Slashdot Mirror
AI Going Nowhere?
jhigh writes "Marvin Minsky, co-founder of the MIT Artificial Intelligence Labratories is displeased with the progress in the development of autonomous intelligent machines. I found this quote more than a little amusing: '"The worst fad has been these stupid little robots," said Minsky. "Graduate students are wasting 3 years of their lives soldering and repairing robots, instead of making them smart. It's really shocking."'"
He has a complete disregard for the question of where the AI engine will run. If an AI is to be of any more use than a curiousity then those "little autonomous robots" must function in a viable manner so that the AI has something to do when it comes to "life".
I understand his frustration in general progress. But, those grad students are building a strong foundation for their later work that may very well meet the goals he is espousing. No need to have design flaws in implementation down the road because the engineer wasn't properly educated in physical design as well as logical design.
-Rusty
The Master (Angelo Rossitto) in Mad Max Beyond Thunderdome, "Not shit, energy!"
Yes, the man is quite brilliant, and possibly the most important voice in the field. That being said, he's also a self-important jerk. Intelligent Systems (what people in the field call AI) aren't where *he* thinks they should be, and he regularly complains about it.
Humans measure intelligence by gauging replys to questions that have quantified answers. Giving an advanced computer a IQ test is sinply a matter of recalling the appropriate information from memory to answer. The true form of AI isnt about intelligence, its about reason. But how do you teach a computer to respond with an answer to a question that the computer has never encountered before... When we build a machine that can answer a question based on incomplete imput we will have made the first step in creating a machine that can "think"
reply to MYCROFTXXX@lunaauthority.com
Fire in the hands of the village idiot is no tool, but a weapon of mass destruction
How about we more thoroughly study and understand how human intelligence operates before we even presume to design something that imitates or rivals it in depth and complexity.
Come off it, stop making escuses. Robitics and AI are two entirely different branchs. AI is almost pure math and computing, and robotics is an engineering discipline. Why are AI researching building little robots?
Building any form of AI system is not easy, but copping out of it by building toys is not the answer. We already have platforms for AI; they're called line terminals. Things like pattern matching do not require a fully autonomous robot, after all.
Minsky is right; whats new to come out of actual AI research in the last 30 years?
So true. Minsky's Good Old-Fashioned AI (GOFAI) has been a dead field since the 70s, after they figured out that getting a computer that could move blocks around in an idealized simple world was not a small first step, and Eliza showed them how easy it was to conflate intelligent with clever. The "successes" they had towards AI were, as one author has written, like climbing to the top of a tall tree and claiming you've made progress getting to the moon.
Now Minksy, never wanting to admit his life's work has been a dead end, comes out saying that it's all these other researchers working in other directions that are at fault for there being no progress. I imagine he believes that if only they'd all climb the tree with him, the trip to the moon could really start.
"As soon as we solve a problem," said Pollack, "instead of looking at the solution as AI, we come to view it as just another computer system."
;)
This is the most interesting comment to me. Because we understand the nature of the process that produces supposedly 'intelligent' results, (and we don't understand the same process in ourselves), we perhaps rightly just view the resulting system as just an application.
Seems like Minsky is throwing all his toys out of the pram because he doesn't want to admit to what everyone else has been saying for a while: that whether a computer can think is at best an astonishingly difficult question to answer and at worst meaningless. I'm a grad student who's just spent a year looking at computational linguistics and semantics (amongst other things), and the most debilitating restriction on the semantic side of things is the problem of so-called 'AI-completeness', which essentially says that if you solve this problem you have a, externally at least, thinking computer. Really simple things like anaphora resolution are AI-complete in the general case. If we could have solved this problem by now, I think it's fair to say we would have done, given its massive importance. However, we know that the brute-force case is ridiculously intractable, and we can't figure out how to do it any more cleverly. Roger Penrose argues that this is due to the fundemental Turing-style restrictions that we place on our notion of computing. Until we get a paradigm shift at that level, we're likely never to solve the general case.
And I'm sure that Minsky is aware that attempts to solve constrained domain inference and understanding have been taking place for a good long time now. I just don't see why he's so upset that the field of 'AI' (which is a nebulous catch-all term at best) has shifted its focus to things that we stand a cat in hell's chance of solving, and that have important theoretical and practical applications (viz. machine learning). Replicating human thought is not the be-all and end-all, and you can argue that it's not even that useful a problem.
