're right about the reason human learning is so slow. It is not just hardware inefficiency, it is the fact that a lot of trial-and-error-based algorithms are used in the brain. *** That's true, but human brain is an accident of incremental & obviously unfinished evolution, not some grand design. Besides, I think to some extent these different areas are specialized not so much by genetic design but by the impact of the input types they recieve. In any case, you must admit, this stone age 'design' doesn't perform very well now & it will get worse as the changes accelerate. *** The human brain has many flaws and is not a perfect guide for AGI, but it has far more general intelligence than any existing computer program, and so it is certainly worth carefully studying when designing a would-be AGI system. Novamente is intended to ultimately go beyond what the human brain can accomplish, but for version 1 we'll be contented to achieve human-level general intelligence ;-) -- Ben Goertzel
Re: [agi] Intelligence by definition
>> Well, in Novamente we are not coding *specific knowledge* that is learnable... but we are coding implicit knowledge as to what sorts of learning processes are most useful in which specialized subdomains... << I'm reminded of an AI pioneer who once commented on this same situation - he closed his eyes, pretending that there wasn't a grad student in the room. ===bob briggs
Re: [agi] Intelligence by definition
Well, in Novamente we are not coding *specific knowledge* that is learnable... but we are coding implicit knowledge as to what sorts of learning processes are most useful in which specialized subdomains... *** I don't know, from where I sit this distinction is artificial. Learning is generally defined as projected compression, complexity of methods to achieve it can be sequentially increased as long as it produces positive additional compression minus the expense,- until it matches complexity of the inputs. In other words, optimal methods themselves should be learned. The Novamente design is mathematically formulated, but not mathematically derived. That is, individual formulas used in the system are mathematically derived, but the system as a whole has been designed by intuition (based on integrating a lot of different ideas from a lot of different domains) rather than by formal derivation. In my view, we are nowhere near possessing the right kind of math to derive a realistic AI design from definitions in a rigorous way. To select formulas you must have an implicit criterion, why not try to make it explicit? I don't believe we need complex math for AI, complex methods can be universal, - generalization is a reduction. What we need is a an autonomously scalable method. Juergen Schmidhuber's OOPS system is an attempt in this direction, but though I like Juergen's work, I think this design is too simplistic to be a functional AGI. http://www.idsia.ch/~juergen/oops.html Thanks, I am looking at it. I noticed that he starts with a known probability distribution, to me that suggests that the problem is already solved ;-) I don't think you're right about the reason human learning is so slow. It is not just hardware inefficiency, Of course, that's only part of it. it is the fact that a lot of trial-and-error-based algorithms are used in the brain. I call it search. The human brain has many flaws and is not a perfect guide for AGI, but it has far more general intelligence than any existing computer program, and so it is certainly worth carefully studying when designing a would-be AGI system. How about using it? ;-) Boris.
RE: [agi] Intelligence by definition
*** Well, in Novamente we are not coding *specific knowledge* that is learnable... but we are coding implicit knowledge as to what sorts of learning processes are most useful in which specialized subdomains... --- I don't know, from where I sit this distinction is artificial. Learning is generally defined as projected compression, complexity of methods to achieve it can be sequentially increased as long as it produces positive additional compression minus the expense,- until it matches complexity of the inputs. In other words, optimal methods themselves should be learned. *** Yes, if you have a huge amount of space and time resources available, you can start your system with a blank slate -- nothing but a very simple learning algorithm, and let it learn how to learn, learn how to structure its memory, etc. etc. etc. This is pretty much what OOPS does, and what is suggested in Marcus Hutter's related work. It is not a practical approach, in my view. My belief is that, given realistic resource constraints, you can't take such a general approach and have to start off the system with specific learning methods, and even further than that, with a collection of functionally-specialized combinations of learning algorithms. I could be wrong of course but I have seen no evidence to the contrary, so far... *** The Novamente design is mathematically formulated, but not mathematically derived. That is, individual formulas used in the system are mathematically derived, but the system as a whole has been designed by intuition (based on integrating a lot of different ideas from a lot of different domains) rather than by formal derivation. In my view, we are nowhere near possessing the right kind of math to derive a realistic AI design from definitions in a rigorous way. --- To select formulas you must have an implicit criterion, why not try to make it explicit? I don't believe we need complex math for AI, complex methods can be universal, - generalization is a reduction. What we need is a an autonomously scalable method. *** Well, if you know some simple math that is adequate for deriving a practical AI design, please speak up. Point me to the URL where you've posted the paper containing this math! I'll be very curious to read it ;-) Juergen Schmidhuber's OOPS system is an attempt in this direction, but though I like Juergen's work, I think this design is too simplistic to be a functional AGI. http://www.idsia.ch/~juergen/oops.html --- Thanks, I am looking at it. I noticed that he starts with a known probability distribution, to me that suggests that the problem is already solved ;-) He starts with a known pdf for theoretical purposes. He is proving that his system can work effectively for ANY given probability distribution. He is not assuming that his system is somehow fed the pdf in advance. -- Ben
Re: [agi] Intelligence by definition
Yes, if you have a huge amount of space and time resources available, you can start your system with a blank slate -- nothing but a very simple learning algorithm, and let it learn how to learn, learn how to structure its memory, etc. etc. etc. This is pretty much what OOPS does, and what is suggested in Marcus Hutter's related work. It is not a practical approach, in my view. My belief is that, given realistic resource constraints, you can't take such a general approach and have to start off the system with specific learning methods, and even further than that, with a collection of functionally-specialized combinations of learning algorithms. I could be wrong of course but I have seen no evidence to the contrary, so far... *** A fixed collection of methods won't scale, - power of a method should correspond to generality (predictive power) of a pattern. The whole point of such pattern-specific & level-specific scaling of methods IS computational efficiency, - it's a lot less expensive to incrementally scale methods for individual patterns than to indiscriminately apply a fixed set of them on patterns most of which are either too complex or too simple for any given method. *** To select formulas you must have an implicit criterion, why not try to make it explicit? I don't believe we need complex math for AI, complex methods can Sorry, that was a typo, it should be "can't" be universal, - generalization is a reduction. What we need is a an autonomously scalable method. *** Well, if you know some simple math that is adequate for deriving a practical AI design, please speak up. Point me to the URL where you've posted the paper containing this math! I'll be very curious to read it ;-) *** We both know that there is no practical general AI yet, I'm trying to suggest a theoretically consistent one. Given that the whole endeavor is context-free it should ultimately be the same thing. I don't have any papers, when the theory is finished I'll write a program, not a paper. My method is ultimately simple: sequential expansion of search for correlations of sequentially increasing arithmetic power/derivation, for inputs which had above-average compression over the shorter range of search / lower arithmetic power/derivation. What's new here (correct me if I'm wrong), is how I define compression, which determines value of a pattern, & encode these patterns to preserve restorability & enable analytical comparison (between individual variable types within patterns). Both are necessary to selectively scale the search, & I don't see it in OOPS It's in my introduction, someplace, but I realize it must be mental torture to try to figure it. Why would you work on it? Only if you agree with my theoretical assumptions, I suppose, the method is uniquely consistent with them. Boris.