Re: [agi] Enumeration of useful genetic biases for AGI
Ben, I am not sure the question has been stated clearly enough to be answered meaningfully, yet. The list given by your correspondent was extremely vague: what does it mean to talk about an implicit set of constraints on ontologies that can be discovered by systematic 'scientific' investigation? For example, there are things I can only perceive *directly* (whatever that means) if they are in 3-D space, but systematic scientific investigation allows me to think about spaces with other numbers of dimensions, in all kinds of ways. Same goes for causality. Having said that, I know what you mean at an intuitive level (and I do believe there are built in biasses) but I think the problem is deeply tangled up with what you think the machinery is, that is getting biassed. I am not even convinced that the question can be properly asked unless you can talk in terms of that machinery. And what is the boundary between an ontological bias and a lesser tendency to learn a certain kind of thing, which can nevertheless be overridden through experience? Richard Loosemore. Ben Goertzel wrote: Hi, In a recent offlist email dialogue with an AI researcher, he made the following suggestion regarding the inductive bias that DNA supplies to the human brain to aid it in learning: * What is encoded in the DNA may include a starting ontology (as proposed, with exasperating vaguess, by developmental psychologists, though much more complex than anything they have thought of) but the more important thing is an implicit set of constraints on ontologies that can be discovered by systematic 'scientific' investigation. So it might not work in an arbitrary universe, including some simulated universes,e.g. 'tileworld' universes. One such constraint (as Kant pointed out in 1780) is the assumption that everything physical happens in 3-D space and time. Another is the requirement for causal determinism (for most processes). There may also be constraints on kinds of information-processing entities that can be learnt about in the environment, e.g. other humans, other animals, dead-ancestors, gods, spirits, computer games, The major, substantive, ontology extensions have to happen in (partially ordered) stages, each stage building on previous stages, and brain development is staggered accordingly. ** My response to him was that these genetic biases are indeed encoded in the Novamente design, but in a somewhat unsystematic and scattered way. For instance, in the Novamente system, -- the restriction to 3D space is implicit in the set of elementary predicates and procedures supplied to the system for preprocessing perceptual data on its way to abstract cognition -- the bias toward causal determinism is implicit in an inference control mechanism that specifically tries to build PredictiveAttractionLink relationships that embody likely causal relationships etc. I have actually never gone through the design with an eye towards identifying exactly how each important genetic bias of cognition is encoded in the system. However, this would be an interesting and worthwhile thing to do. Toward that end, it would be interesting to have a systematic list somewhere of the genetic biases that are thought to be most important for structuring human cognition. Does anyone know of a well-thought-out list of this sort. Of course I could make one by surveying the cognitive psych literature, but why reinvent the wheel? -- Ben G - 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] Enumeration of useful genetic biases for AGI
On 14/02/07, Ben Goertzel [EMAIL PROTECTED] wrote: Does anyone know of a well-thought-out list of this sort. Of course I could make one by surveying the cognitive psych literature, but why reinvent the wheel? None that I have come across. Biases that I have come across are things like paying attention to face like objects(1) and the on going debate over language centres, that is we are biased to expect language of some variety. These two biases I think are parts of the very important general bias to expect other intelligent agents that we can learn from. Without that starting bias, or the ability to have the general form of that bias (the ability to learn almost arbitrary facts/skills/biases from other agents), I think an AGI is going to be very slow at learning about the world, even if its powers of inference are magnitudes above humans. Will Pearson 1. http://info.anu.edu.au/mac/Media/Research_Review/_articles/_2005/_researchreviewmckone.asp - 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] Enumeration of useful genetic biases for AGI
On Tue, 13 Feb 2007 21:28:53 -0500, Ben Goertzel [EMAIL PROTECTED] wrote: Toward that end, it would be interesting to have a systematic list somewhere of the genetic biases that are thought to be mostimportant for structuring human cognition. Does anyone know of a well-thought-out list of this sort. Of course I could make one by surveying the cognitive psych literature,but why reinvent the wheel? Your email acquaintance mentioned Kant. You may want to look at Kant's categories, in his Critique of Pure Reason. These are the 'Categories of the Understanding' by which Kant thought the mind structures cognition: Quantity *Unity *Plurality *Totality Quality *Reality *Negation *Limitation Relation *Inherence and Subsistence (substance and accident) *Causality and Dependence (cause and effect) *Community (reciprocity) Modality *Possibility *Existence *Necessity -gts - 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]
