|
Author
|
Topic: Erik Larson: Ray Kurzweil's Impossible Vision
|
Moderator
Administrator
Member # 1
|
posted 23. January 2003 07:03
Ray Kurzweil's Impossible Vision
by Erik Larson
ABSTRACT—Ray Kurzweil, he really does say the craziest things. So crazy, in fact, that if he weren't Ray Kurzweil, we might stop listening. Kurzweil, a computer technology genius and head of Kurzweil Technologies, an R&D technology company that specializes in Artificial Intelligence applications that include computer vision and speech recognition systems, has made a fortune showing the world just what computers can do. As an author and visionary in the field of AI, he's also had no small success telling the world what computers will do. His record thus far isn't bad: in 1990 he predicted the year a computer would defeat the world chess champion, and as anticipated in 1997 Gary Kasparov waved the white flag to IBM's Deep Blue supercomputer. That, and Kurzweil's status as a pioneer in AI software, has given his more radical predictions an air of added credibility. With that in mind, try this one: By 2029, machines will be as smart as us.
To read the entire paper, please click here
[ 23. January 2003, 07:05: Message edited by: Moderator ]
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 23. January 2003 08:10
This is an interesting review, and I fully agree that Kurzweil's unreasonably optimistic in his timetable; there are many hard problems that need to be solved before we have workable AI.
However, I think the pessimism about a lack of solutions to hard problems is somewhat misplaced. AI researchers originally assumed the problem was too easy--probably because of a lack of knowledge of the vast complexity of biological systems that performed the same function that they were trying to emulate.
That doesn't mean that the hard problems aren't being addressed, though. I've personally done research into overcoming the "hard" perceptual input problems (e.g. the cocktail party problem, where multiple people are speaking at once, yet you want to listen to only one; humans can do this with ease, and speech recognition systems produce utter garbage even with much less interference than that) by using simple expectation-based architectures that can in principle be extended to simple analogs of "human level cognition such as thinking and reasoning". (The claim is made in the paper that "we really don't understand at all how a computer could ever reason in context", but I tend to disagree, since the whole point of the research I was involved in was doing context-dependent perception, and there are a few (very few, but more than none) people who understand how to perceive in context and extend that to very simple analogs of reasoning.
So I think it is at least as unwise to assume that AI will fail as it is to accept Kurzweil's wildly optimistic predictions. There will still be philosophical questions about the mind/brain problem, and whether computers are really intelligent or just give the appearance of being intelligent. But I am fairly optimistic that potent AI will be realized within most of our lifetimes (though not as carbon copies of humans--why do that, even if you could?...we already have plenty of humans!), whether there are issues of ontological committment or no.
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 23. January 2003 08:40
quote:
So I think it is at least as unwise to assume that AI will fail as it is to accept Kurzweil's wildly optimistic predictions. There will still be philosophical questions about the mind/brain problem, and whether computers are really intelligent or just give the appearance of being intelligent. But I am fairly optimistic that potent AI will be realized within most of our lifetimes (though not as carbon copies of humans--why do that, even if you could?...we already have plenty of humans!), whether there are issues of ontological committment or no.
Rex,
I'm sure that Erik doesn't think that AI will "fail." What is the criteria for success by the way? When you say "I am fairly optimistic that potent AI will be realized within most of our lifetimes" what does potent AI consist in? These are important questions when discussing AI because they help us to distinguish between different visions of what AI can do. The Utopian AI of Kurzweil and others is one vision. The practical AI that is being realized in various systems around the world today is a completely different species. I think the potential of AI as it is currently being realized is potent. I don't think that we will see anything like robots that go to the bar with us, have a drink and hold engrossing three hour conversations.
Personally, I think that the Frame Problem will prove to be unsolvable and thus you will see AI succeed at the level of highly specialized systems but failing at fairly simple tasks which require global reasoning.
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 23. January 2003 10:39
Potent meaning that the "frame problem" will be solved. Most AI will probably be specialized since that is what is economically valuable, but I expect to see generic verbal communication and reasoning modules in everything important.
For instance, I might go into the kitchen and ask my fridge, "I'm hungry, but I had chicken for lunch. What can I make?" And the fridge would reply, based on what is in it. Or I might tell it, "I'm in a bad mood today," and it would tell me, "I have a quart of mint chip ice cream in the freezer."
Whether such a system is centralized or split between various components will probably depend on how expensive it is--if it is expensive, you'd probably tend towards a centralized system with more general capabilities. But something like that ought to be attainable, and it is far beyond the capabilities of current multilayer perceptrons and whatnot.
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 23. January 2003 11:53
Rex,
Systems such as the one you describe already exist in the AI/AL labs of MIT. Such systems, however, do not come close to solving the frame problem. For example, in the 1970s systems already existed that were able to be put into a room with colored boxes, find the boxes and order them based on colors. The frame problem arises when a system is faced with novel scenarios (what happens when one of the boxes has a bomb glued to it?) that it was not programmed to deal with. Well, we could keep programming in explicit directions to deal with explicit situations. Well, that won't work unless we can think of all the possible situations before hand. How about a logic that considers every variable in the environent: is the wall orange -> if yes, do this ->if no, do that. That won't work either because, first of all, we can't account for all possible environmental variables and second if we did, we can't come up with an action scheme that accounts for all the possible reactions to a given environmental state. Humans intelligence is able to deal with novel situations, to hone in on the "salient" information. What is the mechanism for doing this? Can we simulate the mechanism in silico? If a human is stacking blocks and sees a bomb, it can recognize the anomaly without considering all the other environmental information like whether the ceiling is 4 feet above its head or 7 feet. Human intelligence is able to hone in on the relevant information for a given situation without relying on brute force search. [Please recognize that this is in no way an argument against the biological basis for this ability. It is just an argument that *we don't even know where to start* to simulate the ability in silico*]
AI is great when placed in controlled environments but when faced with anomlies it breaks down...fails. Daniel Dennet has acknoweldged as much and thinks it is the greatest problem facing AI moving forward.
