Artificial intelligence reborn at MIT

MIT has launched a new $5 million, 5-year project to build intelligent machines. To do it, the scientists are revisiting the fifty year history of the Artificial Intelligence field, including the shortfalls that led to the stigmas surrounding it, to find the threads that are still worth exploring. The star-studded roster of researchers includes AI pioneer Marvin Minsky, synthetic neurobiologist Ed Boyden, Neil "Things That Think" Gershenfeld, and David Dalrymple, who started grad school at MIT when he was just 14-years-old. Minsky is even proposing a new Turing test for machine intelligence: can the computer read, understand, and explain a children's book. More details after the jump.

From MIT News:

Gershenfeld says he and his fellow MMP members “want to go back and fix what’s broken in the foundations of information technology.” He says that there are three specific areas – having to do with the mind, memory, and the body – where AI research has become stuck, and each of these will be addressed in specific ways by the new project...

One of the projects being developed by the group is a form of assistive technology they call a brain co-processor. This system, also referred to as a cognitive assistive system, would initially be aimed at people suffering from cognitive disorders such as Alzheimer’s disease. The concept is that it would monitor people’s activities and brain functions, determine when they needed help, and provide exactly the right bit of helpful information – for example, the name of a person who just entered the room, and information about when the patient last saw that person – at just the right time.

The same kind of system, members of the group suggest, could also find applications for people without any disability, as a form of brain augmentation – a way to enhance their own abilities, for example by making everything from personal databases of information to all the resources of the internet instantly available just when it’s needed. The idea is to make the device as non-invasive and unobtrusive as possible – perhaps something people would simply slip on like a pair of headphones.

"Rethinking artificial intelligence"

• [pkalina]

  In #28, SamSam writes, “Old people aren’t known for their abilities to embrace new ideas…”

  What an offensive stereotype! Each of us is an individual — I have known close-minded people in their twenties and wise and wonderful eighty-year-olds.

  In addition to being offensive, your assertion lacks supporting data. See, for example, Charness, Gary, and Marie-Claire Villeval. 2009. “Cooperation and Competition in Intergenerational Experiments in the Field and the Laboratory.” American Economic Review, 99(3): 956–78

  To the editors: Your policy on comments precludes racists, sexist, and homophobic remarks. Good! Please know that ageism is just as offensive.

  • [SamSam]

    Actually, while I was being joking flippant by calling them “old farts” (I actually received my MSc in AI and respect these people), my comments about old people weren’t entirely without backing.

    There is quite a lot of anecdotal evidence of scientists doing their most seminal work at a young age, although of course one can come up with numerous counter-examples.

    Here are the results of a very quick article search:

    At What age do scientists do their best work? (pdf) (pub. in FASEB Journal, 2008):

    Shows scientists do their “best” work (by articles with most citations) between the ages of 33 and 35, and declines there-after, in almost all life sciences.

    An economic model of the life-cycle research productivity of scientists, (pub. in Scientometrics, 2005):

    Shows “number of citations made to a scientist’s previous work will decline with age.”

    Another good page here on the correlation between youth and scientific accomplishments, which also makes the good point that our perception of scientists are always as older people (think of the iconic images of Einstein and Darwin), even though they most commonly do their seminal work in their youth (both Einstein and Darwin made their greatest contributions to their fields in their 20s).

    So, though I understand your asking the administrators to censor me because you find me offensive, please recognize that not everyone talks out of their asses when they write things.

    Anyway… it sounds like the team will be quite a mix! Sounds great.

    • [pkalina]

      In #40, SamSam writes, “There is quite a lot of anecdotal evidence of scientists doing their most seminal work at a young age, although of course one can come up with numerous counter-examples.”

      That is a much more reasoned statement than your comment in #28, and one with which I agree.

      Ageism is rampant so I am very sensitive to blanket assertions about the capabilities of older workers. Look it up — folks in their fifties do fine in a bad economy as long as they hold on to their jobs. But if they happen to lose a job through layoff or because their employer goes out of business, it takes older workers much longer to find new work. And that new work tends to pay far less.

      In fact, Charness and Villeval found, “First, seniors are not more risk-averse, as opposed to the conventional stereotype. Second, both juniors and seniors react to the competitiveness of the environment and there is no significant difference in performance in the real-effort task across the generations when they are competing. Third, seniors are typically more cooperative than juniors in a team-production game. Cooperation is highest in groups in which there is a mix of juniors and seniors, suggesting that there are indeed benefits in maintaining a work force with diversity in age.”

      I agree — I’m eager to see what the new MIT team comes up with.

      Btw, I think many scientists do their most seminal work at a young age because there’s more pressure on them when they’re younger. Also, after they’re tenured, much of their creativity is channeled into their work with their grad students.