bcc:[EMAIL PROTECTED] Subject: Re: [agi] Enumeration of useful genetic biases for AGI In-Reply-To: [EMAIL PROTECTED] References: [EMAIL PROTECTED] [EMAIL PROTECTED] X-Mailer: VM 7.17 under 21.4 (patch 19) Constant Variable XEmacs Lucid Reply-To: [EMAIL PROTECTED] --text follows this line-- Ben Matt Mahoney wrote: I don't think there is a simple answer to this problem. We observe very complex behavior in much simpler organisms that lack long term memory or the ability to learn. For example, bees are born knowing how to fly, build hives, gather food, and communicate its location. Ben Indeed, and we observe complex behaviors in turbulent fluid flow, Ben plasmas, and other nonliving self-organizing systems as well Ben But I don't see any of this as terribly relevant to the question Ben I was asking ;-) Ben Bees are born knowing how to build hives, but are children born Ben knowing how to build houses? I have a feeling a human's Ben cognitive architecture and dynamics are quite different from Ben those of a bee... If a bee is born knowing how to build a hive, that implies, I expect, that it is born with a program library containing many objects, classes, methods, etc that would be extremely useful for constructing a program to build houses. Building houses no doubt requires adding a bunch more well-organized code, but that code is likely to be a lot easier to write starting with the bee's library. I expect that code discovery is only possible when it can be broken down into steps each of which is not too large, and starting with the bee library it may be that relatively small steps can take you a long way. So I expect that there are biases that are a lot like a killer object oriented code library. The complexity of inductive bias is bounded by the complexity of your DNA, about 6 x 10^9 bits. This is probably too high by a few orders of magnitude, just as the number of synapses overestimates the complexity of AGI. Nevertheless, we risk repeating the error of GOFAI. Early AI researchers were led astray by the successes of explicitly coding knowledge into toy systems. Now we know to use statistical and machine learning techniques, but we may still be led astray by oversimplified models of inductive bias. Certain aspects of the cerebral cortex are highly uniform, which suggests a simple model. But the rest of the brain has a complex structure that is poorly understood. Ben I'm not thinking that a systematic list of known human inductive Ben biases could be derived from genetics neuroscience (in the near Ben term), but rather from cognitive psychology. In the near term, I am trying introspection. To actually build these biases into a system will, I expect, involve a collaboration of human programmers and evolutionary programming. We also have some windows on these biases from ethology (eg bees, see above), and imaging, etc. But working out the genomics could turn out to be the way that gets the most data the soonest. Ben And, I'm not thinking to use such a list as the basis for Ben creating an AGI, but simply as a tool for assisting in thinking Ben about an already-existing AGI design that was created based on Ben other principles. My suspicion is that all the known and Ben powerful human inductive biases are already built into Novamente Ben in various ways, I much doubt Novamente has the library of procedures that a Bee is born with. - 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]
Eric Baum wrote: Ben And, I'm not thinking to use such a list as the basis for Ben creating an AGI, but simply as a tool for assisting in thinking Ben about an already-existing AGI design that was created based on Ben other principles. My suspicion is that all the known and Ben powerful human inductive biases are already built into Novamente Ben in various ways, I much doubt Novamente has the library of procedures that a Bee is born with. Correct. This is an apparent point of disagreement between us. My own working hypothesis is that the hard-coded inductive biases needed for achieving AGI are at a higher level than, say, specific navigation routines. As two, semi-random examples: We do build in a bias to look for patterns among percepts that appear to originate from physically nearby locations. And we build in a bias toward imitative behavior. And our program learning component has a bias for hierarchical learning that will bias the system toward e.g. learning recursive dynamic-programming-like algorithms for navigation (but is still different than supplying the system with navigation algorithms). I think that the old book Rethinking Innateness http://crl.ucsd.edu/~elman/Papers/book/index.shtml got a lot of things right about the nature/nurture controversy. The genome definitely encodes a lot of biases that direct learning in appropriate directions, but my suspicion is that you overestimate the specificity and concreteness of the genetically inbuilt biases. However, the Novamente architecture does in fact support import of more specific biases and code routines as you suggest. So, if you create them, we can plug them in and see if the Novamente learning mechanisms are able to take them as building-blocks and utilize them effectively. Thus, I believe the Novamente architecture is going to be suitable for experimenting with both of our working hypotheses. -- 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