I plan on writing up a summary of the frame problem to post tomorrow so that we are all on the same page when we make reference to the term.
[ 23. January 2003, 12:07: Message edited by: Micah Sparacio ]
IP: Logged
|
|
Mark Szlazak
Member
Member # 391
|
posted 23. January 2003 20:30
I think the field needs a name change, that will solve most of "A.I.'s" problems.
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 23. January 2003 20:52
I agree with comments that imply that Kurzweil was a bit self serving and has overstepped what can be known. I agree that Kurzweil made possibly unwarranted philosophical assumptions (from the information I have so far -- subject to change), and that the claims go beyond what will ever be demonstrated. I agree wholeheartedly that Kruzweil’s timetable is too optimistic. Indeed it will not matter how responsive a machine can be, there will be critics that will make claims of distinctions on a philosophical basis. (And those who know me as having a high belief in the future of AI, and as a severe critic of Dembski may be surprised to hear this.)
But on the “frame problem”:
I think that the issue of the “frame problem” will always be a matter of degree. The reason is that humans also have problems with the “frame problem”.
For one thing, to pose the frame problem as a single problem is a mistake. Therefore there won’t be a stroke of solving “the frame problem”, as the unlimited list of nested problems that make up the “frame problem” will most likely be solved or improved upon incrementally. And to claim that no aspect of the nested “frame problems” have been solved is also a mistake, as many current AI systems definitely make use of “focusing” decision making aspects to improve search.
From review:
quote:
The reasons for this are involved and would require lengthy explanation, but here is a synopsis: any procedure that is locally specified (read: written out in a computational format) seems, ipso facto, an impossible candidate for performing global belief-based reasoning (in other words, for performing everyday human thinking!). To put it another way, the construction of a computer program using any architecture we choose becomes a candidate for interesting cognition only by "knowing" many facts about its world. But by the very fact of its having such deep, rich knowledge, it apparently won't have any means of picking out the small, relevant subsets of those facts needed for particular instances of reasoning. It is not simply that this problem is "hard", and will take a while to solve, but rather that it seems to lack a solution entirely, given our concept of computation. And since we really don't understand at all how a computer could ever reason in context, a fortiori we have no clue how a computer could ever reason like a person. So not only is there no known theoretical bridge to reproducing genuine human thinking, there is no such imagined bridge either!
I think this is an “argument from ignorance”. The fact that the writer does not feel in possession of this kind of knowledge does not mean that no one has made progress. And it certainly does not imply that such progress will not be made in the future. Most attempts at “proving” or “demonstrating” that human thought are outside of all possible computational limitations actually just wind up showing that human’s probably don’t solve the problems in the way the demonstration was claiming.
For example I was involved in solving NP complete problems of a specific nature in more efficient manner. (Solutions that are still being argued over 20 years later). Part of the question is how could a human recognize patterns that potentially would require an exponentially large amount of information to be processed. What we found is that real problems that we solve in the real world aren’t the difficult NP complete problems that take exponential time -- at least the solutions we come up with are not done so in a manner equivalent to the exponential growth algorithm.
Here is an example that relate to this, wherein finding the proper combinations of relationships in following the curves of the image affects what you see. This is a prime example of the “frame problem” and of human failure to solve it immediately:
[And sorry if these aren’t very permanent, I took the first copies of this I could find]
If you don’t see both consistent images, click here for a slightly different version of this.
In these pictures, the human often doesn’t solve the “frame problem” for both consistent images: a young girl looking away left, or an old woman in profile slightly toward us left. The second picture I included is more easily recognized than the first, so I put it as a link. Once you know what to look for (side information) then solving the problem (“seeing the picture”) is easy.
How many times have you realized: “I just wasn’t thinking about the problem with the proper perspective!” This often happens when one serendipitously gets a suggestion from random external events -- thus a “random” search happens to “land” on the constraint problem’s solution.
The “frame problem” certainly won’t be “solved” in the near future. But improved solutions will be found for various “frame problems.” Neither will the frame problem be “solved” for human thought. We should be careful of claiming that AI won’t progress because computational problems are inherently unsolvable. Don’t claim that something that cannot be shown to be a necessity for human thought is a requirement for AI to perform in a similar manner.
Interesting links (Google search):
Good short paper by Dan Sperber and Deirdre Wilson
Psycoloquy (electronic journal) -- “The Complete Description of the Frame Problem”
Masters student Peter Doomen, interesting short
a page on “MITECS”
Hays and Ford
J. van Brakel on Ford & Hays
more Hays and Ford
apparently Dennet: COGNITIVE WHEELS: THE FRAME PROBLEM OF AI
[ 23. January 2003, 23:35: Message edited by: gedanken ]
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 24. January 2003 10:00
Micah, I think you may have misunderstood my example.
