Manage updates with the Download App
- Related Stories
-
Photo galleries: Robots in action
March 16, 2007 -
This is your brain on a microchip
May 11, 2006 -
Blueprinting the human brain
May 10, 2006 -
The best of times in science and tech
April 3, 2006 -
Stanford wins $2 million in robotic car race
October 9, 2005 -
FAQ: Keeping pace with robots
October 5, 2005 -
Ray Kurzweil deciphers a brave new world
September 29, 2005 -
Your brain: Search engine, or calculator?
June 29, 2005 -
IBM computing algorithm thinks like an animal
March 22, 2005
(continued from previous page)
Could you elaborate on that--on nonmonotonic reasoning?
McCarthy: OK. In ordinary logical deduction, if you, say, have a sentence P that is deducible from a collection of sentences--call it A--and we have another collection of sentences B, which includes all the sentences of A, then it will still be deducible from B because the same proof will work. However, humans do reasoning in which that is not the case. Suppose I said, "Yes, I will be home at 11 o'clock, but I won't be able to take your call." Then the first part, "I will be home at 11 o'clock,"--you would conclude that I could take your call, but then if I added the "but" phrase, then you would not draw that conclusion.
So nonmonotonic reasoning is where you draw a conclusion, which may be a correct conclusion to draw, but it isn't guaranteed to be true because some added facts may prevent it. Now, that was around 1980, or a little bit before, that formalizing nonmonotonic reasoning began, and it's turned into a fairly big field now.
What would be the biggest achievements in the last 50 years? Or how much of the original goals were accomplished?
McCarthy: Well, we don't have human-level intelligence. However, I would say driving the car 128 miles shows a considerable advance. (Editors' note: In last fall's DARPA Grand Challenge, the winning vehicle--Stanford's robotic car, "Stanley"--drove itself 131.6 miles across the Mojave Desert.)
What's the next big thing, then, to accomplish?
McCarthy: I would like to see further progress in formalizing commonsense knowledge and reasoning, taking context into account. That's something I've been working on for a long time and that some other people also work on, and which DARPA supports, but I think the ideas that are available are not sufficient to reach human-level intelligence.
A goal in AI is not so much to make machines be like humans, having human intellectual capabilities, but to have the equivalent of human intellectual capabilities, correct? In other words, not reinventing the human but creating something that thinks similar to humans and surpasses human thought?
McCarthy: That's the way I see the problem. There are certainly other people who are interested in simulating human intelligence, even aspects of it that are not optimal. In particular, Allen Newell and Herbert Simon tended to look at it that way.
Another sort of high-level goal that may or may not be reachable seems to be to try to program originality into machine thinking.
McCarthy: Yes. That would be worth some efforts. I did something that was so to speak part way to that in 1963, in which I talked about a creative solution to a problem, a solution that involved elements that were not in the problem, the statement itself. But that was just a start.
And originality--is that as simple as trying to introduce some randomness into the programs, or was it a different order of magnitude?
McCarthy: Well, in principle, in a logical system, you could generate sentences systematically or randomly...and any idea would eventually turn up, but the "eventually" is likely to be extremely far in the future. So that hasn't done much, either using randomness or otherwise. What's needed is to figure out good ways of constructing new ideas from old ones.
Going back for a second to the notion of having machine capability versus programming and the right source of ideas--today we have so much more computational capability than was available 50 years ago. What difference is that making, with the state of the art of computer chips and memory these days?
McCarthy: I would say that 50 years ago, the machine capability was much too small, but by 30 years ago, machine capability wasn't the real problem.
The real problem still being the basic ideas?
McCarthy: Yes.
How do robots factor into thinking about artificial intelligence? I guess in the popular vision (in movie images of humanoid robots), that's where people would tend to see human-level intelligence, but are robots a real factor, or does it really matter what shape or form the machine takes?
McCarthy: Certainly, robots present some problems. That is, they have to operate in an environment, and some of the even rather elementary problems have not been solved yet--that is, combining the ability to walk the way a human walks, which is falling forward rather than just shuffling, and with the ability to understand a three-dimensional scene and so forth. These ideas have been worked on sort of separately, but there still isn't a robot that could move around confidently in a cluttered room and climb stairs, let alone climb trees.
See more CNET content tagged:
Artificial Intelligence, conference, intelligence, idea
33 comments
Join the conversation! Add your comment
think better than we do. That wouldn't seem hard.
Next, they could program unmanned aerial vehicles to see and
avoid other air traffic, just like a live pilot does. So far, they are
dumb and blind, and so a danger to all other air traffic.
think better than we do. That wouldn't seem hard.
Next, they could program unmanned aerial vehicles to see and
avoid other air traffic, just like a live pilot does. So far, they are
dumb and blind, and so a danger to all other air traffic.
think better than we do. That wouldn't seem hard.
