|
Author
|
Topic: Erik Larson: Ray Kurzweil's Impossible Vision
|
Light Jaguar
Member
Member # 599
|
posted 25. January 2003 19:44
I hope this is not divergently off the topic. I would just like to say that it seems evident (to me anyway) that if computational devices are unable to become conscious, then there will always remain a raft of things they are unable to do. So: will they be able to become conscious?
Your probable answer to that, of course, depends to a large extent on what you believe consciousness is. My belief concerning consciousness is on record: it is the ground of existence, and not a "process". If I am wrong. If consciousness is a "process" (as I am assuming that gedanken and others believe), then obviously there is no barrier in principle to consciousness in vivo silica.
I would speculate, though, that there is a barrier, and it goes like this:-
I maintain that it is not possible to "trick" or "coerce" the will of the world to sum over, or create out of, any specified stuff, an "organism" or living, sentient form, unless that will itself, and for its own reasons, becomes possessed of the urge to do so. There is always the remote possibility that it could become so possessed, but my bet would tend to be against it.
That ends my modest contribution, in which, much to my relief, I avoided having to think about the "frame problem" ![[Wink]](wink.gif)
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 25. January 2003 23:15
Micah said:
quote:
If we attempt to describe the effects of actions or events using logic, then we must describe both changes and non-changes in order to draw any useful conclusions.
and Erik added:
quote:
To be accurate and technical (not necessarily the same thing), we should say that the "frame problem" is a problem that arises when representing a domain using a logical language.
The solution seems pretty obvious here, just given the statements. Don't use a logical language. There is very little evidence that the core processing functions used in human cognition are well represented as a logical language.
Micah also said:
quote:
In other words, how do we program AI systems to handle the common sense law of inertia: inertia is normal, change is exceptional. Is there a generalized representational system, with only a few rules (because we simply can't program in all the rules), that will allow the system to disregard non-change and focus on the salient change in various, unexpected and novel environments?
Again, the answer seems to be obvious given the question. "How do we program AI systems to disregard non-change?" The answer is simple: learn what is going on, and only pay attention to changes once it's learned. We already have a wide variety of machine learning algorithms.
How do you pay attention to the salient changes? Well, you have to have some notion of which changes are salient and which are not. We instinctively are afraid of snakes and spiders (apparently these are programmed in in advance), and have to learn to be afraid of bombs. Note here that fear is one method of weighting relevance of an object based on instinct or past experience. Using emotions to control attention is an area that has not been explored as extensively in AI research as it deserves to be. And it is less than entirely useful without a system with powerful abstraction capabilities, of which there are relatively few (but ideas for several--Hopfield, for instance, is working on a system that should provide a portion of this capability).
quote:
Rex, again, I see no reason why the system you describe would represent a solution to the frame problem, unless you introduced the possibility of a proper response to unforeseen, unexpected and wildly divergent environmental stimuli.
The main reason why language recognition has been so problematic is exactly because you need to give a proper respons to unforseen, unexpected, and wildly divergent verbal stimuli. If you solve that problem, the generalization of the solution should solve the generalized frame problem. (Also, speech recognition is hampered by an inability to focus on the relevant auditory signals (i.e. what the speaker is saying, as opposed to the whirring of the dishwasher).)
IP: Logged
|
|
Erik Larson
Member
Member # 313
|
posted 26. January 2003 22:43
Hi everyone,
In my previous post I tried to give an adequate (barely adequate) treatment of the technical FP, thus providing a starting point for fixing the meaning of a generalized version. Such a generalized version would be usable for discussions about the limitations of AI. As I mentioned before, a natural explication of generalized FP is in terms of "relevance." Relevance is uniquely suited, I think, because it is clear that the essence of the technical FP is a system's inability to exclude or "ignore" irrelevant facts when reasoning in a dynamic environment. I hope that these observations are uncontroversial enough.
Now an idea put forth earlier was that all cognitive systems have a version of the FP (hereafter I mean the generalized FP unless explicitly stated):
(1) for all x if cognitive_system(x) then hasFP(x)
I believe here that we are quantifying over too many individuals. Humans are not part of the FP. The idea is not that all reasoning systems are imperfect (exluding from the FP then only God), but rather that computational systems lack a capability that humans have: a grasp of context when reasoning in dynamic environments. This point is obvious (or ought to be) because the "frame problem" is by definition just that capability that machines lack relative to humans. To say that there are many frame problems (i.e., that we also have a FP) is to render the concept meaningless. (Likewise it it makes no sense to say that humans have "solved" the FP, etc.)