Can the MIT system recognize my voice, in a noisy kitchen environment, and come up with solutions to my queries when I phrase them in arbtrary ways, including in areas where it is not specifically specialized (e.g. mood)? If so, I'd be extremely impressed, and think that Kurzweil's timeframe was pessimistic.
Inasmuch as the frame problem consists of brittleness and lack of context, my collaborators and I already have a good idea of how to proceed and have demonstrated a system that exhibits very simple solutions to these problems. So I do not think it is accurate to say that "we don't even know where to start".
(Also, note that if a person has never seen a bomb before and knows nothing about it, they will probably fare little better than an AI system when dealing with one glued to a box. We cannot fault AI systems for lack of capability when we don't give them any training! We can, however, fault them for being unable to be trained except on one very narrow task, which is the case for most AI now.)
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 24. January 2003 10:56
Rex, Gedanken,
I highly recommend the Dennett paper though, as others have noted, it is a philosophically liscenced interpretation of the frame problem.
The two books that I most highly recommend on the subject are:
The Robot's Dilemma Revisited: The Frame Problem in Artificial Intelligence
The Robots Dilemma: The Frame Problem in Artificial Intelligence
There are several interpretations of the Frame Problem and so the tough part is always going to be: exactly what are we talking about.
Here's the best technical summary I've happended upon, taken from Murray Shanahan:
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.
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?
BTW, please remember that the frame problem is a computational/logic problem.
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.
[ 24. January 2003, 10:59: Message edited by: Micah Sparacio ]
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 24. January 2003 12:15
[Warning, stories here contain violence]
From Micah’s early post:
quote:
The frame problem arises when a system is faced with novel scenarios (what happens when one of the boxes has a bomb glued to it?) that it was not programmed to deal with.
From first paragraphs of Dennet’s COGNITIVE WHEELS: THE FRAME PROBLEM OF AI:
quote:
Once upon a time there was a robot, named R1 by its creators. Its only task was to fend for itself. One day its designers arranged for it to learn that its spare battery, its precious energy supply, was locked in a room with a time bomb set to go off soon. R1 located the room, and the key to the door, and formulated a plan to rescue its battery. There was a wagon in the room, and the battery was on the wagon, and R1 hypothesized that a certain action which it called PULLOUT (Wagon, Room, t) would result in the battery being removed from the room. Straightaway it acted, and did succeed in getting the battery out of the room before the bomb went off. Unfortunately, however, the bomb was also on the wagon. R1 knew that the bomb was on the wagon in the room, but didn't realize that pulling the wagon would bring the bomb out along with the battery. Poor R1 had missed that obvious implication of its planned act.
Back to the drawing board. `The solution is obvious,' said the designers. `Our next robot must be made to recognize not just the intended implications of its acts, but also the implications about their side-effects, by deducing these implications from the descriptions it uses in formulating its plans.' They called their next model, the robot-deducer, R1D1. They placed R1D1 in much the same predicament that R1 had succumbed to, and as it too hit upon the idea of PULLOUT (Wagon, Room, t) it began, as designed, to consider the implications of such a course of action. It had just finished deducing that pulling the wagon out of the room would not change the colour of the room's walls, and was embarking on a proof of the further implication that pulling the wagon out would cause its wheels to turn more revolutions than there were wheels on the wagon - when the bomb exploded.
A story:
Nai and Eve lived on a nice plot of land near a village in a very remote region. The existence of “bombs” had never affected their simple lives.
One day Nai and Eve both went into town with their hand carts. Unbeknownst to them, Nai’s extremely evil brother had returned from the West and wanted to inherit all their land -- so he put bombs in their hand carts.
Nai was a plodding type who didn’t think very far ahead. Nai went to the butcher to obtain meat to fuel his hunger for the dinner meal. When he brought out the meat to put it in the cart, he saw the bomb. He wondered what to do with it, so he brought it into the shop to show it to the butcher and his assistant. They were all blown up when the bomb went off.
Eve was different from Nai, as Eve was always planning for the future. Eve arrived at a friend’s dwelling with her hand cart, and her friend was going to give her some vegetables for their precious food supply. When she entered her friend’s dwelling, she noticed the colors of her friend’s room’s walls. So she sat down to think about how that color would look in her dwelling. As she considered the implications of how much work she would have to spend scraping her walls, the bomb went off in her cart. She was spared but had lost Nai and both of her family’s hand carts.
The evil bother was not done though, because he would not inherit their land. And the next day the sad Eve needed to go into town for another errand and no longer had a hand cart, so she hooked up her horse Horse to the horse drawn two wheeled cart. While in town, the evil brother went to place another bomb in the back of the cart. But Horse sensed the fear and unusual activity in the evil brother, and reared up, dumping the cart and contents, galloping away. In the ensuing tumble, the evil brother was destroyed by his own contrivance.
But the town people had learned about bombs from the repeated experiences. They learned to focus on what to do when they saw such a suspicious device that matched a pattern that they had learned to recognize from the gear they discovered in the evil brother’s possessions. What they learned is to run -- a simple fear response that works in most circumstances. It served them in the future in their simple lives.