• [nanuq]

  “Minsky is even proposing a new Turing test for machine intelligence: can the computer read, understand, and explain a children’s book.”

  Considering some of the children’s books out there, this might not be such a good idea. Designing a computer to understand something by Dr Suess or Lewis Carroll seems unbelievably hard. Just getting a computer to understand the difference between a Snark and a Boojum would take a whole new science of AI.

  • [cymk]

    Lets start with See Spot Run, and work our way up from there.

  • [Brainspore]

    You just made me imagine a scenario where a computer short circuits and burns out “Star Trek” style while trying to understand why the residents of Whoville would continue to sing even after the Grinch stole their presents.

• [Anonymous]

  There are some who say that the problem with the AI community and their inability to make any progress was precisely Marvin Minsky’s hectoring attitude an unassailable genius. If you were a grad student trying to counter any of his beliefs or theories you had to go up against a solid wall of theoretically backed “genius” to make any headway. This led to AI researchers cowering in corners trying to find different models of intelligence merely to escape his vitriol (q.v. insect-based bottom-up AI techniques).

  Rebooting the field by reintroducing Minsky may be sounding the death knell of this project right from the beginning.

• [Flugelmeister]

  I prefer artificial stupidity

• [octopod]

  rebooting ai research is a good idea, it’s been going nowhere for the longest time. the stupid robot dog was just lame. they should get some decent writers in, and trisha helfer.

• [Kimmo]

  BRING IT

• [anansi133]

  If they want to fix what’s broken, they’ll do some elementary philosophy first. Questions like, “What kind of thinking *don’t* we ever want machines to do for us”? And, “When is a human always going to be a better solution than a machine?”

  To my mind, most of what was broken about artificial intelligence, was the search for a new kind of slavery without the moral qualms.

  ‘Weak’ AI theories always seemed to have more practicality than the strong stuff.

  • [Karl Jones]

    To my mind, most of what was broken about artificial intelligence, was the search for a new kind of slavery without the moral qualms.

    I’m reminded of John Brunner’s novel A Maze of Stars, in which an artificially intelligent spacecraft (the Ship) seeds humanity on planets throughout the Arm of Stars, then periodically (over vast periods of time) revisits the planets.

    The Ship’s AI is fundamentally based on the brain of a giant squid (a lonely, solitary creature — the kind of mind that might endure the loneliness of the voyage), with an upper-brain overlay of AI based on an amalgamation of human minds to keep the squid-AI in check.

    Bottom line: the designers wanted an intelligent AI with free will, yet obedient to humanity’s needs.

    And what do we call a creature that possesses free will yet remains an obedient possession of another creature?

    A slave.

    • [arkizzle / Moderator]

      Karl, I’m reminded of Ian M. Banks Culture series.

      Similarly, the ship Minds have free will. But they have no more particular compunction to engage with society than humans do. Their chains are only social, just like the humans in that universe.

• [zapakh]

  Never mind Alzheimer’s, I’ll take one of those social prompters and have it remind me to ask people things, because I always forget by the time I see them.

• [PathogenAntifreeze]

  The history of effective AI will probably begin with this: “On April 30, 2008 a team at HP Labs announced the development of a switching memristor.”

  That news last year was probably one of the biggest, game-changingest, completely-unnoticed-at-the-time developments in human history so far. If you know much about neural networks, how model neurons work, and the “current” limitations, (I took AI courses during Computer Science curriculum circa 1999, but I haven’t seen significant progress since) you will be very excited when you learn what a Memristor is and about them going from theory to reality.

  http://en.wikipedia.org/wiki/Memristor

  • [Mike Estee]

    The history of effective AI will probably begin with this: “On April 30, 2008 a team at HP Labs announced the development of a switching memristor.”

    Well put.

    I would phrase it another way: The problem with modern AI is that it is a problem people have largely been trying to solve with fixed architecture computers and procedural programming languages. We’re going to have to fundamentally change the way computers, and the languages used to program them, work before we crack this nut. Massively parallel programable memristor nets seem like a good place to start research.

    On the other hand, it seems to me that Artificial Intelligence is much more likely to arise as a byproduct of Artificial Life, than it is from Computer Science.

    • [Xenu]

      The problem with modern AI is that it is a problem people have largely been trying to solve with fixed architecture computers and procedural programming languages. We’re going to have to fundamentally change the way computers, and the languages used to program them, work before we crack this nut. Massively parallel programable memristor nets seem like a good place to start research.

      It seems like you’re making an assertion that there is a fundamental difference between two things which are, in fact, logically equivalent.

• [SamSam]

  Bah. I see this getting hailed everywhere as “rebooting” the field and “starting anew” and all that.