Hence we can safely exclude people from consideration:
(2)for all x if computational_system(x) then hasFP(x)
Now (2) is debatable, as I can see from the other posts. But also from reading the other posts it seems we are in agreement on this claim:
(3)for all x if representational_logic_system(x) then hasFP(x)
The burden then (if there is to be any teeth to the FP criticism of AI) is to demonstrate that the FP is not architecture-dependent; it applies to all computational systems. By all computational systems I mean classical (symbolic), connectionist, or any other system that is Turing computable.
In fact I believe that (2) is not only demonstrable by reasonable argument, but that inclusion of the non-classical architectures such as neural networks make things worse for the critic of the FP argument against AI. These architectures are even more dismal at providing a workable cognitive framework that obviates FP-type objections.
In my next post I will try to develop the general argument as I (and others) see it.
Erik
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 26. January 2003 23:47
quote:
I believe here that we are quantifying over too many individuals. Humans are not part of the FP. The idea is not that all reasoning systems are imperfect (exluding from the FP then only God), but rather that computational systems lack a capability that humans have: a grasp of context when reasoning in dynamic environments. This point is obvious (or ought to be) because the "frame problem" is by definition just that capability that machines lack relative to humans. To say that there are many frame problems (i.e., that we also have a FP) is to render the concept meaningless. (Likewise it it makes no sense to say that humans have "solved" the FP, etc.)
Hence we can safely exclude people from consideration:
I understand fully that we can’t literally classify people as having the “frame problem” in a literal sense, because people don’t operate by limited “representational logic system” constraints. So it was never my intention to show that people can’t solve the “frame problem” as it was posed.
My point was rather to show that people don’t “inherently” have a grasp of the “context” of a given problem. That grasp, when it is gained, is gained by learning by way of experience. They gain that experience (learning) over time. When comparing the capabilities of people and capabilities of some future hypothetical AI, one must allow that AI to gain experience in the area of the problem equivalent to the experience that is needed by humans to grasp the “context”.
In the Dennett “robot” example, the problem description denies the robot experience with bombs and other aspects of the problem, and rather demanded that the “representational logic system” be able to find “context” without that experience. This is inherently hamstringing the potential characteristics of that future AI by the rules of the comparison so that it is inherently bound to fail. (In other words no AI that is limited by the rules of comparison will be able to meet the test, but an AI that can be developed without those limitations in rules may indeed be able to meet a test without those limitations. Part of this is the nature of the “test” itself must not ask more of the AI than is asked of the person.)
quote:
The burden then (if there is to be any teeth to the FP criticism of AI) is to demonstrate that the FP is not architecture-dependent; it applies to all computational systems. By all computational systems I mean classical (symbolic), connectionist, or any other system that is Turing computable.
We must also be careful here. “Turing computable” can be a very strange limitation, in the range of behaviors we wish to discuss. Are you asking for there to be a “recognition state” of a Turing machine that meets some specific characteristic? Or are you asking that a reasonable person would think that the AI was exhibiting fairly strong recognition of “context” in its response?
Even if the future AI system should work by some analog (and thus not exactly “Turing computable”) principles, I am willing to consider that we are limited to digital simulation of those principles. So in that sense the AI system is amenable to a Turing simulation that must in principle produce comparable results -- even though the computational resources of that Turing simulation would be far beyond the ability to be constructed, while the possibly analog AI implementation (of neural networks for example) is within the range of construction in the next 35 years. One thing that I expect to be left out of this is any argument of determinism of the computation (e.g. all “Turing machines” are deterministic), but I have not seen any indication that anyone is trying to raise this issue -- I simply mention it to head it off in the future.
quote:
(2)for all x if computational_system(x) then hasFP(x)
and
quote:
In fact I believe that (2) is not only demonstrable by reasonable argument, but that inclusion of the non-classical architectures such as neural networks make things worse for the critic of the FP argument against AI. These architectures are even more dismal at providing a workable cognitive framework that obviates FP-type objections.