Moral of this story:
The “frame problem” affected Nai, Eve, and Horse in different ways. Novel situations were not handled in optimal ways. Some solutions were effective for unexpected reasons. Learning took place in the humans, where they adapted to a new pattern -- but it was not simply a matter of inherent knowledge of how to plan. Most importantly, we realize that complete exhaustive analysis of all possibly relevant facts is not important, and that recognition of certain patterns should trigger a change of priorities in behavior as a learned response.
Micah, why do you assume that intelligent responses must encompass computational logic? I’ve worked on computational logic search strategy. Results are used in planning systems that can plan projects far better than any human can schedule the events, because the planning algorithms can see many more consequences than the limited human mind. (This is similar to my hand calculator being able to calculate many more digits faster than I could ever do by hand -- it does not imply my hand calculator is “intelligent”.)
[ 24. January 2003, 13:37: Message edited by: gedanken ]
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 24. January 2003 14:17
quote:
Micah, why do you assume that intelligent responses must encompass computational logic? I’ve worked on computational logic search strategy. Results are used in planning systems that can plan projects far better than any human can schedule the events, because the planning algorithms can see many more consequences than the limited human mind. (This is similar to my hand calculator being able to calculate many more digits faster than I could ever do by hand -- it does not imply my hand calculator is “intelligent”.)
I'm not sure I understand what you are saying when you tell me that I am assuming "that intelligent responses must encompass computational logic." I'm simply stating what I understand to be the frame problem, and the inability to solve it via computational logic. I'm not saying anything about how this bears upon intelligence nor whether intelligence is computational (indeed, I think that computationalism is misguided). I am also not saying that the frame problem is not solvable outside of AI. After all, the problem is not a problem for billions of organisms around the world.
Gedanken, search strength is totally irrelevant to the frame problem. If the frame problem was related to search capabilities then we would be talking about the combinatorial explosion problem. We are not.
Also, what is this business about "inherent knowledge of how to plan." Was this ever a part of the frame problem? Who is saying that "inherent knowledge of how to plan" is essential to solve the frame problem? Knowledge is not the issue, never was. The issue is ability.
The solution to the frame problem does not lie in some unique conscious activity on our part but in something fundamentally organismal that has eluded computational simulation. That is all that I am claiming. I'm not sure that I see the point of your story...it is dealing with a strawman far off in left field.
[ 24. January 2003, 14:26: Message edited by: Micah Sparacio ]
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 24. January 2003 18:24
Micah,
quote:
The solution to the frame problem does not lie in some unique conscious activity on our part but in something fundamentally organismal that has eluded computational simulation. That is all that I am claiming. I'm not sure that I see the point of your story...it is dealing with a strawman far off in left field.
Actually I think that “I'm not sure that I see the point of your story...it is dealing with a strawman…” is just what I think about the “frame problem” description as a limitation on effective AI.
The description of “the frame problem” in terms of an AI system failing to deal with the real world is put in terms of search. The search (in computational logic terms) is suggested to spend much time focusing on irrelevant aspects. Then aspects are suggested like “…[the AI] was embarking on a proof of the further implication that pulling the wagon out would cause its wheels to turn more revolutions than there were wheels on the wagon - when the bomb exploded” from Dennet’s description. That description of “embarking on a proof” is a claim that the AI program was slow to pass over and discard that issue. There are two aspects: why is it slow, why is it even looking there? The reason it is slow (having bothered in the first place) is precisely combinatorial explosion -- not the kind that makes it intractable, just the kind that makes it take a significant amount of time so it is not passed over in a millisecond. Why is it even looking there is a problem of focus.
I think the major point is the focus problem, how does the AI learn where to spend its focus. This is where I find the greatest disparity between the claim and the reality of the research. First the problem is real, it is not imagined by the reviewer, Dennet, Micah, or myself. But a large amount of focus in the last 20 years in AI research has been in within discipline expert systems. These don’t (as described) know how to get outside of their domains -- just as the example provides. But within their domains they do very well at selecting focus! That is part of what has been learned.
So I think that there is a “frame problem” in humans trying to solve the very real underlying problem of focus. Descriptions like the “frame problem,” which clearly are accurately described (even by Micah) as “a computational/logic problem” are probably missing the point in terms of pointing us to solutions.
One aspect is that the critiques suggest a need for a programmer to enter new information for each new case in the “frame problem” based critiques. This is really pointing to the problem of learning, not the problem of focus exemplified by the descriptions of the frame problem as given. New GA, neural network, and fuzzy logic techniques are partially addressing the learning problems.
But these should not ignore that consistency checks are inherent in AI problems, even for a NN type process -- and some search processing will still be essential. But these are “mini-“ searches, not exhaustive and therefor intractable searches. In addition there is another kind of search -- a low order layered pattern matching search -- that is only now becoming feasible to simulate with modern CPU’s power and potential massive parallelism at a low cost.
quote:
Also, what is this business about "inherent knowledge of how to plan." Was this ever a part of the frame problem? Who is saying that "inherent knowledge of how to plan" is essential to solve the frame problem? Knowledge is not the issue, never was. The issue is ability.