  Seems to me that the last people you want “rebooting” a field are the same old farts who have been working on it for the past sixty years. The same old farts who invented it, no less.

  Old people aren’t known for their abilities to embrace new ideas and have new sparks of inspiration at the best of times, and even less so when they each have spent their whole careers defending their dogmas and reputations.

  I do hope they have some people below the age of 30 on their team, or I really don’t see it going anywhere. Done right, though, an all-star team of the old pioneers and young genius undergraduates and such, structured in a way that everyone could actually contribute and not be overawed by fame, you might actually get somewhere.

• [Ito Kagehisa]

  Rebooting the field by reintroducing Minsky may be sounding the death knell of this project right from the beginning.

  I don’t know Minksy personally, but it seems to me that he keeps making confident predictions and proclamations that haven’t panned out. I wouldn’t want him on my project, because I don’t see any evidence that he can deliver anything.

  I keep hearing him hailed as a genius, also, but I never see any actual evidence that this is so.

• [Rusty]

  Ed Boyden was in my class in high school – although he’s not “under 30″ he should be about my age, 32. So they’re not all old farts.

  Marvin Minsky is a smart old dude, though. And regardless of the fact that people seem to think of him as trumpeting the coming golden age of AI his seminal works to me seem to be about what computers/AI can NOT do (the only one of his book I’ve read, admittedly, is somewhat old, about the “perceptron” and what it was not capable of)

• [dculberson]

  Why is the Turing test still seen as a good test of artificial intelligence? It’s a good test of artificial human intelligence, but that’s a scale of intelligence that took billions of years to evolve. That coupled with the grounding in “practical” application screams a misdirected focus to me. Why not focus on making systems that learn on a rudimentary level first? If they can come up with a system that acquires knowledge, then optionally a way to make it curious (as a cat!) then to me that’s a functioning AI.

  A chimpanzee can’t read and understand and explain a children’s book, but there’s no denying that they’re intelligent. Same with a gorilla, or a dog, or a cat.

  It’s like deciding we want to generate electricity, and starting with a fusion reactor first.

  • [friendpuppy]

    We’ve got systems that learn on rudimentary levels. It’s the “scaling up” that is the problem. Ideally we could make a working cat brain and try to evolve that but there are no automated tests that could say, “well, this is the best candidate at this stage, we’ll move it up to the next level” If we could only automate the detection of intelligence…

• [Brainspore]

  Synthetic neurobiology is an interesting field but as an approach to A.I. it seems kind of like building an airplane based on the way a bird flies (which is exactly what people did for centuries with very little success). Sometimes I wonder if Da Vinci would have been able to conquer flight if someone had told him to abandon the flapping wings altogether and focus on a simple glider first.

• [Anonymous]

  People want robots who will clean their homes, get their mail, get the hammer out of the garage, put the laundry from the washer to the drier: the problem with the AI world is that they aren’t solving the right problems.

  We need special task robots, not general AI.

• [pjcamp]

  This looks a lot smaller than it sounds.

• [Clay]

  Or perhaps even:

  We have been tugging AI along on planks of increasingly sophisticated lubrication. The memristor poses a wheel.

• [Lifewish]

  “There is also the elegance of running a neural network *on* a neural network.”

  This is only elegant if your underlying “neurons” happen to look something like the “neurons” you’re modelling. Otherwise you’re probably going to have to have a seriously complicated (and unintelligible) intermediate layer.

  Anyway, any attempt to model biologically realistic neurons is pretty much doomed to failure at the moment, cos we don’t know how the buggers work. The “weighted sum + threshold function” model is easy to implement, but demonstrably does not provide a complete description of the underlying system. And last I heard, the whole back-propagation strategy for NN programming was basically unverified.

  The current approach seems to be to model small sections such as simple visual systems* in great depth, and work outwards from there. This seems like it might actually add value – even if we can’t get it to work, these systems are small enough that it should be possible to figure out why.

  This project, by contrast, seems more like a return to the old quasi-philosophical approach to AI. That produced lots of beautiful computer science, and hopefully this new project will do the same, but in terms of actually modelling/implementing realistic intelligent systems I consider it a backwards step.

  * For an example of the state of the academic art, see: http://www.ploscompbiol.org/article/info%3Adoi%2F10.1371%2Fjournal.pcbi.1000555

• [Anonymous]

  The last thing we humans will invent is AI. Then its lights out for us. Some things shouldn’t be done.