Here first of all one must consider the combination of classical, neural network, and genetic algorithm approaches. GA approaches have found to produce advances in forms of “learning” in novel situations. So to leave GA out of the list, then claim that things are seemingly worse when including the new techniques seems a little disingenuous. And if the claim is being made that no GA system is showing promise in constructing solutions in novel situations, I suggest doing a better search of the literature.
The major conflict I see here is that goal of comparison is to see if the AI will be able to exhibit strong capacities to operate in novel situations. But the specified comparison being given is once again a highly technical check that an algorithm meets a condition (hasFP(x)) that is not considered as testable in the human. So inherently the machine is being held to a standard that the human is not, and that is being considered as a fair comparison. I disagree on methodology.
More specifically, no Turing machine algorithm x is ever not going to have the “hasFP(x)” for some problem y, when problems y are allowed complete range of all possible problems. Then if problems y are a limited set, there certainly can be Turing machine algorithms that can deal with a limited set of problems effectively -- as there are already AI systems that deal with a set of different problem types. The question is being posed as one of type when the comparison that is needed is one of degree. It is one of degree because we need to see if the AI can function strongly in a range of new problem types -- but in which the AI has been able to gain experience equivalent to type of experience that a human would have gained in working in the same problem area.
What we must consider is not “a grasp of context when reasoning in dynamic environments.” Rather what we should consider is a grasp of context when in dynamic environments, after being given an appropriate opportunity to learn about the new environment, such as the experience that has either explicitly or implicitly been given for the human to learn. And remember that we are talking about a hypothetical future AI -- no system has yet achieved the goal we are specifying, at the level of performance that we are asking..
This issue of “level of performance” is extremely important. It relates both to making sure we are giving the hypothetical AI the same opportunities to learn that we are giving the human, and then we are considering the likelihood that humans will fail in a similar situation. (This must include the learning opportunities given the human when we examine any failures of the AI to perform.)
These are why I gave my little “story”. Dennett’s “robot” was programmed (in the description) without allowing for benefit of years of experience to accumulate in ways of the world. (Such experience could, in principle, give the AI learned patterns that would allow it to “focus” more quickly on finding solutions).
In the end, the real question is really about how we could develop this ability to learn from experience to focus better on solving problems. But this was not the statement of the problem given, and has not been incorporated into what is suggested for comparison for the future AI system.
[ 27. January 2003, 00:59: Message edited by: gedanken ]
IP: Logged
|
|
Erik Larson
Member
Member # 313
|
posted 28. January 2003 22:07
Gedanken,
I appreciate your views and am glad that you've taken the time to discuss these issues in the context of my review. What I've been trying to do is put in place a "filter" so that we (you, me, and everyone else) can discuss whatever implication the 'frame problem' might have for the future of AI without climbing Babel. I think the "frame problem" (or whatever you want to call the relevance problems with AI) has profound implications. We ought, therefore, to get things straight.
Now, I want to skip ahead in my promises to you and everyone else, and assume that you're willing to grant that the frame problem applies to computational systems. That is the problem we are looking at.
Now, you ask, so what? I think what you really care about is: what will happen with computational reasoning tomorrow? Will computational systems STILL have the frame problem (the silly inability to deal with irrelevance), or will they begin to close the gap on human reasoning over time. This is the view most proponents of AI have. And it is natural enough. Kurzweil, he just has MORE of this view I suppose. He thinks we are on a steep curve, we are making rapid progress. Success is imminent (well, imminent in 2029).
So proponents of AI don't need to accept Kurzweil's timeline, as I believe you and others have pointed out. Whatever the actual dates, the point is that we can hold this beguiling idea of progress toward human reasoning abilities. "Not sure when we'll get human reasoning, but we are sure we're making progress-if incremental progress--toward the goal." Call this the "incremental progress" view of AI. (Also, "inch by inch everything's a cinch" will do).
I want to say something about this view. First off I admit that it IS, afer all, a natural way of thinking. It seems just one more example of the march of science. Science has done pretty well these days. So who should object? They don't call it "progress" for nothing, right?
But you see that the incremental progress view supposes that the progress is a step by step thing, that systems will get better and better (better at reasoning like humans, that is). It presupposes a view of human reasoning that precludes the possibility that, instead of incremental steps, there are huge principled divides--conceptual gaps not fillable by more computing power, clever engineering, etc.