Robots, though, I agree with. Can't stand the critters
Henry
i don't do sigs. oops.
I very much enjoy the works of Markram and Tsodyks. What they mainly analyze is how two nerve cells can interact with each other. They showed how they change their connection weights and how the timing of spikes, nerve impulses, affect how neurons connect to each other and how they transmit information.
While these studies tell us a lot about the underlying biology they do not tell us what these modes of information transmission are used for. For years it had been known that synapses have complex nonlinear properties. Biology pretty much does not constrain what functions neurons compute.
Thats why I do not believe that such studies will bring us nearer to real AI anytime soon. The algorithms coming from these systems are severely underconstrained. A lot of modelling has followed the pioneering works of Markram and Tsodyks, one of them being Maas. All these algorithms are very fascinating and might yield insight into the functioning of the nervous system.
The algorithms however are lightyears from being applicable to real world problems. The field of AI is old and in some sense quite mature. None of the "biologically inspired" algorithms today can compete with state of the art machine learning techniques.
Googlefight "Slashdot Troll" against "BSD is dying" 303:229. BSD thus cant die.
* One might claim LTP or LTD as some sort of neuronal knowledge. Ok, that's fair, but my point stands if you apply it to the building blocks of neurons. Do ion channels "know"? Do amino acids? It's turtles all the way down.
Give me Classic Slashdot or give me death!
That's all there is to it. Minski is just sore that his theories from 30 years ago aren't proving themselves, and the decentralized models being implemented by his rivals at MIT (e.g. Rod Brooks and his graduate students) are demonstrating remarkably sophisticated behaviors and advancing the state of the art.
It can pick me out in a crowd, and it can show a number of emotions, such as surprise, anger, and boredom.... yawn.
to an extent yes it has decent pattern recognition. can it pick you out from the rear? no. side? no.
Can it simulate and fool you into thinking it is showing emotions ? yes. is it anytihng but an expensive toy? no.
the Abio is amazing, but it hardly does what people think it does. and that is the key with the abio.. it does alot of things that fool humans quite well.
it will get better, but it is hardly near AI material.
Do not look at laser with remaining good eye.
Ask some 3-year-old kids which way is up, and they will all know, but they won't be able to define it. Yet, since computers don't have bodies, they don't have anything like the semicircular canals that we have, which act as gyroscopes and give us an intuitive, non-intellectual sense of which way is up.
Trying to program intelligence purely with software puts the researcher at a disadvantage, since even the most fundamental rules and attributes of things (fire is hot, water is wet) have to be explicitly entered as constant variables.
Once robotics advances to the point where mobility, vision, and speech recognition can be taken for granted, then AI can be programmed as an add-on.
Body first, mind second - That's how animals evolved on this planet, and it's how, I believe, Rodney Brooks approaches this field.
every stain tells a story
"Graduate students are wasting 3 years of their lives soldering and repairing robots, instead of making them smart. It's really shocking."
Yeah, much more shocking than the -- decades -- he (and others in the 'hard AI' camp) have been spending? They've made oh-so-much more progress, haven't they?
Rodney Brookes made more progress with his robots in the early nineties than the whole hard AI camp did in 3 decades. I remember seeing a documentary once comparing this huge robot which used a traditional procedural program to navigate through a boulder-strewn field. It took about 3 HOURS to decide where to put its foot next. Meanwhile, little subsumption architecture-based robots were crawling around like ants, in real-time. (Oh, and some of them had to learn to walk from first principles every time they were turned on -- only took about half an hour!) That's the most damning evidence of the failure of hard AI I can think of.
As others here have said, what good is a brain until we get a useful BODY working? Manueverability and acute senses are a must before an artificial intelligence can do anything useful, or learn from its environment effectively.
ERROR 144 - REBOOT ?
If Marvin wants to know why AI hasn't made major strides in the past 30 years (and, by the way, I would say that it has) he should look no further than his own bullying, arrogant approach.
Promising subfields like perceptrons were intentionally quashed by him... he went out of his way to strangle investment and research in areas he considered to be a dead end. We're not literature majors: we can't just all say the same thing in a party over wine and cheese and call it progress.
Even bad ideas, when well explored, can give new meaning and better approaches to a field. And since this is research, noone knows the correct answer: even a dumb-seeming idea may turn out to be the right one-- or give us clues about features the right answer needs to have.
