ld231782@longs.LANCE.ColoState.Edu (Lawrence Detweiler) (03/04/90)
>Note that babies bootstrap, so that "knowing what to look for" becomes >increasingly sophisticated. This deserves a thread of its own. This is what is so remarkable about the human neural net, in my opinion: it doesn't have something feeding it the "right" answers to its neurons. The whole thing is self-contained. The accomplishments of conventional computational neural nets are not so grandiose considering that they require such intensive and explicit training. This is not the tired point that "computational neural nets pale in comparison to the brain" and therefore they can never hope to rival or surpass it (begging the question). Rather, it is suggesting that there may be something fundamentally different about current NNs and the one in the brain. Sure, an artificial NN can "learn" to speak by an EXTERNAL CONTROL that modifies its weights. But the human can do it with an INTERNAL one. Can somebody invent a black box that can learn to speak (and make sense!) solely from interaction with the outside environment? As long as there is anything outside the box that BYPASSES THE INPUT when it modifies weights you don't have a neural network in the most realistic sense. And any network that does is about the same in concept as a little man inside everyone's head. There is something crucial about the roles of interaction and association in learning that computational nets seem to miss. A specious argument is that babies DO have "models", namely their parents. But in the beginning it is all just so much input. Babies must "learn to learn." How are they to know that when mommy says "say mommy" they are to say mommy? Is there some set of "mommy" neurons that are being changed based on how well they say "mommy"? The answer is yes, but whether this is true from the BEGINNING...? Likewise one could argue that in current computational nets there is only input (the stimulus and correct response, so to speak). But the relationship between stimulus and response is firmly established by the EXTERNALLY DERIVED learning rule. In any real brain this is precisely what must be learned! It must be INTERNALLY derived. How does one neuron learn what influences it, what it is modeling? I think McClelland and Rumelhart make an allusion to this (namely the difference between computational NNs and the human one exposed by the "baby bootstrap") in the beginning of the first volume of PDP (of the assumption that neurons are molded by some "correct" guide, "we think this is nearly correct..."). One might argue that the role of heredity is to create a minimal structure that experience can build on (the kind of heredity that makes a baby respond differently to a smile vs. frown, track an object with eyes, etc.). Permit me some indulgence in the radical: suppose that intelligence can exist INDEPENDENT of heredity! That is, heredity is only a way of optimizing learning (by starting from more than scratch), but it is not necessary for learning to take place. The development of the child suggests this idea, which is in direct opposition to every "supervised" learning scheme in use. (I suggest the terms "autonomous" or "automatic" vs. "manipulated" instead of supervised vs. unsupervised.) The ultimate neural network would probably allow direct modification of its weights (like in artificial models) and also "automatic" mode (like in the brain). You'd get the best of both worlds. This is a convincing argument (among others) that if an artificial brain is ever devised with the capabilities of the human one, then there also exists one of the former that is superior to any of the latter. Now, don't take any of this as a criticism of neural networks in their present form. These are only observations. Even though results of learning from manipulative techniques such as back-propagation have been less than spectacular (requiring lots of training), manipulative techniques will probably always have a role in cases of identifiable ("right") output.
andrew@dtg.nsc.com (Lord Snooty @ The Giant Poisoned Electric Head ) (03/04/90)
In article <5061@ccncsu.ColoState.EDU>, ld231782@longs.LANCE.ColoState.Edu (Lawrence Detweiler) writes: > >Note that babies bootstrap, so that "knowing what to look for" becomes > >increasingly sophisticated. > > This deserves a thread of its own. This is what is so remarkable about > the human neural net, in my opinion: it doesn't have something feeding > it the "right" answers to its neurons. The whole thing is > self-contained. The accomplishments of conventional computational > neural nets are not so grandiose considering that they require such > intensive and explicit training. This is not the tired point that > "computational neural nets pale in comparison to the brain" and > therefore they can never hope to rival or surpass it (begging the > question). Rather, it is suggesting that there may be something > fundamentally different about current NNs and the one in the brain.[...] Well, I have to disagree with this. I used the word "bootstrap" intentionally; even a bootstrap needs a seed. In the case of a baby, there is strong evidence to suggest that certain vigilance functions are hardwired from day one, an important one being face recognition. The effect of this "vigilance filtering" is to effectively provide a narrow subset of "environmental inputs to which attention must be paid". Then there is the mimicry response, whereby the baby learns to mirror the filtered facial expressions, etc. The upshot of all this hardwired stuff is that the baby is effectively acting as a supervised net (no offence intended!) and is far away from acting as a totally unsupervised autonomous entity. The key point about the bootstrap is that, from small attentionally- concentrated beginnings, the remarkable panoply of behaviours emerges finally. > Likewise one could argue that in current computational nets there is > only input (the stimulus and correct response, so to speak). But the > relationship between stimulus and response is firmly established by the > EXTERNALLY DERIVED learning rule. In any real brain this is precisely > what must be learned! It must be INTERNALLY derived. How does one > neuron learn what influences it, what it is modeling? Mothers typically indulge in mimicry themselves (it seems sometimes as this is nearly instinctual, albeit adult behaviour!). Observing correlations between one's own responses and the responses of the object of attention (by the baby) could even remove the requirement for "hardwired mimicry", which is better parsimony. > One might argue that the role of heredity is to create a minimal > structure that experience can build on (the kind of heredity that makes > a baby respond differently to a smile vs. frown, track an object with > eyes, etc.). Permit me some indulgence in the radical: suppose that > intelligence can exist INDEPENDENT of heredity! That is, heredity is > only a way of optimizing learning (by starting from more than scratch), > but it is not necessary for learning to take place. The development of > the child suggests this idea, which is in direct opposition to every > "supervised" learning scheme in use. (I suggest the terms "autonomous" > or "automatic" vs. "manipulated" instead of supervised vs. unsupervised.) Well, I'm attempting to refute this. You are postulating extra entities - in this case, brains which possess eldritch properties no artificial net possibly could (a bit like Penrose) and there really seems to be no need for the extra axioms. > The ultimate neural network would probably allow direct modification of > its weights (like in artificial models) and also "automatic" mode (like > in the brain). You'd get the best of both worlds. This is a convincing > argument (among others) that if an artificial brain is ever devised with > the capabilities of the human one, then there also exists one of the > former that is superior to any of the latter. In that manipulating people is often deemed necessary, and requires a massive investment of effort in many cases (ask any salesman), I'd have to agree that you'd have the Ultimately Pliable Person there! This might well give a healthy boost to any consumer-oriented economy... Or is that a level of intelligence below what you meant? :-) -- ........................................................................... Andrew Palfreyman andrew@dtg.nsc.com Albania before April!
ld231782@longs.LANCE.ColoState.Edu (Lawrence Detweiler) (03/07/90)
>I used the word "bootstrap" >intentionally; even a bootstrap needs a seed. In the case of a baby, >there is strong evidence to suggest that certain vigilance functions >are hardwired from day one, an important one being face recognition. >The effect of this "vigilance filtering" is to effectively provide a >narrow subset of "environmental inputs to which attention must be >paid". The idea of "a narrow subset" sounds inconsistent with supervised (manipulative) learning, where instead EVERY function is implied by the selective reinforcement from the backpropagation rule. I acknowledge that there are hardwired vigilance functions. I suggest that they are not necessary for learning: they only optimize it. >Then there is the mimicry response, whereby the baby learns >to mirror the filtered facial expressions, etc. When I said that a baby "learns to learn" I mean that everything that looks like learning--including the mimicry response--may be a simple consequence of laws that are predominantly dependent on interactive sensory experience (Piaget's `circular reaction' referred by Gaudiano) and NOT (perhaps whatsoever) on mental heredity. >The key point about the bootstrap is that, from small attentionally- >concentrated beginnings, the remarkable panoply of behaviours emerges >finally. That you call them "vigilance functions" suggests not. They only serve to focus attention, the precursor to learning; they are not the mechanism of learning itself. Thus the central issue lies elsewhere and attention-focused beginnings are merely a peripheral consideration. >Observing correlations >between one's own responses and the responses of the object of attention >(by the baby) could even remove the requirement for "hardwired mimicry", >which is better parsimony. This is more evidence for an autonomous system. The idea that the final outcome depends not on some intrinsic model but on the stimulii in the environment precludes some overseeing, manipulating mechanism. [intelligence without heredity?] >Well, I'm attempting to refute this. You are postulating extra entities - >in this case, brains which possess eldritch properties no artificial net >possibly could (a bit like Penrose) and there really seems to be no need >for the extra axioms. No, I think you have it backwards. It is supervised mechanisms that would be extraordinarily difficult to encode computationally--we would have to understand the entire function of the brain's heredity. If intelligence is autonomous (as I propose) it would be drastically simpler to duplicate it in synthetic forms. You see, in the supervised approach one might say "learning depends on an intrinsic ideal model toward which the system converges." In autonomous learning one might say "learning depends on individual sensory experiences arising from the environment." Surely discovering the rules that govern the former would be much more difficult to discover than that of the latter. There's only so many ways something can react (change in accordance) to a stimulus, but an overwhelming complexity in heredity. >The upshot of all this >hardwired stuff is that the baby is effectively acting as a supervised >net (no offence intended!) and is far away from acting as a totally >unsupervised autonomous entity. The most plausible case for "supervised" (manipulated) learning in the baby is in the sensations of pleasure and pain. Presumably these are "hardwired". But how much of a baby's behavior is goverened by merely direct (physical) sensation of pleasure and pain? I think one can see that the vast majority of human experience is free of these direct reinforcements that appear to be the only possible biological analogues of manipulative learning. Don't you think you would feel it if all your learning was governed by some overseeing mechanism of reward and punishment? Pain and pleasure themselves may be functions that are conditioned into us by our environment in the classical Pavlovian way. In my first message I couldn't think of a term that was applicable. I remember it now: the idea of manipulative learning sounds like just another scheme of the homunculus--a little man inside our heads. In this case the man is disguised in our heredity that "oversees" the brain's correct development through manipulative techniques. This shows how the whole idea of supervision in the brain is hard to believe. What guides the guide? The idea that heredity is the homunculus is a clever idea because we reach a dead end.