Knowledge of where to focus is precisely the issue. Ability depends on knowledge. If you don’t know or have experience with the kinds of relationships that actually exist in nature (like something connected to an object moving along with that object when it is moved) then you can’t solve the new problems that are associated with the new and unusual case. And if you do have that knowledge -- then the search programs suggested in the critiques are easily solved by today’s planning AI systems. The “frame problem” as posed is how to deal with new cases when in a state of lacking knowledge of the important relationships. Then programming was suggested to add knowledge of some of those relationships, but lacking focus. And that was either inclusive of the knowledge of the relationship needed, or exclusive of that knowledge. If inclusive, it is indeed a “search” problem and its tractability, and if lacking it is also a knowledge problem. The “ability” that is considered to be lacking can be viewed in part as a lack of knowledge of where to focus and prioritize one’s attention. This is a matter of experience in humans -- and “experience” is knowledge. Beyond that the problem becomes an intractable search problem, both for human and machine. And I’m the one saying this.
My story is just as much “a strawman far off in left field” as Micah is suggesting. That’s just what I think the “frame problem” descriptions and implied limitations as compared to actual solutions that will and are being found to the problems of focus and knowledge that are suggested by the frame problem. My story shows humans who actually use learning to change their focus (plus some level of instinct built in by a kind of “programming”). It shows how they also have problems with the frame problem, but can eventually improve with learning. AI systems will also have to improve with learning -- they can’t be “programmed” to solve every problem.
[ 24. January 2003, 19:25: Message edited by: gedanken ]
IP: Logged
|
|
Erik Larson
Member
Member # 313
|
posted 24. January 2003 22:35
Hi everyone,
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. McCarthy coined the term in the 1960's (if memory serves) to describe a problem in capturing the set of propositions an agent must know when forming a belief about its environment. For instance, I need to know that a knife is a solid object well suited for spreading mayo if I am making a sandwich. I don't need to spend time thinking about whether a knife could be used as a transportation device for hamsters, and all of this kind of thing. The problem practitioners in AI were having--what led to the phrase "frame problem"--was in preventing a system with a representation of a domain from spending its time trying to prove absurd and irrelevant facts using the assertions in its domain. Specifically, the problem arose when using a representation to reason about change (i.e., in environments where things change and so new states of the environment must be considered--what are the new beliefs that the agent should keep using?).
So Micah is exactly right--if we want to wear the computer scientist hat then we need to restrict the discussion to representational logic systems (where the language is something like first-order logic and the inference mechanism is some suitable theorem prover). With this hat on, we might (read: should) all agree that the frame problem is an unsolved research problem. Also we might be able to agree that complexity considerations (e.g., are we in NP, or P-SPACE, or whatever) aren't really germane, since the sort of problems we are modeling and the expressivity of our language (e.g., FOL or even quantifying over classes i.e., HOL) mean that our worst case performance is undecideability. Admittedly, no one worries about that bug-bear in practical engineering puzzles, since there are mechanisms for terminating queries. Agreed. But more to the point people do worry about how exactly to "frame" an agent's set of beliefs to permit reasoning about change. And when we sit down and think about this problem, the whole idea that there is some way to specify what beliefs to use when reasoning about changing states (and how many state changes are there in a realistic scenario?) becomes really mysterious.
Other comments:
* The idea that there are a "bunch of frame problems" seems to me nonsense. The frame problem is a type-level idea--there are many instances of the frame problem but one problem of this type.
* If we expand what we mean by the "frame problem" then there are different names for this type of problem: the problem of relevance, or the problem of change, etc. I think a natural philosophical explication here is the notion of relevance--how do we restrict reasoning about change to relevant sets of beliefs, while ignoring irrelevant ones? That usage seems okay to me, since it's easy to see that McCarthy's "technical" frame problem arose precisely because a computational system can't keep irrelevant beliefs from affecting its reasoning about changes in the environment. Some would disagree with this philosophical treatment; in that case we need to talk about representational logic systems (a bit restrictive, no?).
* The idea that humans have their own "frame problem" supposes, of course, the non-technical or philosophical treatment of the problem. No problem. But it also commits itself to the idea that flaws in our cognitive abilities (i.e., in pattern recognition, belief-based reasoning, discrimination, etc) are somehow in the same boat as problems discovered about computational systems. At one level there are parallels: no person or machine is a perfect reasoner. Everyone--every known thing--fails at something. But the problem with this line of thought is that pointing to flaws is both obvious (and therefore not much help in addressing anything) and beside the point. Beside the point because the existence of "flaws" presupposes that there is some functioning, core capability. In people, this is true enough. But the problem with designing computational systems lies in finding some suitable functioning capability in the first place. Perhaps we can find one, perhaps not. At any rate human and machine cases are different.
* In fact I don't think that there is much hope in finding anything like a "functioning core capability" for solving the (philosophical or technical) frame problem, at least given our current concept of "computation." I want to develop this point, but I will have to do this in a follow-up post.
Erik
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 25. January 2003 00:11
I want to discuss some selected quotes from Erik Larson at length, but taking them as a whole, so I will quote them first with minor introductory comments:
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. McCarthy coined the term in the 1960's (if memory serves) to describe a problem in capturing the set of propositions an agent must know when forming a belief about its environment. …
Basically this was a statement that the domain we must work in for solving the problems is “a domain using a logical language”. Indeed McCarthy undoubtedly did coin this term then, and was undoubtedly using “a logical language” representation of knowledge and AI in general. (In fact LISP was developed in that time and in his lab if I’m not mistaken, specifically to be such a “logical language” or for processing such. EVAL function could evaluate the truth of a logical function in current state of reasoning or computation.)