[agi] hard-coded inductive biases
... various comments ... It more fundamental than that: The design of your 'senses' - what feature extraction, sampling and encoding you provide lays a primary foundation to induction. Peter - 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] hard-coded inductive biases
Peter Voss wrote: ... various comments ... It more fundamental than that: The design of your 'senses' - what feature extraction, sampling and encoding you provide lays a primary foundation to induction. Peter That is definitely true, and is PART of what I meant by saying that the inductive biases of the human mind are largely inbuilt **implicitly** in Novamente. Some are inbuilt implicitly in the design of the perception module... -- 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
[agi] the birth of the mind
http://www.amazon.com/Birth-Mind-Creates-Complexities-Thought/dp/0465044069/sr=8-1/qid=1171483943/ref=pd_bbs_sr_1/105-4534151-3528451?ie=UTF8s=books A good easy account of the developing brain, wherein it is described where the (many) bits missing from the genome come from. Might be of interest to some AGI folks. -- Eugen* Leitl a href=http://leitl.org;leitl/a http://leitl.org __ ICBM: 48.07100, 11.36820http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE - 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 signature.asc Description: Digital signature
Re: [agi]
Ben Eric Baum wrote: Ben I think that the old book Rethinking Innateness Ben http://crl.ucsd.edu/~elman/Papers/book/index.shtml Ben got a lot of things right about the nature/nurture controversy. Ben The genome definitely encodes a lot of biases that direct Ben learning in appropriate directions, but my suspicion is that you Ben overestimate the specificity and concreteness of the genetically Ben inbuilt biases. There are monkeys that are born pre-programmed to learn, from a single episode of seeing a video of another monkey shrieking at a snake, to fear snakes. The monkey will not learn from seeing another monkey shriek at a flower to fear the flower. Nor will the monkey fear snakes if it hasn't previously seen another monkey do so. It also seems clear that monkeys are programmed to learn social interaction routines, because if they miss social input during a critical period in their development, they never develop them. Just as humans are programmed to learn language. Navigation is pretty important to creatures, and its not likely to be easy to build those programs unless there's a lot built in, so you might see evolution having incentive to build in routines. Even very simple creatures do some navigation, so evolution has been perfecting these routines for a long time. Evolution had way more computational power than a creature does during life, so I can't see why if one thinks the creature could learn it during life, one wouldn't think evolution could build it in, which would likely be fitter. I guess you'd credit that the birds, who if they don't see the heavens during the critical period in their development never learn to navigate by the stars, and if they do see them do, are programmed to develop a navigation instinct. I don't see why its surprising if the kind of navigation routines that are useful for playing Sokoban are programmed in as well. But I should clarify-- I don't mean the final routines are explicitly coded in exactly. The genomic code runs, interacts with data in the sensory stream, and produces the mental structures reflecting the routines. That's how it evolves, because as the genome is being mutated, what survives is what works in development which takes place in contact with the sensory stream. If the monkey doesn't see the other monkey shrieking, it won't build the snake fear routine. There will thus be a sense in which what is genomically coded is a bias to develop routines, rather than explicit routines in final form. But as I try to think what kinds of bias I can write down that will be useful, and what kinds accord with introspection, big chunks of code like scaffolds come to mind. Ben However, the Novamente architecture does in fact support import Ben of more specific biases and code routines as you suggest. So, if Ben you create them, we can plug them in and see if the Novamente Ben learning mechanisms are able to take them as building-blocks and Ben utilize them effectively. Ben Thus, I believe the Novamente architecture is going to be Ben suitable for experimenting with both of our working hypotheses. That's good. Ben -- Ben Ben - This list is sponsored by AGIRI: http://www.agiri.org/email Ben To unsubscribe or change your options, please go to: Ben 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] Priors and indefinite probabilities