Of course we've had major advances in AI. One of the challenges of AI, as the article points out, is that once something is well understood, it is defined as being outside the AI field. Computer vision, face recognition, voice and speech recognition. Conversation engines like SmarterChild. No, this isn't HAL, but they are good, positive steps in the right direction.
It was still pretty much explicitly programmed to do those things. It's not really bored, it was just programmed to act bored, etc. Even the image recognition is a testament to the intelligence of the programmers, but not really to the AIBO itself.
Please mod the parent up.
:-(
I am an AI researcher and the parent poster is speaking truthfully.
The main challenges in AI at the moment do not concern building the physical robots -- e.g. a piece of kit on wheels with IR sensors or such things.
The main challenges in AI concern applying some very complicated math to solve problems like pattern recognition, density estimation and other forms of machine learning.
It seems to me that a large number of AI PhD students spend their lives tinkering with the mechanics and electronics of the robots that will ultimately be used to test their algorithms. This is wasted time; a good electronics graduate should be able to do the tinkering, it shouldn't require a prospective AI PhD student to do it.
I can see the point in the PhD student learning a little about the hardware that they want to run their algorithms on (so that they know the limitations and common problems with real hardware), but they should not spend all their time doing that and wasting the opportunity to spend their time contributing to their field (i.e. AI, not mechanics or electronics).
That said, many AI labs do not have the funding to be able to pay full time hardware technicians, so in many cases the PhD student *has* to do the tinkering
"The noble art of losing face will one day save the human race"---Hans Blix
Before you yell flamebait or troll, let me explain.
I have been following the progress of various AI technologies, including neural nets and adaptive logic networks, for many many years now. Years ago perceptions were first developed and it was shown that they could learn simple patterns. Perceptrons were basically two layers of software simulated neurons. They worked, and researchers were fascinated and worked on them regularly.
Minsky, being the "highly regarded" and "leader" in AI, wrote a paper that proved that these perceptrons could never learn more complicated patterns, and threw a bunch of math at the reader. So people stopped. After all, there was a mathematical proof that perceptrons weren't going anywhere. Research skidded to a halt for decades because of Minsky.
Of course, then someone developed the (gasp!) THREE layer perceptron/neural net and sure enough with the right formula it could learn much more complicated tasks.
Minsky, in my opinion, does this regularly. The problem is, that he has a reputation in the industry as being a leader (I'm not sure why).
He's already lost us two or three decades of research because of his "leadership" -- I am terrified that he might cost us more development into the future.
Where could we have been if Minsky wasn't always going around half cocked, screaming that he is right? "Robots are useless!" is history repeating itself and him trying to get more press. Keep developing guys, just ignore the peanut gallery. There's always someone who says it can't work (ahem, Minsky) -- it can and it will.
The baby's fine -- please stop sending business cards.
-----
Free P2P Backup, Windows & Linux
It can pick me out in a crowd, and it can show a number of emotions, such as surprise, anger, and boredom.... yawn.
My dog can do the same thing.
It's not a dead end. He is complaining that there seems to be no interest in studying the big picture. We are focusing in solving specific problems and creating "smart" systems that operate on a set of rules, maybe "learning" on the way to enhance the predefined rules originally installed by the creators.
But the real AI comes when you create a 'stupid' system that is able to become smart through learning and training. He feels dissapointed because nobody (or almost nobody) is focusing in this direction.
You have smart toys ala Aibo, and smart systems ala Eliza, and a lot of people is working towards creating smarter toys and smarter systems, but the real breaktrough will come when somebody manages to create the dumbest system possible.
Life isn't like a box of chocolates. It's more like a jar of jalapenos. What you do today, might burn your ass tomorrow.
"The problem with that is that no one really *knows* how the brain works beyond a very, very basic and limited understanding. No one has ever been able to satisfactorily create/reproduce one. There's more going on than just synapses in there, that much most scientists can agree on. What they don't agree on is *what* else is going on in there."
That's what Minsky is getting at. Few people are working on that problem.
Research talent in universities seem to be striving for business solutions. But IMO, such research should be primarily done by businesses, not AI labs. Universities should create new science.
So Minsky has an opinion, he expresses it, he provides data, calculations and evidence to support it.
People just accept it, and progress is delayed.