Though in the same paragraph, I’m separating this out as the remainder:
quote:
… For instance, I need to know that a knife is a solid object well suited for spreading mayo if I am making a sandwich. I don't need to spend time thinking about whether a knife could be used as a transportation device for hamsters, and all of this kind of thing. The problem practitioners in AI were having--what led to the phrase "frame problem"--was in preventing a system with a representation of a domain from spending its time trying to prove absurd and irrelevant facts using the assertions in its domain. Specifically, the problem arose when using a representation to reason about change (i.e., in environments where things change and so new states of the environment must be considered--what are the new beliefs that the agent should keep using?).
I separate this out, because the problem to be solved does not necessarily have to be solved using the methods from the 60s that were proposed when the problem type was initially identified.
Then this:
quote:
So Micah is exactly right--if we want to wear the computer scientist hat then we need to restrict the discussion to representational logic systems (where the language is something like first-order logic and the inference mechanism is some suitable theorem prover). With this hat on, we might (read: should) all agree that the frame problem is an unsolved research problem. …
and
quote:
… But more to the point people do worry about how exactly to "frame" an agent's set of beliefs to permit reasoning about change. And when we sit down and think about this problem, the whole idea that there is some way to specify what beliefs to use when reasoning about changing states (and how many state changes are there in a realistic scenario?) becomes really mysterious.
Once again, the first reinforces the idea that we must solve this problem using “representational logic” or something like “first order logic”. Then the second part returns to the general nature of the problem to be solved. We must ask: Is the question whether the problem can be solved? Or is the question whether the problem can be solved using “representational logic” or “first order logic”. The reason is that these questions have entirely distinct answers. The answer to the second is undoubtedly no, but I believe that we will see significant abilities of computer systems to solve the first question in the next 30 years. (Even though as I have commented before I don’t agree with Kurzweil’s time frame on the completeness of the “solution”.)
A quote from Erik Larson’s “Ray Kurzweil's Impossible Vision” paper introducing the thread:
quote:
According to Kurzweil, gargantuan computing power (much more than currently available) is a necessary condition for realizing the Strong AI thesis. Blazing fast processors are necessary to run the massively parallel neural networks and .genetic. algorithms that will reproduce our minds.
Here clearly Erik Larson is recognizing that the mode of computation to accomplish Kurzweil’s vision would include “neural networks” and “genetic algorithms”. These are not subject to the limitations of “representational logic” or something like “first order logic”. So I don’t understand why limitations of such logic systems are being used as an argument that in the vision of a stronger AI system researchers will not make very significant progress in the “frame problem”. This is why I feel that the argument is a “straw man” argument. Erik Larson wants to reinforce that the presentation is about a limitation of these “logic” systems (especially “first order”), a point I can’t dispute -- but it is hardly relevant to whether AI systems will have these limitations.
The differences of “logic systems” and the developing methods of dealing with problem solving is showing how different our normal abilities at solving problems is from “theorem proving”. The Fuzzy Logic, Neural Networks, and Genetic Algorithms methods use approximate reasoning -- and thus avoid limitations of hard logic systems. One of the enduring problems of AI has been the “brittleness” of such hard logic systems, while a human can easily gloss over a mispeling for example. (SIC-Yes that was intentional). And reasoning by analog (a human trait for solving many problems) will be much enhanced over hard logic methods.
If the difficulty of the “frame problem” is a tendency to focus incorrectly on what the problem is, so as to not find a solution within a reasonable time (as the robot and bomb example suggests), then we are similarly committing a “frame problem” type of error to demand that it must be solved using the techniques of the 1960s. It is an error to claim that intelligent agency is “computation”, and it is an error to demand that creating AI systems should be based on “computational” methods. That Von Neuman machines might be used to execute neural network and other fuzzy systems does not define the problem type as “computation”.
In the post above, Erik Larson continues with some bullet remarks:
quote:
The idea that there are a "bunch of frame problems" seems to me nonsense. The frame problem is a type-level idea--there are many instances of the frame problem but one problem of this type.
If there are many “instances” of the “frame problem”, and some sub-classes of such problems are solved one way, and other sub-classes of such problems are solved another, then there are “separate problems” being solved with different methods. It does not matter to me if you want to classify them all as a single “frame problem” in generalization -- just don’t claim that researchers have to find a single solution to all its parts to create a highly functional AI system. (Remember that the brain consists of different sections that seem to operate somewhat differently, and operate on different problem types. Surely there are different instances of “frame problems” involved in each type of problem solving capacity which may have distinct “solutions” or methods that improve ability in that area.