jstern@orion.oac.uci.edu (Jeff Stern) (03/15/90)
Regarding the importance of curiousity in a baby, in neural nets, and in ourselves, I recall a quote I came across a few years ago: "Lao Tzu believed that everything that exists comes into reality through the polarity of *yin* and *yang*. He called the specific physical laws and cycles that control and govern reality the *Tao*, and suggested that the actions of the *Tao* reflect the purpose of a larger entity (the Absolute). So if reality came about because the Absolute wanted to know itself, then our evolutionary destiny must be to help it get a good look by investigating, observing, and emulating nature." --From R.L.Wing's new translation of the "Tao Te Ching" ISBN 0-385-19637-7 ...just a thought. Jeff. ---------------------------------------------------------------------- | Internet: jstern@orion.oac.uci.edu | Same? Different? | | CompuServe: 72647.1527@compuserve.com | I can't tell anymore. | ---------------------------------------------------------------------- -- ---------------------------------------------------------------------- | Internet: jstern@orion.oac.uci.edu | Same? Different? | | CompuServe: 72647.1527@compuserve.com | I can't tell anymore. | ----------------------------------------------------------------------
edwards@cogsci.berkeley.edu (Jane Edwards) (03/15/90)
In article <5130@ccncsu.ColoState.EDU> ld231782@longs.LANCE.ColoState.Edu (Lawrence Detweiler) writes: >The most plausible case for "supervised" (manipulated) learning in the >baby is in the sensations of pleasure and pain. Presumably these are >"hardwired". But how much of a baby's behavior is goverened by merely >direct (physical) sensation of pleasure and pain? I think one can see >that the vast majority of human experience is free of these direct >reinforcements that appear to be the only possible biological analogues >of manipulative learning. In article <4972@newton.praxis.co.uk> cdh@praxis.co.uk (Chris Hayward) writes: [lots deleted] >. . . >The key point, I think, is that if we include *curiosity* as a primary >motivator, everything else falls into place. It provides the bootstrap. In article <fdwB026V91DI01@amdahl.uts.amdahl.com> kp@amdahl.uts.amdahl.com (Ken Presting) writes: >I can't help with the discussion itself, but here is pointer >to some (relatively obscure) literature. > >George Kelly wrote a book called "A Theory of Personality" which does >almost exactly what Chris suggests. Kelly's work seems to have been >swamped by the larger cognitivist movement in psychology, but he is >still mentioned in introductory survey texts on personality. It seems to me that there are a couple of things getting mixed up here. The idea of a need for varied experience is alive and well in personality and social psychology and as well as a reputable history (in such areas as sensory deprivation research, and the interesting book _Functions of varied experience_, Fiske & Maddi 1961). In fact, even food-deprived rats seem to have a need for it - i.e., will press a lever for light stimulation rather than merely for food - so it is certainly not specific to only a few species. But a need for varied experience is quite different from curiosity, I think. It wouldn't seem right to say that the rats pressed the lever out of curiosity. There are two further questions, though: Are there any independent manifestations of curiosity besides the behavior which it is being invoked to explain (i.e., sensation-seeking or cognitive mastery)? If not, then curiosity as a motivator or drive runs into a problem of _circularity_ which is faced by all "drive"-type explanation. It's like explaining why people drink wine by positing a wine-drinking drive. It is only superficially an explanation - in actuality, though, really just a relabelling of what needs to be explained. Can "curiosity" increase the accuracy of our predictions of behavior? Probably not. This is probably the main reason that emotions are seldom mentioned in cognitive research. They tend to have global effects which are context dependent, diffuse and poorly understood, espcially in humans. In order to gain precision by using curiosity to explain/predict behavior, we would first need a theory of curiosity, including what is most and least attention-getting under which situations, etc. And once we had those kinds of hierarchies perhaps that would be theory enough so that we wouldn't need to invoke curiosity in addition. Or so it seems to me. Jane Edwards (edwards@cogsci.berkeley.edu) UC Berkeley
markh@csd4.csd.uwm.edu (Mark William Hopkins) (03/18/90)