Next, they could program unmanned aerial vehicles to see and
avoid other air traffic, just like a live pilot does. So far, they are
dumb and blind, and so a danger to all other air traffic.
The only way that we are ever going to create an intelligent machine that approaches the level of thought found in the human mind is through pulse coded neural networks used to implement positive feedback cortical planes. Current Von-Neuman machines don't have the power to emulate a neural network containing billions of neurons with trillions of interconnectsion - maybe they never will.
The problem is that current AI research still revolves around how to get Von Neuman machines to act intelligently. This is moronic! Its like looking for your eyeglasses out by the streetlight when you lost them in a dark alley because the light is better there. It just ain't going to happen! Of course, I am not saying that such research is useless - that would also be moronic. It just won't lead to a concious machine.
If you want a concious machine, there is no alternative, but to create specialized hardware to emulate the neural network. Three dimensional circuitry implemented with proximity field effect gates might work. However, this is a chicken and egg problem because there will be little motivation to develop the proper circuitry since you get so much bang for the buck with von neuman machines.
Of course if you had enough Von Neuman machines that could handle trillions of interconnections maybe it would be possible to make a concious machine. However, this would be prohibitively expensive --- wait, what about the internet....
The problem is that even if we had the specialized hardware you propose, we wouldn't know what to do with it. We understand too little about the problem. We don't even uderstand what questions to ask. Technology normally evolves in small incremental steps, but we still don't know what the incremental steps to get from "Elisa" to "HAL". It is today just a big step from Simulated Intelligence to Artificial Intelligence.
Since Von Newman machines have proven to be so good at emulating anything they allow us to emulate massively parallel systems in a cheap lab. There's nothing, NOTHING, a neural net or parallel machine can do that a Von Newman cannot do slowly. Maybe they are orders of magnitude too slow to do any practical AI, buit that doesn't matter. For research purposes slow AI is the same as AI, so if we can use the VN machines to emulate a machine that can, in a few months of computation, deduce something simple that requires actual intelligence, it would be a great step.
If we'd rather concentrated on creating specialized machines we would instead be spending millions and probably some of the best brains on earth in building thousands of generations of machines that don't work and that are probably trying to solve problems we don't understand, machines that could have been emulated, even if slowly, by other machines we already have.
Once we get Von Newman machines to do some intelligent things we would have achieved something: understanding some of the fundamental questions and mechanisms. And then, only then, investing in designing a platform to do that at a bigger and faster scale, is worth the effort.
Why waste billions on designing hundreds of generations of machines to do tasks we don't even understand?
It is better to rely on emulation until we understand the questions, until we can say "this is what I want the machine to do", and then we can move our efforts into building efficient and scalable machines.
Speed is not of the essence. Slow AI is still AI.
Today we don't know how to go from Elisa to Hal. We don't have a roadmap or a path, we have no incremental steps to go by. Building extremely capable hardware such as your proposed neural nets is not going to help us get there.
Once we've built emulations that can show some signs of being on that path it is worth to concentrate on building specialized hardware to do practical implementations.
The only way that we are ever going to create an intelligent machine that approaches the level of thought found in the human mind is through pulse coded neural networks used to implement positive feedback cortical planes. Current Von-Neuman machines don't have the power to emulate a neural network containing billions of neurons with trillions of interconnectsion - maybe they never will.
The problem is that current AI research still revolves around how to get Von Neuman machines to act intelligently. This is moronic! Its like looking for your eyeglasses out by the streetlight when you lost them in a dark alley because the light is better there. It just ain't going to happen! Of course, I am not saying that such research is useless - that would also be moronic. It just won't lead to a concious machine.
If you want a concious machine, there is no alternative, but to create specialized hardware to emulate the neural network. Three dimensional circuitry implemented with proximity field effect gates might work. However, this is a chicken and egg problem because there will be little motivation to develop the proper circuitry since you get so much bang for the buck with von neuman machines.
Of course if you had enough Von Neuman machines that could handle trillions of interconnections maybe it would be possible to make a concious machine. However, this would be prohibitively expensive --- wait, what about the internet....
The problem is that even if we had the specialized hardware you propose, we wouldn't know what to do with it. We understand too little about the problem. We don't even uderstand what questions to ask. Technology normally evolves in small incremental steps, but we still don't know what the incremental steps to get from "Elisa" to "HAL". It is today just a big step from Simulated Intelligence to Artificial Intelligence.
Since Von Newman machines have proven to be so good at emulating anything they allow us to emulate massively parallel systems in a cheap lab. There's nothing, NOTHING, a neural net or parallel machine can do that a Von Newman cannot do slowly. Maybe they are orders of magnitude too slow to do any practical AI, buit that doesn't matter. For research purposes slow AI is the same as AI, so if we can use the VN machines to emulate a machine that can, in a few months of computation, deduce something simple that requires actual intelligence, it would be a great step.