I don’t deny that there is value in a potential overall solution to all instances of “frame problem”. An instance of such a generalized solution could include a learning system (possibly GA and Neural Network based) that learns to solve new types of “frame problems” piecemeal. Though there was one single systemwide “programming” solution, strength in dealing with the individual types of problems would be gained as “learning experiences”. The obvious claim could be that such “learning experiences” are really a form of “programming”, so that the claim could then be made that the system really has distinct “programming” solutions to each group of instances, and no overall solution to the “frame problem” has been found. This supports the fine print technical wording of the claim of the article, while the spirit of the argument that a very strong AI won’t be developed would be incorrect. Note my comparison to human learning from several of my posts and below -- if the standard is an ability which is somehow comparable to humans, you can’t use a standard that exceeds that of humans as the reference.
quote:
The idea that humans have their own "frame problem" supposes, of course, the non-technical or philosophical treatment of the problem. No problem. But it also commits itself to the idea that flaws in our cognitive abilities (i.e., in pattern recognition, belief-based reasoning, discrimination, etc) are somehow in the same boat as problems discovered about computational systems. At one level there are parallels: no person or machine is a perfect reasoner. Everyone--every known thing--fails at something. But the problem with this line of thought is that pointing to flaws is both obvious (and therefore not much help in addressing anything) and beside the point. Beside the point because the existence of "flaws" presupposes that there is some functioning, core capability. In people, this is true enough. But the problem with designing computational systems lies in finding some suitable functioning capability in the first place. Perhaps we can find one, perhaps not. At any rate human and machine cases are different.
Yes, it is a “non-technical” and somewhat “philosophical” treatment of the problem. This is to point out that the standard by which the performance of the AI system to be judged cannot be made higher than that of a human, so that some sort of fine print “technical” judgment is used as the criterion. To do so would be a “straw man” argument, and I assume the participants want a realistic standard of comparison and not just to win some sort of highly technical argument. Far beyond everyone failing at something, everyone learns gradually to do the logic and planning and behavior capacities that they develop over a long period. How well are humans at solving truly novel problems? I claim that humans have significant limitations, and learn to solve such problems over a long period of time -- including taking “serendipidous” information as hints. The history of invention is repleat with examples of this -- as the James Burk series “Connections” shows with some clarity.
The bottom line:
By specifying that the problem is in the domain of computational “representational logic systems”, and by specifying that there is only a single general “frame problem” to be entirely solved, Erik Larson has presented a technical thesis that is undoubtedly correct. However it is irrelevant to the issue of whether a future strong AI system will exhibit human-like characteristics and handily deal with the issues suggested by the “frame problem”.
[ 25. January 2003, 00:42: Message edited by: gedanken ]
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 25. January 2003 13:02
I wanted to expand on my comments with regard to something that Micah said, because I think this can be illuminating on how future very strong AI systems may be accomplished:
quote:
Gedanken, search strength is totally irrelevant to the frame problem. If the frame problem was related to search capabilities then we would be talking about the combinatorial explosion problem. We are not.
Repeating a quote of part of the “robot” story from Dennett:
quote:
Back to the drawing board. `The solution is obvious,' said the designers. `Our next robot must be made to recognize not just the intended implications of its acts, but also the implications about their side-effects, by deducing these implications from the descriptions it uses in formulating its plans.' They called their next model, the robot-deducer, R1D1. They placed R1D1 in much the same predicament that R1 had succumbed to, and as it too hit upon the idea of PULLOUT (Wagon, Room, t) it began, as designed, to consider the implications of such a course of action. It had just finished deducing that pulling the wagon out of the room would not change the colour of the room's walls, and was embarking on a proof of the further implication that pulling the wagon out would cause its wheels to turn more revolutions than there were wheels on the wagon - when the bomb exploded.
In this model, the ‘designers’ incorporated study of ‘the implications about [the proposed action’s] side-effects, by deducing these implications from the descriptions it uses in formulating its plans’.
So we have a very important question in understanding the nature of the problem. In the list of mathematically generated “side effects”, was the side effect that pulling the wagon out of the room would also carry the bomb with it one of the “side effects” that would have been analyzed?
Within this are a couple of very important distinctions. First, were there an infinite or are there a finite number of side effects? Clearly to be even remotely possible for the AI to analyze these exhaustively, the side effects to be studied must be limited to a finite set. So would the problem here of the bomb being carried with the wagon appear in the finite set of all side effects that the AI could possibly analyze, by the coding method that would be used? Answering this question is part of the issue of basic “ability” aside from computational power.
Because if computational power were unlimited, then the only issue would be whether the issue in question was represented somewhere in all the possible side effects! If it were somewhere in that set, then the AI would have noticed the problem and potentially solved it.
The “blocks world” planning problem is a very old problem. (See example here.) If the aspect of the bomb being on the cart had been recognized as a side effect, the AI could easily have made a plan to remove the bomb from the cart before moving the cart -- and this is by way of very old planning techniques.
But what was the description of what happened? “It had just finished deducing that pulling the wagon out of the room would not change the colour of the room's walls, and was embarking on a proof of the further implication that pulling the wagon out would cause its wheels to turn more revolutions than there were wheels on the wagon - when the bomb exploded.” In other words, the AI was searching a lot of side effects that were not relevant, and ran out of time.
So there are two possible answers to the question of whether the new problem could be found in the finite set of ‘all side effects’ that could be analyzed -- either yes or no. If the answer was ‘yes’, then the issue is clearly one of search time. If the answer was no (or the number of side effects was infinite in the representation), then we have a problem of representation.
But why does the search time take so long? The reason is simple. Combinatorial explosion! (Micah, the aspect of combinatorial explosion was not made explicit, but was partially implicit in what was stated.)