This presupposition, you should be aware of, and it ought to be defended with arguments. So, what are the arguments? What makes the human and machine cases in principle the same, so that progress in AI is just progress toward closing a knowably contingent gap? Minds and machines sure SEEM different, after all, might it be the case that they really are?
The incrementalist is assuming quite a bit, perhaps. But then, perhaps not. At any rate, I think it's reasonable to start with a bit of skepticism about the future of AI, rather than unbridled optimism (even if it is optimism in only incremental progress). For one, because the field is notoriously "behind schedule." Things just might be a bit harder--or more precisely, different--than we originally all assumed.
Now if there are actually principled arguments for why the field is so "behind schedule", why progress toward human-like reasoning is so hard (how do we do it again?), in addition to the historical observations of failure that provide good empirical support for those arguments, why should you be so optimistic? On what, exactly, does your optimism about AI rest?
For my part, (I feel I should say this now), I'm no Luddite. I've worked as an AI engineer (both in academia and in the private sector), and I'm profoundly interested in making progress on AI. But in my view, it is not in replacing (or trying to replace) distinctively human types of thinking with machine computation that has much value. It is in finding interactionist approaches where people and machines contribute to solving problems by each providing what the other lacks (or doesn't want to do). Human minds are not computational machines, so what? Here is a whole interesting field of AI anyway. And I don't even have to bang my head trying to "solve" the frame problem...
As for my arguments, I have to defer, though they're not now as interesting as one might imagine. Fodor has put the point well enough. I am just a copy cat. Perhaps I've got some uniqueness by uniquely combining (existing arguments). But if anyone is interested I will try to lay out the problem of abduction, the global nature of beliefs, and the inability of computational systems to get this right.
Thanks for your messages.
Erik
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 29. January 2003 02:30
My optimism comes from personal experience with novel approaches to AI research. I agree that incremental advances that build only upon architectures that we have how are doomed to failure, among other things, because of the frame problem.
However, I know several people working in the area (and some proprietary information that I can't yet share) that encourages me that we now know how to implement one of the core capabilities that had been missing. This capability is that of expectation--how do you use what you already know, or what you want to find, in order to modify your current processing? This is not the single holy grail to AI, of course, but it's an important step. At this point, incremental progress (a lot of it, but incremental all the same) is all that's really needed to have a highly generalizable, robust expectation-fulfilling capability.
Of the other capabilities that I'm aware of that need construction, I see promising avenues along all of them. None of them seems to me to be insoluble (though some are very difficult), and when put together, should yield something with remarkably strong AI capabilities. Expectation and relevancy (which can be implemented similarly to expectation) can serve to pre-select options to consider, which should effectively remove the frame problem. Top level chess-playing programs (e.g. Deep Junior), in particular, are very good at avoiding the frame problem by doing a breadth-first relevancy-weighted search of moves. Of course, chess playing programs have an advantage: they already know what is relevant. However, this in no way precludes learning what is relevant during, and making the initial part of the search a search for relevancy.
Maybe a sketch of "the problem of abduction, the global nature of beliefs, and the inability of computational systems to get this right" will show why this is doomed to failure in an AI context. I would be interested in seeing the arguments.
I didn't find the Brakel article in Psycoloquy very persuasive. Dennett had some interesting points, but I didn't find any fatal roadblocks there either, as he concerns himself mainly with logic processors that were common back when symbolic AI was considered hopeful. For example, in one place he says that the lack of learnable expectation is a fatal flaw; we probably have that now. (We should know in 3-5 years; it looks promising to me in principle, but the real test is using it for an industrial-strength application.)
IP: Logged
|
|
Light Jaguar
Member
Member # 599
|
posted 29. January 2003 16:10
At the risk of not talking about the frame problem again, I resonated with some of what Erik said.
Just as there are different "IDs", so there are different "AI"s. Practical AI, or what I would call practical AI, can abstract certain elements out of huamn reasoning, or create certain analogues for human reasoning, and produce worthwhile, workable systems. Also, we can have systems, quite practical, which "reason" in a rather different way from human beings. Deep Junior, for instance, doesn't do anything like what we do, for the most part.
