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."'"
It can pick me out in a crowd, and it can show a number of emotions, such as surprise, anger, and boredom.... yawn.
"There are more pleasant things to do than beat up people." --Muhammad Ali
It's not the AI which is going nowhere. It's the traditional approaches to AI such as Minsky's symbolic logic which are not going anywhere. Seach google for Henry Markram, Maass, Tsodyks. Their research seems very promising.
> You'll never have real, true intelligence
:-)
Define "real, true intelligence"
> You can try to simulate that, but so far
> simulation consists of what amounts to a
> gazillion 'if' tests
That's what tradiditonal AI school is doing. Yes, you are correct. It won't go anywhere. On the other hand spiking neural networks are very promising. Search google for "liquid state machine". These researches are making progress novadays, not Minsky.
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.
My journal has hot
What they need, then, is for an engineering student to do their masters dissertation on creating a generic physical framework for AI systems, or a computing student to do theirs on a generic simulation environment for virtual AI 'bots. Then this can be re-used by AI students in subsequent years. Alternatively, each year they could team up engineering students working on the physical robots with computing studetns working on the AI systems, that way both departments are working on their core speciality.
I took Minsky's class last year, and let me tell you, the article couldn't print 75% of the irate stuff he has to say about AI, MIT, and life in general. We once spent an hour class session listening to Misky rant about modern science fiction and random things he didn't like about his Powerbook. In fact, most of his classes were extended rants about something or other (you zealots will be happy to know that he too, hates the Microsoft).
He comes across as affable but bitter. I found it strange that though he cointually complains about the leadership of the AI lab, he and his protege Winston were in control of it for some ~30 years without making any groundbreaking progress. In fact, Minsky's latest work "The Emotion Engine" is simply a retread of his decades-old "Society of Mind." I suspect that now that Brooks and the new guard are moving in, the old guard is looking for someone blame its lack of results on.
Here's some perspective from an MIT AI lab grad student who's been inspired by both Minsky and Brooks. (Minsky is on my Ph.D. committee.)
"AI has been brain-dead since the 1970s."
I agree, unfortunately. At least, what was traditionally meant by "AI" has been brain-dead. There is very little focus in the field today on human-like intelligence per se. There is a lot of great work being done that has immediate, practcal uses. But whether much of it is helping us toward the original long-term goal is more questionable. Most researchers long ago simply decided that "real AI" was too hard, and started doing work they could get funded. I would say that "AI" has been effectively redefined over the past 20 years.
"The worst fad has been these stupid little robots."
Minsky's attitude towards the direction the MIT AI lab has taken (Rod Brooks's robots) is well-known. And I agree that spending years soldering robots together can certainly take time away from AI research. But personally, I find a lot of great ideas in Rod's work, and I've used these ideas as well as Marvin's in my own work. Most importantly, unlike most of the rest of the AI world, Rod *is*, in the long run, shooting toward human-level AI.
Curiously, just last month I gave a talk at MIT, tited "Putting Minsky and Brooks Together". (Rod attended, but unfortunately Marvin couldn't make it.) The talk slides are at
http://www.swiss.ai.mit.edu/~bob/dangerous.pdf.
In particular, I shoot down some common misperceptions about Minsky, including that he is focused solely on logical, symbolic AI. Anyone who has read "The Society of Mind" will realize what great strides Minsky-style AI has made since the early days. I also show what seem like some surprising connections to Brooks's work.
- Bob Hearn
Think of it. Only a stupid robot is a good robot.
People who do perform illusions and escape tricks have been doing things mostly the same way for decades. Magic tricks may change slightly, but all the basic principles and tricks are the same. There's no real evolution, just adaptation to please the crowd.
Now, with that in mind, let's look at artificial intelligence. AI has always been about trying to convince an audience that a machine is thinking. This is demonstrated by the very existence of the Turring test and many products (such as the Aibo, Furby, etc) that try to mimick emotions. If the audence is entertained, amused, or convinced, the AI is considered good. Bad AI is when the audience can see right through it.
Artificial intelligence is magic. It's a trick. It's an illusion.
It is no surprise then that AI hasn't really advanced. The trades of showmen are practically unchanged for hundreds of years. Razzle-dazzling an audience involves technological advances, but it remains unchanged. Even in the cases where "artificial intelligence" is used to aid in medical diagnosis ("expert systems") or manufacturing are really only following man-made logical structures. The computers aren't thinking, they're only doing what they're told to do, even if indirectly. The end result is impressed people who think the machine is smart.
Of course, you don't have to take my word for it. If you want to see how badly AI is going nowhere, I hightly recommend reading The Cult of Information by Theodore Roszak. While his focus is not on the fallicy of AI, it covers it in context with the much broader disillusionment of computers by society.
Now, what does AI need in order to progress? Probably AI creating other AI. Something with a deeper embodiment of evolution. As long as it's man-made, it will never be intelligent, just following a routine. Of course, I am going to stop right here... I am not qualified to offer a solution these obstacles.
Join Tor today!
Human Level AI's Killer Application - Interactive Computer Games, John E Laird and Michael van Lent American Association for Artificial Intelligence AI Magazine Summer 2001 pp 15-25
My summary of the above - the AI in games might not be too hot (some would dispute with the academics about that but let it go), but game environments themselves are complex enough to pose a challenge for state-of-the-art AI researchers.
(Bits On Our Mind, an exhibition of some undergrad and graduate computer science work) ...and I headed STRAIGHT for the nematode booth. You see, I had heard that some clever Cornellian had created a simulation of the entire neural network of a nematode. The way I saw it, there was nothing else there that could possibly be more interesting than that.