If we'd rather concentrated on creating specialized machines we would instead be spending millions and probably some of the best brains on earth in building thousands of generations of machines that don't work and that are probably trying to solve problems we don't understand, machines that could have been emulated, even if slowly, by other machines we already have.
Once we get Von Newman machines to do some intelligent things we would have achieved something: understanding some of the fundamental questions and mechanisms. And then, only then, investing in designing a platform to do that at a bigger and faster scale, is worth the effort.
Why waste billions on designing hundreds of generations of machines to do tasks we don't even understand?
It is better to rely on emulation until we understand the questions, until we can say "this is what I want the machine to do", and then we can move our efforts into building efficient and scalable machines.
Speed is not of the essence. Slow AI is still AI.
Today we don't know how to go from Elisa to Hal. We don't have a roadmap or a path, we have no incremental steps to go by. Building extremely capable hardware such as your proposed neural nets is not going to help us get there.
Once we've built emulations that can show some signs of being on that path it is worth to concentrate on building specialized hardware to do practical implementations.
The only way that we are ever going to create an intelligent machine that approaches the level of thought found in the human mind is through pulse coded neural networks used to implement positive feedback cortical planes. Current Von-Neuman machines don't have the power to emulate a neural network containing billions of neurons with trillions of interconnectsion - maybe they never will.
The problem is that current AI research still revolves around how to get Von Neuman machines to act intelligently. This is moronic! Its like looking for your eyeglasses out by the streetlight when you lost them in a dark alley because the light is better there. It just ain't going to happen! Of course, I am not saying that such research is useless - that would also be moronic. It just won't lead to a concious machine.
If you want a concious machine, there is no alternative, but to create specialized hardware to emulate the neural network. Three dimensional circuitry implemented with proximity field effect gates might work. However, this is a chicken and egg problem because there will be little motivation to develop the proper circuitry since you get so much bang for the buck with von neuman machines.
Of course if you had enough Von Neuman machines that could handle trillions of interconnections maybe it would be possible to make a concious machine. However, this would be prohibitively expensive --- wait, what about the internet....
The problem is that even if we had the specialized hardware you propose, we wouldn't know what to do with it. We understand too little about the problem. We don't even uderstand what questions to ask. Technology normally evolves in small incremental steps, but we still don't know what the incremental steps to get from "Elisa" to "HAL". It is today just a big step from Simulated Intelligence to Artificial Intelligence.
Since Von Newman machines have proven to be so good at emulating anything they allow us to emulate massively parallel systems in a cheap lab. There's nothing, NOTHING, a neural net or parallel machine can do that a Von Newman cannot do slowly. Maybe they are orders of magnitude too slow to do any practical AI, buit that doesn't matter. For research purposes slow AI is the same as AI, so if we can use the VN machines to emulate a machine that can, in a few months of computation, deduce something simple that requires actual intelligence, it would be a great step.
If we'd rather concentrated on creating specialized machines we would instead be spending millions and probably some of the best brains on earth in building thousands of generations of machines that don't work and that are probably trying to solve problems we don't understand, machines that could have been emulated, even if slowly, by other machines we already have.
Once we get Von Newman machines to do some intelligent things we would have achieved something: understanding some of the fundamental questions and mechanisms. And then, only then, investing in designing a platform to do that at a bigger and faster scale, is worth the effort.
Why waste billions on designing hundreds of generations of machines to do tasks we don't even understand?
It is better to rely on emulation until we understand the questions, until we can say "this is what I want the machine to do", and then we can move our efforts into building efficient and scalable machines.
Speed is not of the essence. Slow AI is still AI.
Today we don't know how to go from Elisa to Hal. We don't have a roadmap or a path, we have no incremental steps to go by. Building extremely capable hardware such as your proposed neural nets is not going to help us get there.
Once we've built emulations that can show some signs of being on that path it is worth to concentrate on building specialized hardware to do practical implementations.
one) first requires understanding our own evolved intelligence. To
do that we need to start over. See "Wholly holistic evolution, Mr.
Darwin" at <a class="jive-link-external" href="http://we.karleklund.net" target="_newWindow">http://we.karleklund.net</a>.
one) first requires understanding our own evolved intelligence. To
do that we need to start over. See "Wholly holistic evolution, Mr.
Darwin" at <a class="jive-link-external" href="http://we.karleklund.net" target="_newWindow">http://we.karleklund.net</a>.
one) first requires understanding our own evolved intelligence. To
do that we need to start over. See "Wholly holistic evolution, Mr.
Darwin" at <a class="jive-link-external" href="http://we.karleklund.net" target="_newWindow">http://we.karleklund.net</a>.