One branch of projective AI, though, is predicated on the assumption that human consciousness is actgually or somehow "essentially" computational in nature, and that when we reach decisions or make judgements, there are somehow "algorithm-like-things" going on inside our heads. Sometimes, I almost believe that one has to be not involved in AI to appreciate in some standing-back-from-the-canvas sense how terribly unlikely that really is.
I do agree it is a useful way of thinking up to a point. But this variant of strong AI, I think, mixing up an abstraction and the source of that abstracxtion (human mind). There are things that human beings do, which, if looked at out of the side of the eye and with a vivid (though healthy) imagination, could be described as "vaguely computer-like". Also, by constraining human behaviour in certain ways, you can make a person function "as a computer". This is not the same thing of course as saying that we could, in any way, argue back from the abstraction to the source.
I think better AI will be able to do a number of things very well. But I also believe that there are things AI will not be able to do at all, for reasons previously stated.
The sort of thing AI should be capable of doing well:-
Constrained learning tasks of increasing complexity
Finding solutions to problems that are tedious and soul-destroying for humans to work on.
As Erik said, entering into a synergy with human operators to solve problems.
Tasks requiring massive analysis to considerable depth in finite time.
Tasks requiring awesome and "intelligent" traversal of search spaces.
Make choices according to inferences and learnings.
Some of these are already being done, obviously.
SOme of the things I maintain AI will not be able to do, with the most important first.
*Be conscious.
*Experience emotion or sentient thought or a "sense of presence" in any way.
*Make choices as conscious minds make them.
*Create as conscious minds create (I agree: there will never be a "3 hour engrossing conversation at the bar", nor an epic novel that captures the world's imagination, nor a new musical masterpiece to rival our own).
*Have a spiritual awareness.
IP: Logged
|
|
Erik Larson
Member
Member # 313
|
posted 30. January 2003 20:56
Rex,
Well this proprietary theoretical breakthrough sounds really exciting. Say, if it is going public, any chance you could send me a message? I would like to buy stock after due diligence, since this kind of thing could easily lead to a suite of "next generation" applications that would rapidly revitalize the tech market!
Seriously, though (wait, maybe I WAS serious with you Rex), Light Jaguar's list made me think of a distinction that, to my knowledge, never seems to get much play in the AI research literature--descriptive versus generative computations. By 'descriptive' I mean finding some piece of reasoning that a human might do, and simulating it ex post facto using machine computations. I desire x, believe that y will help me get x, and that z will help me get y, and so I construct my plan to do z, then y, then goal x (or what have you--the details aren't important so long as the idea that one might reason in this way is clear enough).
Now some human really did just that piece of reasoning last Wednesday. So now we start laying out some rules, or a training set, or whatever patterns or structures we are using, that will give the result given the conditions. This kind of thing ought to be easy enough to _describe_ computationally, the question is whether it is possible to use these observations to generate novel cases that show that new, coherent plans or reasoning steps are being _generated_. Hence, generative computing.
Generative computing is really hard--for one, what sort of interests do you give the computer, so that it wants to go out and DO reasoning about the world at all? What does it want to reason about? Why? And even if you can side-step your way into an answer for that (and believe me, people WILL side-step--or rather headlong plunge--into discussions about how that's being done all over), how exactly is the computer to actually look for and find the relevant pieces of info, to do the novel generative reasoning? What guides the looking for relevant pieces of info? How does the computer _know_ what it needs to solve the problem, the problem being a new one? Well, it keeps looking. And looking and looking and looking. That is the frame problem of course (you check its reasoning to see what it is "looking" at, and the frame problem becomes clear enough).
Okay, so the distinction between ex post facto descriptive reasoning ("simulating" what some other thing did) versus generating answers to questions, the questions themselves from some set of desires--that is a big difference. On the one hand we crank through a set of steps, pointing the computer, wind up toy-like, at a set of facts and conditions to claim reasoning ("novel" reasoning if new facts and conditions from the last iteration), on the other we have a being with desires and beliefs, and a reasoning faculty devoted to satisfying its ends by pondering and conceptualizing its world. Different, different.
Okay, my "scratch pad" post.