Why is it his fault that there are so many followers? If anything is to be blamed is that these researchers just blindly follow whatever he's saying rather then take a good critical look at what is going on.
If his math and theory "proved" that an area of AI was a dead end, and it wasn't, his math/theory was wrong. It is a sad state when nobody dare challenge the status quo.
What Minsky is actually flaming about here is the damage done to the field by the original "AI winter" and the real possibility of a new AI winter starting in a few years.
;-)
.sig below).
"AI winter" is the name given to the collapse of strong AI as a business model in the mid 80's - expert systems and symbolic AI in general didn't deliver on their promises, and so the money went away. As a guy who got his doctorate in AI in 1985, I can tell you all about it.
One of the major causes of AI winter was researcher hubris - lots of people hacked up systems that appeared to solve 80 percent of certain complex problems and then said "all that stands between us and a complete solution is money and time". For many of those systems, solving the last 20 percent would have taken 2000 percent of the time, if it could have been done at all. The tragedy of AI winter, though, is that basically all of symbolic AI was abandoned, though some of it is creeping back out into the light with obfuscated syntax (see my
What Minsky sees here is a lot of people heading down the same path, but with neural nets and small robots instead of expert systems. The new systems are doing some interesting things relative to the old symbolic AI systems (though they do have the advantage of 20 years of Moore's law to help them). But, will they scale up? Right now, nobody knows. If they don't, the last thing the field needs is another cycle of overpromise/underdeliver/abandon.
Maybe AI is just plain hard, and cracking it will take longer than one or two computer industry business cycles.
To a Lisp hacker, XML is S-expressions in drag.
Minksy belongs to the old school of AI thinking. These guys believe that it is possible to make statements about intelligence itself, without considering the interactions of the organism/agent with its environment or the underlying architecture of the brain/CPU. I think that the total failure of this style of thinking to produce anything interesting in 50 years proves that this approach is sterile. Minsky laments the fact that graduate students build robots, but this activity exposes students to the challenges of constructing a device that must actually interact with the environment. It is ridiculous to assume that you could design a system capable of intelligent behavior without ever confronting the problems of sensors and actuators. Almost every part of the brain is devoted to processing raw sensory input or generating motor output. One cannot simply design an intelligent system without worrying about sensory input and behavioral output. The CYC project of Lenat has the laudable goal of teaching a machine "common sense" by hard coding a vast database of simple statements like "Trees cannot walk". This is a totally wrongheaded approach to learning and reasoning, and is typical of old school, hard AI.
We will only make progress in engineering intelligent, adaptive systems by studying actual examples of intelligent, adaptive systems, namely animals. Neuroscientists and psychologists are beginning to embrace the tools of mathematical modeling and simulation wo help explain nervous system structure and function. Computer scientists would do well to similarly embrace the work of experimental neuroscience.
Minsky is a dinosaur.
What's the difference between thinking and fooling people into believing you're thinking?
The same difference between cheating on a test in a subject you've never taken and actually knowing the material: as soon as you're asked a question that isn't on your crib sheet, the charade is over.
An AIBO has a pre-programmed set of behaviors, and any stimulus it isn't programmed to respond to will not have a realistic effect. The same is true of Loebner Prize winner Alice -- the only difference is the size of the pre-programmed response database. Ask it something it's never heard, and it will choke.
When these AIs are able to produce convincing responses to new stimuli, then I'll say that the difference between "fooling" and actually thinking has become irrelevent.
The enemies of Democracy are
You're making the same leap of faith that most people make (and I now think is incorrect): The human (or Dog) mind is a biological computer therefore an IBM with the right software is also a mind in the same sense.
It doesn't work this way, and yes, there is a difference. Having an outward appearance of intelligence is not enough to show intelligence. Read Searle and Block's discussions on the Chinese Room argument -- it's a fascinating and eye opening read (I think it was Block that -- quite convincingly, IMHO -- makes the case that most of our intelligence is innately biological, and "strong AI" not even possible with what we know today).
IMHO one of the problems with AI is that we don't even know what human intelligence is, and until there is a fundamental advance (not technological but in our understanding of our human/biological mind) then it seems to me the most we can hope for are machines that mindlessly ape intelligent behavior, but are not intelligent in any but very superficial ways or by very loose definitions.
Something that mimics the outward appearance of intelligence is a far cry from what, hopefully we'll be capable of in the (probably still distant?) future.