Tying together recent threads on indefinite probabilities and prior distributions (PI, maxent, Occam)... For those who might not know, the PI (the principle of indifference) advises us, when confronted with n mutually exclusive and exhaustive possibilities, to assign probabilities of 1/n to each of them. In his book _The Algebra of Probable Inference_, R.T. Cox presents a convincing disproof of the PI when n = 2. I'm confident his argument applies for greater values of n, though of course the formalism would be more complicated. His argument is by reductio ad absurdum; Cox shows that the PI leads to an absurdity. (Not just an absurdity in his view, but a monstrous absurdity :-) The following quote is verbatim from his book, except that in the interest of clarity I have used the symbol to mean and instead of the dot used by Cox. The symbol v means or in the sense of and/or. Also there is an axiom used in the argument, referred to as Eq. (2.8 I). That axiom is (a v ~a) b = b. Cox writes, concerning two mutually exclusive and exhaustive propositions a and b... == ...it is supposed that a | a v ~a = 1/2 for arbitrary meanings of a. In disproof of this supposition, let us consider the probability of the conjunction a b on each of the two hypotheses, a v ~a and b v ~b. We have a v b | a v ~a = (a | a v ~a)[b | (a v ~a) a] By Eq (2.8 I) (a v ~a) a = a and therefore a b | a v ~a = (a | a v ~a) (b | a) Similarly a b | b v ~b = (b | b v ~b) (a | b) But, also by Eq. (2.8 I), a v ~a and b v ~b are each equal to (a v ~a) (b v ~b) and each is therefore equal to the other. Thus a b | b v ~b = a b | a v ~a and hence (a | a v ~a) (b | a) = (b | b v ~b) (a | b) If then a | a v ~a and b | b v ~b were each equal to 1/2, it would follow that b | a = a | b for arbitrary meanings of and b. This would be a monstrous conclusion, because b | a and a | b can have any ratio from zero to infinity. Instead of supposing that a | a v ~a = 1/2, we may more reasonably conclude, when the hypothesis is the truism, that all probabilities are entirely undefined except these of the truism itself and its contradictory, the absurdity. This conclusion agrees with common sense and might perhaps have been reached without formal argument, because the knowledge of a probability, though it is knowledge of a particular and limited kind, is still knowledge, and it would be surprising if it could be derived from the truism, which is the expression of complete ignorance, asserting nothing. === -gts - 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]
But I should clarify-- I don't mean the final routines are explicitly coded in exactly. The genomic code runs, interacts with data in the sensory stream, and produces the mental structures reflecting the routines. That's how it evolves, because as the genome is being mutated, what survives is what works in development which takes place in contact with the sensory stream. If the monkey doesn't see the other monkey shrieking, it won't build the snake fear routine. There will thus be a sense in which what is genomically coded is a bias to develop routines, rather than explicit routines in final form. Agreed, yes. This is the main point Elman et al make in their book as well, as you know. But as I try to think what kinds of bias I can write down that will be useful, and what kinds accord with introspection, big chunks of code like scaffolds come to mind. This is where I'm not sure you're right ... I'm not sure the relevant biases are best provided to an AGI system as big chunks of code. For each of your big chunks of code, I might be able to figure out a way to achieve the same bias -- in a more flexible and learning-friendly way -- via subtler mechanisms within the Novamente system (a few parameter tweaks, a few small in-built procedures, etc.) -- 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] Priors and indefinite probabilities
Indeed, that is a cleaner and simpler argument than the various more concrete PI paradoxes... (wine/water, etc.) It seems to show convincingly that the PI cannot be consistently applied across the board, but can be heuristically applied to certain cases but not others as judged contextually appropriate. This of course is one of the historical arguments for the subjective, Bayesian view of statistics; and also for the interval representation of probabilities (so when you don't know what P(A) is, you can just assign it the interval [0,1]) Ben gts wrote: Tying together recent threads on indefinite probabilities and prior distributions (PI, maxent, Occam)... For those who might not know, the PI (the principle of indifference) advises us, when confronted with n mutually exclusive and exhaustive possibilities, to assign probabilities of 1/n to each of them. In his book _The Algebra of Probable Inference_, R.T. Cox presents a convincing disproof of the PI when n = 2. I'm confident his argument applies for greater values of n, though of course the formalism would be more complicated. His argument is by reductio ad absurdum; Cox shows that the PI leads to an absurdity. (Not just an absurdity in his view, but a monstrous absurdity :-) The following quote is verbatim from his book, except that in the interest of clarity I have used the symbol to mean and instead of the dot used by Cox. The symbol v means or in the sense of and/or. Also there is an axiom used in the argument, referred to as Eq. (2.8 I). That axiom is (a v ~a) b = b. Cox writes, concerning two mutually exclusive and exhaustive propositions a and b... == ...it is supposed that a | a v ~a = 1/2 for arbitrary meanings of a. In disproof of this supposition, let us consider the probability of the conjunction a b on each of the two hypotheses, a v ~a and b v ~b. We have a v b | a v ~a = (a | a v ~a)[b | (a v ~a) a] By Eq (2.8 I) (a v ~a) a = a and therefore a b | a v ~a = (a | a v ~a) (b | a) Similarly a b | b v ~b = (b | b v ~b) (a | b) But, also by Eq. (2.8 I), a