So with the answer of “yes” that the new problem could potentially have been properly analyzed in the AI’s system of representation, then the “frame problem” becomes a problem of “focus”. It becomes a problem of focus precisely because of combinatorial explosion prevents all but almost infinite computational capacity systems from analyzing all the irrelevant consequences in any exhaustive way so as to find the relevant consequences.
Now what if the number of side effects was infinite in the representation used? Then clearly we have a limitation of ability, not simply a limitation of combinatorial explosion. But this does not mean that the issue of “focus” has changed -- that the “frame problem” is no longer really the “focus” problem. Because if the consequences of moving the cart had been selected in a “focused” set of consequences to analyze, out of a potentially infinite set of such consequences, it could still have been found and planned for as per “blocks world” examples.
So we have the last issue, were the important aspects totally outside of the possible representation of the problem? Was there no possible recognition that the bomb in the cart would travel with the cart and thus cause further problems, given the representation of knowledge given? Was there no possible solution within the range of motions that the robot could achieve, of picking up and moving objects? (Now clearly the original description was not intended to be a statement that the solution was outside of all possible motions the robot could achieve, or else the whole exercise would have been meaningless as a comparison to human thinking capacity.)
If the “frame problem” is intended to be a statement that the issues to be solved are not even found within the representation of the physical problem that can be sensed in the novel case by the AI, then we have a very different question from the way the question has been presented. In that case, we are asking about the fundamentals of problem representation -- not about the fundamentals of solving those problems. That limitation will also apply to humans.
If the human didn’t understand that separation of himself from the bomb was necessary, he would not act to provide that. (And that comes from experience and learning.) If the human didn’t understand that he could lift and move objects, and that has an effect on how things operate, then the human could also not solve this problem.
I think that humans do not do the elaborate planning operations that are embodied in the predicate calculus format that is used in many AI representations. This is why scheduling systems can outperform human schedulers in finding technical solutions within the given rules, and more quickly, because these sort of predicate calculus calculations can be performed by rote in a machine must faster than any human. That is why posing the issue in terms of solving predicate calculus problems is a mistake at the outset.
I realize that posing the questions this way has been an integral part of the AI community’s methodology, and that discussion of the “frame problem” in those terms has occupied significant intellectual effort of the AI community. That does not mean that the AI community is correct or incorrect in working on this problem in this manner. Indeed humans do a limited amount of planning which resembles vaguely some limited set of predicate calculus decision analysis.
When I go to perform a task, I don’t think through all the implications -- any more than the AI should. Rather I use experience to guide me to a general attempt at solving the problem. I may mentally “simulate” or visualize the consequences of my actions -- and this ability is similar to the predicate calculus search. But the “visualization” or “simulation” of the actions is a vastly different process from the logic formalism of the predicate calculus statement of the problem -- and this could be a process that is implemented by a form of neural network and other system structure.
When I go to perform a task, I usually start with what I know I need. Consider Dennett’s midnight snack example. But suppose I am a kid in the middle of the day, and I’m trying for the first time to construct a sandwich.
I make motions like I have seen my parents make, with moving ingredients to the bread. I learn that layers can be stacked on the bread (in a general way). I re-learn something that I learned form my days in a crib, that usually things that are on top of other things move together with those things they are on top of. But once I have this experience, it is a template that can be called up in some sort of anagogic reasoning in the future. (The visualization of the analogy is some sort of parallel search problem, most likely -- once again something that could be amenable to a neural network type of representation for parallel search).
As I make the sandwich, I don’t necessarily follow a set order of ingredients. As I move them to the bread, I may view mistakes -- like the pickle falls off. So I make corrections by a feedback mechanism -- I have a vague “visualization” of the sandwich ingredients in a stack (goal), and when the pickle falls off I make a correction. This correction was not a matter of properly representing the planning operation in a predicate calculus system, rather it was a matter of a vastly greater complexity system for representing vague notions. The aspect of limited planning (as in the predicate calculus) is simply an aspect that the researcher must recognize as a partial functionality of the neural network system. The knowledge gained form the predicate calculus type of planning system is useful as a guide for the research, but the system should not be expected to do “deep planing” in a manner that is directly comparable to the predicate calculus approach.
The robot AI could have started to move the cart, and then noticed that the bomb was moving with the cart. That could provide feedback that helps focus the attention on the needed planning -- and that is a much more local and learning-based and feedback based approach than trying to get the AI to be able to formulate the plan in whole in predicate calculus form without mistake and without experience in the novel problem.
This is why I think that the issue of how the future AI systems will be implemented is very important. Remember Kurzweil’s comments on how these problems will be solved:
quote:
According to Kurzweil, gargantuan computing power (much more than currently available) is a necessary condition for realizing the Strong AI thesis. Blazing fast processors are necessary to run the massively parallel neural networks and .genetic. algorithms that will reproduce our minds.
Whether or not they will come close to “reproduc[ing] our minds”, I am confident that stronger and more functional AI systems will develop in the future. And they will be able to make use of concepts along the lines of the neural networks, genetic algorithms, and future concepts that are spawned from these. The limitations will not be those of “predicate calculus” representation. This is what people need to understand to begin imagining how these future systems will be able to produce responses that have the appearance and function that we normally associate with intelligence. This is what people need to understand to start to visualize solutions to the limitations that the “frame problem” appear to create.
IP: Logged
|
|
|
Printer-friendly view of this topic