Krulwich-Bruce@cs.yale.edu (Bruce Krulwich) (11/04/88)
In article <349@uceng.UC.EDU>, dmocsny@uceng (daniel mocsny) writes: >Man can build physical mechanisms that can outperform his own physical >work capacity by orders of magnitude. We can't even define intelligence, >much less establish limits for it. I see no reason to doubt that he >will oneday build a machine that is more intelligent than himself, unless >the dualist view is correct (and physico-chemical mechanisms cannot >account for intelligence). However, if you asked me ``Can Man build a >_logic_ machine more intelligent than himself?'' I would laugh. What do you consider to be a "logic machine" ?? You might mean any of: - A formal system based on axioms and inference rules - A rule-based system (ie, OPS) - A standard computer - Lots of other things Which of these you mean determines the correctness of some of the things you say below, such as: >However, logic machines require explicit programming for the most trivial >tasks. They are not self-organizing nor adaptive. They do not learn from >everyday experience in a generally useful way. As long as that is true they >can never possess what we could reasonably call intelligence. This is not necessarily true for any of the definitions of "logical machines" that I gave above. Can you give some more details about exactly what you're saying?? >The connectionist approach to AI may succeed in creating machines that >correct these glaring deficiencies of logic machines. If so, then in >combination with logic machines they may create a hybrid intelligence that >exceeds anything we have yet seen. Especially if that hybrid includes us. If you're claiming that it's possible to do something with connectionist models that its not possible to do with "logical machines," you have to define "logical machines" in such a way that they aren't capable of simulating connectionist models. On the other hand, I think your claim is incorrect even if simulating connectionist models on "logical machines" is ignored. >In any case, discussing whether machines will exceed human intelligence >is a bit premature, rather like arguing over how tall a redwood seedling >might eventually become. Probably none of us will live to see the >question settled, and the seedling has an enormous struggle ahead of >it. Better to pay attention to nibbling away at subproblems... While you're working away on your subproblem, you shouldn't ignore other people's subproblems. While I am the last person to question the validity of connectionist approaches to AI, it looks as if you are unfamiliar with any recent work in the more "classical" areas of AI (ie, machine learning, case based reasoning, etc). Bruce Krulwich
dmocsny@uceng.UC.EDU (daniel mocsny) (11/07/88)
In article <42136@yale-celray.yale.UUCP>, Krulwich-Bruce@cs.yale.edu (Bruce Krulwich) writes: [ in reply to my doubts about ``logic-machine'' approaches to learning ] > If you're claiming that it's possible to do something with connectionist > models that its not possible to do with "logical machines," you have to > define "logical machines" in such a way that they aren't capable of > simulating connectionist models. Good point, and since simulating a connectionist model can be easily expressed as a sequence of logical operations, I would have to be pretty creative to design a logical machine that could not do that. (By ``logical machine,'' I mean any algorithmic device with sufficient generality to implement any of the instances you cited in your article.) I have a vague concept of a ``universal computer,'' gleaned from the occasional Wolfram or Hopfield paper, distorted somewhat through the transfer function of my inadequate understanding, but retaining some conceptual utility nonetheless. A sufficiently capable computer, whether based on a Von Neumann or PDP model, should be able to simulate all other computers, given enough time and memory. A machine works best in its own ``native mode,'' but that does not limit all the things we might kludge it up to do. An occasional human brain can (under appropriate duress) be made to operate at least momentarily much like a logical machine -- pushing symbols around, performing elementary operations on them one at a time, until the input vector becomes the output vector. I have trouble imagining that is what is going on when I recognize a friend's face, predict a driver's unsignaled turn by the sound of his motor, realize that a particular computer command applies to a novel problem, etc. Upon a microsecond's reflection I must admit that all connectionist models require explicit programming of some sort. Before they can start learning, someone must specify their structure, to ``get the ball rolling,'' so to speak. Indeed, our own brains start off with explicit genetic programming. The difference, I suppose, is all in the amount of programming required, compared to the total information gain. The information content of the human genome is ~750 MB, of which a sizable fraction determines our basic brain structure. The human brain goes on to absorb a terrific amount of information during its service life. (Terabytes? With electric stimulus, your brain can recall past experiences in vivid detail -- sights, sounds, smells, textures. If you've done any graphics or audio work, you'll know that's scary.) Can a system that only does logical inferences on symbols with direct semantic significance achieve a similar information gain through experience? Can we really, truly, specify a set of logical constructs that will fit on a Maxtor, turn it loose in the real world, and have it come back twenty years later to regale us with its discoveries? > On the other hand, I think your claim is incorrect even if > simulating connectionist models on "logical machines" is ignored. Time will tell. I long to be proven wrong. I would dearly love to have a computer that was not so brittle and helpless as the ones to be had today. I hope that I did not sound too critical of logical machines in my earlier post. I did say that they have many strengths where we have weaknesses. But the original question was whether they would exceed human intelligence. And that is a very tall order. > it looks as if you are unfamiliar with > any recent work in the more "classical" areas of AI (ie, machine learning, > case based reasoning, etc). I will appreciate pointers to significant results. Is anyone making serious progress with the classical approach in non-toy-problem domains? (One serious problem with the logical machine approach is that the bigger these systems get, the more likely they are to collapse. Success in toy domains is not easy to scale up.) Can a purely logical machine demonstrate a convincing ability to spot analogies that don't follow directly from explicit coding or hand-holding? Is any logical machine demonstrating information gain ratios exceeding (or even approaching) unity? Are any of these machines _really_ surprising their creators? Dan Mocsny