Erik
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 31. January 2003 01:33
Erik, you can find a very basic form of the architecture in the article Sagi et al, "A biologically motivated solution to the cocktail party problem", Neural Computation, 13:1575-1602 (2001).
Of course, this is only one of the pieces that we need for AI-as-we-like-to-dream-about-it. You quite rightly point out that generative computing is hard, and never seems to get much play. You raise some very good questions that need to be answered before the problem can be solved.
However, the existence of hard problems is the reason why we don't have AI now, not a reason for us to never have AI. If we have impossible problems, then we will never have AI, and if we were making no progress on any of the hard problems, we would have good reason to be pessimistic. But the system I mentioned above goes a long way towards solving the expectation problem, so we seem to be making progress on some hard problems.
I don't see any intrinsic reason why the same type of scoring models used in chess-playing programs can't be used to make a computer "want to" go out and do reasoning about the world. Chess programs don't operate by comparing how close an outcome is to their goal state--chess is too deep a problem for that. Instead, the program cares about good positioning and tries to achieve that.
Generative computing is hard. It's hard for us. If I am faced with a new problem, how do I know what I need to solve the problem? I don't, for sure. Typically I think about the problem as such: what other problems are similar to this? How have I solved similar problems before? Which strategies were productive? What aspects of the problem are likely to be critical (based on my knowledge of solutions to similar problems)? I don't see any inherent reason why this can't be implemented as part of a weighted, fuzzy, breadth-first search strategy in a computer.
LJ said:
quote:
SOme of the things I maintain AI will not be able to do, with the most important first.
*Be conscious.
*Experience emotion or sentient thought or a "sense of presence" in any way.
*Make choices as conscious minds make them.
*Create as conscious minds create (I agree: there will never be a "3 hour engrossing conversation at the bar", nor an epic novel that captures the world's imagination, nor a new musical masterpiece to rival our own).
*Have a spiritual awareness.
I'm not sure we know how to address, philosophically, whether other humans are conscious, experience emotion, or have spiritual awareness. It's the whole first-person third-person problem. Trying to do this with computers just makes argument from analogy-with-self that much more problematic.
As far as making choices and being creative goes...I don't really know whether AI ever will--there are too many intervening problems that need to be solved to have any certainty. I imagine it is possible, and I imagine that we will probably do it, but if it is difficult, one has to ask the question: "why bother"? There are plenty of humans around who can do these things.
IP: Logged
|
|
Jesse
Member
Member # 112
|
posted 31. January 2003 03:48
One thing that some of you seem to have missed about Kurzweil's arguments for human-level AI is that they do not depend on our being able to solve the "frame problem", or even being able to understand much of anything about how the brain works at a high level. His main argument is that by 2030 or so we will have the technology to map out entire brains at the synaptic level, along with enough knowledge of the way individual neurons work to perform a fairly accurate simulation of those brains in computers. For example, see the section on "Reverse Engineering the Human Brain" in his article The Law of Accelerating Returns:
quote:
The most compelling scenario for mastering the software of intelligence is to tap into the blueprint of the best example we can get our hands on of an intelligent process. There is no reason why we cannot reverse engineer the human brain, and essentially copy its design. Although it took its original designer several billion years to develop, it's readily available to us, and not (yet) copyrighted. Although there's a skull around the brain, it is not hidden from our view.
The most immediately accessible way to accomplish this is through destructive scanning: we take a frozen brain, preferably one frozen just slightly before rather than slightly after it was going to die anyway, and examine one brain layer--one very thin slice--at a time. We can readily see every neuron and every connection and every neurotransmitter concentration represented in each synapse-thin layer.
The basic assumption here is that high-level intelligence emerges from the interactions of lots and lots of relatively "dumb" neurons arranged in the right sort of pattern; most neuroscientists today believe this is true, although dualists who believe in a supernatural soul would dispute it, along with those who think quantum effects are essential to intelligence (such as Roger Penrose and Stuart Hammeroff). But if the assumption is correct, and if Moore's law continues for another few decades, and if interactions between individual neurons are simple enough that we can learn pretty much everything there is to know about the factors that cause a neuron to fire at a given time, then Kurzweil's prediction is pretty reasonable. I think most people who disagree with Kurzweil would object to the first premise, although there is room for doubt about the others as well.
IP: Logged
|
|
|
Printer-friendly view of this topic