v ~a and b v ~b are each equal to (a v ~a) (b v ~b) and each is therefore equal to the other. Thus a b | b v ~b = a b | a v ~a and hence (a | a v ~a) (b | a) = (b | b v ~b) (a | b) If then a | a v ~a and b | b v ~b were each equal to 1/2, it would follow that b | a = a | b for arbitrary meanings of and b. This would be a monstrous conclusion, because b | a and a | b can have any ratio from zero to infinity. Instead of supposing that a | a v ~a = 1/2, we may more reasonably conclude, when the hypothesis is the truism, that all probabilities are entirely undefined except these of the truism itself and its contradictory, the absurdity. This conclusion agrees with common sense and might perhaps have been reached without formal argument, because the knowledge of a probability, though it is knowledge of a particular and limited kind, is still knowledge, and it would be surprising if it could be derived from the truism, which is the expression of complete ignorance, asserting nothing. === -gts - 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] Priors and indefinite probabilities
Chuckling that this is still going on, and top posting based on Ben's prior example... Cox's proof is all well and good, but I think gts still misses the point: The principle of indifference is still the *best* one can do under conditions of total ignorance. Any other distribution would imply some latent knowledge. The subtle and deeper point missed by gts, and unacknowledged by Cox, is that while it is logically true you can't get knowledge from ignorance, as a subjective agent within a consistent reality, sometimes you just gotta choose anyway, or you don't get to play the game. LEADING TO THE ONLY THING REALLY INTERESTING ABOUT THIS DISCUSSION: The deeper philosophical point that, as subjective agents, we can't actually ask a fully specified question without having a prior of some kind, and that by playing the game we tend always to move toward a state of less ignorance. The principle of indifference, or as Jaynes put it, equal information yields equal probabilities, is beautiful in its insistence on consistency, and there's an even greater beauty in what it says about our place in the universe. Ben, thanks for your diplomatic acknowledgement of both sides, below. - Jef Ben Goertzel wrote: Indeed, that is a cleaner and simpler argument than the various more concrete PI paradoxes... (wine/water, etc.) It seems to show convincingly that the PI cannot be consistently applied across the board, but can be heuristically applied to certain cases but not others as judged contextually appropriate. This of course is one of the historical arguments for the subjective, Bayesian view of statistics; and also for the interval representation of probabilities (so when you don't know what P(A) is, you can just assign it the interval [0,1]) Ben gts wrote: Tying together recent threads on indefinite probabilities and prior distributions (PI, maxent, Occam)... For those who might not know, the PI (the principle of indifference) advises us, when confronted with n mutually exclusive and exhaustive possibilities, to assign probabilities of 1/n to each of them. In his book _The Algebra of Probable Inference_, R.T. Cox presents a convincing disproof of the PI when n = 2. I'm confident his argument applies for greater values of n, though of course the formalism would be more complicated. His argument is by reductio ad absurdum; Cox shows that the PI leads to an absurdity. (Not just an absurdity in his view, but a monstrous absurdity :-) The following quote is verbatim from his book, except that in the interest of clarity I have used the symbol to mean and instead of the dot used by Cox. The symbol v means or in the sense of and/or. Also there is an axiom used in the argument, referred to as Eq. (2.8 I). That axiom is (a v ~a) b = b. Cox writes, concerning two mutually exclusive and exhaustive propositions a and b... == ...it is supposed that a | a v ~a = 1/2 for arbitrary meanings of a. In disproof of this supposition, let us consider the probability of the conjunction a b on each of the two hypotheses, a v ~a and b v ~b. We have a v b | a v ~a = (a | a v ~a)[b | (a v ~a) a] By Eq (2.8 I) (a v ~a) a = a and therefore a b | a v ~a = (a | a v ~a) (b | a) Similarly a b | b v ~b = (b | b v ~b) (a | b) But, also by Eq. (2.8 I), a v ~a and b v ~b are each equal to (a v ~a) (b v ~b) and each is therefore equal to the other. Thus a b | b v ~b = a b | a v ~a and hence (a | a v ~a) (b | a) = (b | b v ~b) (a | b) If then a | a v ~a and b | b v ~b were each equal to 1/2, it would follow that b | a = a | b for arbitrary meanings of and b. This would be a monstrous conclusion, because b | a and a | b can have any ratio from zero to infinity. Instead of supposing that a | a v ~a = 1/2, we may more reasonably conclude, when the hypothesis is the truism, that all probabilities are entirely undefined except these of the truism itself and its contradictory, the absurdity. This conclusion agrees with common sense and might perhaps have been reached without formal argument, because the knowledge of a probability, though it is knowledge of a particular and limited kind, is still knowledge, and it would be surprising if it could be derived from the truism, which is the expression of complete ignorance, asserting nothing. === -gts - 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