:P
So I found myself standing in front of a computer screen. It was a worm swimming through water! In 3D! In real time! After I pushed my jaw shut, I began to ask the genius student some questions...
"Is that real-time?" "Well, actually, no, that is a 10 second looping clip that took a week to calculate."
"Well, I see a neural map there. Is that complete?" "Well, actually, no, that is a simplified version of the real nematode nervous system, on the order of about 1 simulated neuron to 10 actual neurons."
"So you simulate neurons! That's awesome. Let's see the code." (He proceeds to flip through 4-5 pages of very sophisticated-looking mathematical equations to describe the behavior of ONE neuron.)
What a let-down! No wonder Minsky is pissed, real AI is HARD!
We will soon have hardware that has the number of connections or processing power of a human brain. The problem is nobody's come up with the software to run on it. In humans this is what makes the brain more than big organ ... the "soul" if you are religiously inclined. Maybe a human soul can be reduced to nothing more than a program with an enourmous propesity to learn and adapt over years of training / habituation ... say from the years 0 to 18.
There are many authors that have written and demonstrated that the brain probably doesn't function as a mass of context-free predicate logic rules -- including my favorite, Hubert Dreyfus.
Dreyfus argument is old, and its rebuttals are well-known. Consider that symbolic systems are not limited to context-free predicate logic.
The progress of AI is uncertain, but it is certain that there's no future for symbolic logic AI.
It is not certain for me.
Both connectionist and symbolic approaches may succeed if given enough time. However, I think that obsession with neural nets of many people here is of the same nature that obsession of numerous early aviation enthusiasts with wind-flipping devices. Certainly you can mimic mechanics of nature with some effort, but there are usually better ways to do the job.
Lisp is the Tengwar of programming languages.
Actually, they have made significant strides already in figuring out how the brain works. Check out the Levy Lab at the University of Virginia. They've trained a computer model of a rat's hippocampus to do all sorts of intelligent things, such as transitive inference, sequence completion/combination/disambiguatuion, goal finding, etc. While these are not difficult problems for humans to solve or hardcode into a program, the fact that a single network can do these different and sometimes contradictory things represents something that I would call intelligence. As far as I know, they don't plan on having a model of a human brain very soon, since U.Va. lacks NSA-scale compute servers, but even rat-level learning is pretty cool.
WARNING: there is a trojan on your
Try and name an AI researcher who is not a self-important jerk...
... all at MIT/AI ... need I continue?
Oh, say, Rod Brooks, Tomas Lozano-Perez, Hal Abelson, Gerry Sussman, Eric Grimson, Pat Winston, Tom Knight
The difference between Minsky and the rest is precisely as the first poster asserted. Having read Minsky's books, known him professionally and personally, and having taken his course, I must agree that the amount of weight placed on his words are not equal to their value. As others have observed (I forget whom and where), Minksy's original contributions were interesting ramblings at the edge of a new field which happened to pinpoint rich veins of research in some cases, and kill off valuable paths in others (think perceptrons which are, yes, in fact, very useful things, and yes, in fact, do model real neurons reasonably well, and no are not computationally impoverished unless you abide by Minsky and Papert's artifice of only single layers). In otherwords, in some cases, he got lucky, in others he fell flat. This initial success led him to continue pontification (think "Society of Mind", a book of little real contribution), while doing marginally small amounts of actual research. Rod Brooks, in contrast, has made far more, and far deeper, contributions working on his subsumption architecture.
Minsky's course (at the advanced graduate level) consists of students listening to his musings and ramblings which he often repeats through the term, since he has no syllabus, no agenda, and no apparent desire to teach. When he gives talks, they are all extemporaneous; someone like Churchill could pull that off, Minksy's stream-of-consciousness style keeps his acolytes happy, but leaves those with real thirst for knowledge quite parched. Does this not fit the accusation?
So what if Minksy thinks graduate students shouldn't be soldering robots? Does that matter? So what if the current AI field isn't following his pet projects, is he making any contributions himself? We've made tremendous strides in AI over the past decade; they just haven't been where Minksy thinks they should be, despite his questionable over-all track record. Exactly why should anyone care that much?
Put my fist through my alarm clock with its ding-dong death inside my ear. - The Blackjacks.
This reminds me of a quote from French mathematician Henry Poincare: "just as houses are made out of stones, so is science made out of facts; and just as a pile of stones is not a house, a collection of facts is not necessarily science."
Applied to the cyc project: a collection of facts is not necessarily intelligence.
If you were put inside a little white box where you had to flip millions of switches on and off according to certain simple rules, you would look like an idiot next to a computer. A computer can't walk around and recognize things, and doesn't know what an apple is, so what? In my opinion, machine intelligence should be focused on making computers able to make themselves better at what they do best. I'm not sure what a super intelligent computer system would be used for, and I don't think that I would even be able to imagine what would be possible. I would be interested to know what other people think about this idea. Most of the things that I can think of tie back into the "real" world somehow. What would a self-organizing non 3-dimension oriented intelligence be able to do?
Saying that AI is impossible because computers can't come into "our world" of three dimensions, or understand our literature is kind of intelligence chauvinism.
Just guessing here, but I believe the grandparent post was refering more to the work of people like Douglas Hofstader than work being done on artifical neural networks. I can highly recommend Fluid Concepts and Creative Analogies: Computer Models of the Fundamental Mechanisms of Thought as an introduction to this branch of AI research.
I get the impression though that many AI researchers see Hofstader as a heretic. That's too bad because I think the ideas he and his team have developed hold more promise than any other approach to AI currently extant.