ray@bcsaic.UUCP (Ray Allis) (11/16/88)
In article <393@uceng.UC.EDU> dmocsny@uceng.UC.EDU (daniel mocsny) writes: >In article <42136@yale-celray.yale.UUCP>, Krulwich-Bruce@cs.yale.edu (Bruce Krulwich) writes: > >[ in reply to my doubts about ``logic-machine'' approaches to learning ] > >> If you're claiming that it's possible to do something with connectionist >> models that its not possible to do with "logical machines," you have to >> define "logical machines" in such a way that they aren't capable of >> simulating connectionist models. > >Good point, and since simulating a connectionist model can be easily >expressed as a sequence of logical operations, I would have to be >pretty creative to design a logical machine that could not do that. Whoa! Wrong! (Well, sort of.) I think you conceded much too quickly. 'Simulate' and 'model' are trick words here. The problem is that most 'connectionist' approaches are indeed models, and logical ones, of some hypothesized 'reality'. There is no fundamental difference between such models and more traditional logical or mathematical models; of course they can be interchanged. A distinction must be made between digital and analog; between form and content; between symbol and referent; between model and that which is modelled. Suppose you want to calculate the state of a toy rubber balloon full of air at ambient temperature and pressure as it is moved from your office to direct sunlight outside. To do a completely accurate job, you're going to need to know the vector of every molecule of the balloon and its contents, every external molecule which affects the balloon, or affects molecules which affect the balloon, the photon flux, the effects of haze and clouds drifting by, and whether passing birds and aircraft cast shadows on the balloon. And of course even that's not nearly enough, or at fine enough detail. To diminishing degrees, everything from sunspots to lunar reflectivity will have some effect. Did you account for the lawn sprinkler's effect on temperature and humidity? "Son of a gun!" you say, "I didn't even notice the lousy sprinkler!" Well, it's impossible. In any case most of these are physical quantities which we cannot know absolutely but can only measure to the limits of our instruments. Even if we could manage to include all the factors affecting some real object or event, the values used in the arithmetic calculations are approximations anyway. So, we approximate, we abstract and model. And arithmetic is symbolic logic, which deals, not directly with quantities, but with symbols for quantities. Now with powerful digital computers, calculation might be fast enough to produce a pretty good fake, one which is hard for a person to distinguish from "the real thing", something like a movie. But I don't think this is likely to be really satisfactory. Consider another example I like, the modelling of Victoria Falls. Water, air, impurities, debris and rock all interacting in real time on ninety-seven Cray Hyper-para-multi-3000s. Will you be inspired to poetry by the ground shaking under your feet? No? You see, all the ai work being done on digital computers is modelling using formal logic. There is no reason to argue over whether one type of logical model can simulate another. The so-called "neurologically plausible" approach, when it uses real, physical devices is an actual alternative to logical systems. In my estimation, it's the most promising game in town. >much like a logical machine -- pushing symbols around, performing >elementary operations on them one at a time, until the input vector >becomes the output vector. I have trouble imagining that is what is >going on when I recognize a friend's face, predict a driver's >unsignaled turn by the sound of his motor, realize that a particular >computer command applies to a novel problem, etc. Me, too! >Can a system that only does logical inferences on symbols with direct >semantic significance achieve a similar information gain through >experience? Key here is "What constitutes experience?" How is this system in touch with its environment? >I will appreciate pointers to significant results. Is anyone making >serious progress with the classical approach in non-toy-problem >domains? [...] > Can a >purely logical machine demonstrate a convincing ability to spot >analogies that don't follow directly from explicit coding or >hand-holding? Is any logical machine demonstrating information gain >ratios exceeding (or even approaching) unity? Are any of these >machines _really_ surprising their creators? > >Dan Mocsny Excellent questions. I'd also like to hear of any significant results. Ray Allis, Boeing Computer Services, Seattle, Wa. ray@boeing.com
ok@quintus.uucp (Richard A. O'Keefe) (11/18/88)
In article <8673@bcsaic.UUCP> ray@bcsaic.UUCP (Ray Allis) writes: >Whoa! Wrong! (Well, sort of.) I think you conceded much too quickly. >'Simulate' and 'model' are trick words here. Correct. A better would would be _emulate_. For any given electronic realisation of a neural net, there is a digital emulation of that net which cannot be behaviourally distinguished from the net. The net is indeed an analogue device, but such devices are subject to the effects of thermal noise, and provided the digital emulation carries enough digits to get the differences down below the noise level, you're set. In order for a digital system to emulate a neural net adequately, it is not necessary to model the entire physical universe, as Ray Allis seems to suggest. It only has to emulate the net. >You see, all the ai work being done on digital computers is modelling using >formal logic. Depending on what you mean by "formal logic", this is either false or vacuous. All the work on neural nets uses formal logic too (whether the _nets_ do is another matter). >>much like a logical machine -- pushing symbols around, performing >>elementary operations on them one at a time, until the input vector >>becomes the output vector. I have trouble imagining that is what is >>going on when I recognize a friend's face, predict a driver's >>unsignaled turn by the sound of his motor, realize that a particular >>computer command applies to a novel problem, etc. >Me, too! Where does this "one at a time" come from? Most computers these days do at least three things at a time, and the Connection Machine, for all that it pushes bits around, does thousands and thousands of things at a time. Heck, most machines have some sort of cache which does thousands of lookups at once. Once and for all, free yourself of the idea that "logical machines" must do "elementary operations one at a time".