In article <5061@ccncsu.ColoState.EDU> ld231782@longs.LANCE.ColoState.Edu (Lawrence Detweiler) writes: >>Note that babies bootstrap, so that "knowing what to look for" becomes >>increasingly sophisticated. > >Sure, an artificial NN can "learn" to speak by an EXTERNAL CONTROL that >modifies its weights. But the human can do it with an INTERNAL one. >Can somebody invent a black box that can learn to speak (and make >sense!) solely from interaction with the outside environment? A child who gets a spanking for swearing is not undergoing internal control :) The point is, you just have to look in the right places to find the human analog to a neural net's explicit training. The instinct to survive surely figures in here too in terms of providing "external control". You hit upon the key idea: there is an INTERACTION with an external environment. So there's going to be a teaching agent somewhere. If it's not a parent or authority, it'll be the environment itself or something in it. After all, teachers don't have to be animate. >A specious argument is that babies DO have "models", namely their >parents. But in the beginning it is all just so much input. Babies >must "learn to learn." The ability to "learn to learn" must be present at least in part from birth. Plus, there must be certain invariant fundamental ways of domain independent learning that we have all our lives from birth that cannot be changed, but that can be used as the basis for further bootstrapping: ways that are inherent in the very neuro-chemistry of our brain. The best way to resolve these issues is to start thinking of a human's mind as an integral part of a combined mind-body system. The human is an intelligent control system first, logic machine second if at all. All of our learning is based on this ancient attribute of our species and other species. We develop concepts and abstractions out of our actual concrete experiences of mobility, eating, sleeping, etc. If you think of a control function such as walking as being C function, the symbol for walking that we gradually build up would be a pointer to that C function, which itself can then be used in other functions as data. So we abstract routines into symbols and build up a towering hierarchy so much so that the concrete basis of the tower very nearly falls out of sight. That ability to create abstractions is built-in, and is probably uniquely human in terms of our ability to abstract on anything (and by exclusion, it must be deduced to take place in the neocortex). The famous broom-balancing neural net controller ideally would not have needed anything more than the learning algorithm (representing the built-in genetic endowment) a representation of the goal (which to measure errors against), and a physical device to control (a broom carrying cart). Without any external intervention it would have "learned" to do the balancing trick. The environment representing implicitly the laws of physics would have been the teaching agent, so its training would really be external. I'm not aware of the details of that experiment, but I don't think it actually had any external agent training it (except maybe to rebalance the broom).
bill@boulder.Colorado.EDU (03/20/90)
Let me point out that this is an issue philosophers have been arguing about for centuries. Locke believed that a baby is a "tabula rasa", with no prespecified cognitive structure. On the other hand, Kant, stimulated by Hume, decided that the ability to organize experience requires "synthetic a priori" concepts, which are not derivable from experience, but are not logical truths either; "space" is a prototypical example, "causality" is another. Also, there is increasing evidence that babies are "prepared" to learn certain kinds of things but not others. One of the most compelling examples, in my opinion, is phoneme learning. Each language uses a set of some thirty-odd phonemes (i.e., vowels and consonants, roughly speaking), and creates words by stringing them together. Different languages use different sets of phonemes. A baby, in order to understand language, must learn to analyze the stream of sound coming into its ears well enough to distinguish the phonemes of its language. This is a task so difficult that the best of modern computer technology does it very poorly; yet it is done almost automatically by virtually mindless three- year olds. If you were to show these same three-year olds pictures of the spectrograms of the sound (which contain the same information), you could show them pictures forever and they would never begin to be able to distinguish phonemes. (Adult experts can't do it.) The clearest examples of prepared learning, though, are seen in non-human species. Rats, for example, can easily learn to associate a novel taste with a later feeling of sickness, even when the sickness comes several hours later; and they can easily learn that a flashing light signals that they will be shocked unless they move away; but it is almost impossible to teach them to associate taste with shock, or a flashing light with a feeling of sickness. It seems that the nervous system has evolved to make strong assumptions about what sorts of things are likely to be causally connected, and connections that violate the assumptions are very difficult to learn. If this is true for rats, it is likely to be true for humans as well. -- Bill Skaggs