  • [Anonymous]

    You’re absolutely right. Whenever people bring pets into my home the first thing I do is destroy them. Children too. Anything less intelligent than me. I can’t even help it. I had a brother, but he wasn’t too good at math…

• [Anonymous]

  I think the real AI will come when they can solve undecidible problems (i.e. problems that computers cannot solve by definition)
  more information at http://en.wikipedia.org/wiki/Undecidable_problem

• [jtf]

  I’m totally high-fiving David for getting Boing Boing’d.

• [friendpuppy]

  Is there any reason a memristor has not or cannot be emulated with software? How much different is using a memristor than weighting synapses? I’ve always held that strong AI is a software problem, not hardware.

• [davidad]

  Minsky was on my master’s thesis committee, and he was easier to please than either of my other readers. Yes, it’s true, he’s sometimes flippant, maybe a bit too confident at times, but he does have an unusual power to see the obvious things that most do not, as well as a priceless oral tradition of stories about famous academics dating back to B.F. Skinner, and the mistakes they made.

  I’m 18, and we have a 15-year-old on the team, along with more grad students (under 30) than professors, and I can assure you the dynamics are as far from stifling as it gets.

  On the memristor stuff, I’m a programming language theory guy, and my master’s thesis was on computer architecture. Let me make it clear that the Mind Machine Project is a vertically integrated group – we have a chip designer on the team, a couple programming language people, some software engineers and some folks like Minsky who do the top-level conjectures. I absolutely believe that AI will not happen until we have more appropriate lower-level tools (and making those new programming languages is my job). Unfortunately, we haven’t been able to get access to HP’s research technology directly, but we are working with a reconfigurable machine architecture called RALA that could be seamlessly transitioned to a massively parallel memristor net when they become commercially available. It needs some work, still, as do the programming language tools, but I think we’ll be able to finish that theoretical work as the material science of memristors progresses to a reasonable cost basis.

  #29: My personal belief – not affiliated with MMP in any way – is that brain modeling is actually more likely to win the race to a complete, real-time, human-equivalent AI. Nonetheless, we’ll need the computer science to get there, and we might come out with something really interesting, even if it doesn’t act like a human does.

  #30: We are actually looking at contacting some sci-fi writers – Greg Egan, for instance. No Trisha Helfer, though, sorry. :D

  jtf: Thanks! Unfortunately, I can’t tell who you are. >_>

• [Anonymous]

  5M is not a ton of money/ cool they are cool enough to run a sanity check.

• [jfrancis]

  When I was in college I invented SI – Superficial Intelligence. You could converse with it for hours. It would repeat the last 2 or 3 words of your previous sentence and add a question mark.

  And add a question mark?

  Yes, apparently when you– d’oh!

• [Anonymous]

  check out bogosity, measured in micro lenats

• [John Lupien]

  Oh, they have $5 million? Give that AI project
  one white chip.

  • [Trent Hawkins]

    You know, PS3s dropped in price. Could build one hell of a super cluster for 5 million dollars.

• [imag]

  Good call Pathogen…

  friendpuppy: The problem is one of horsepower. Emulating a massively parallel network on a few (even very fast) processors is difficult. A neural network like the one in your brain is composed of billions of very simple “processors” that function essentially like memristors – use stimulates the process that occurs when that neuron is used.

  In other words, with software on traditional hardware, bigger networks take exponentially more time for a serial processor to emulate, while memristors offer simple linear scaling. It should be said that Google datacenters are beginning to approaching the number of operations that take place in a human brain, which means they could theoretically begin emulating networks of that complexity.

  However, Memristors offer the potential to have billions of very small processors, so no software emulation is necessary. Ideally the array could be orders of magnitude more compact and much cheaper. There is also the elegance of running a neural network *on* a neural network. It means that, like the brain, the way you structure the array could determine how it functions, with different cortexes, hemispheres, etc…

  • [arkizzle / Moderator]

    I vote we get Google to pick one day a year, let’s call it AI day.. (but not G[oogle]AI day, that’s something else) and everyone agrees to not need Google services for that day.

    Then they can dedicate their datacenters to testing full human brain emulation. It could be quite a thing. Like a new world holiday, where everyone waits to see what pops out this year, see what sort of character the GOOG is becoming. It’d probably be pretty shit for the first ten years or so, but then it’d be everyone’s second favourite observance of the year (after Rabies Day! obviously).

• [Nash Rambler]

  I, for one, welcome our new brain-augmenting AI overlords.

• [Keneke]

  Minsky’s Turing test is cute, but you could just as easily say that a Turing-positive AI could interpret metaphors or song lyrics. It’s this ephemeralness of quantification that makes reverse engineering the brain so hard in the first place.

• [technogeek]

  Since we’re finally starting to learn a bit about brain architecture, it may not be a bad time to start thinking about testing some of those theories by implementing them.

  I’d be sufficiently impressed if they can build something as intelligent as a cat or dog… including social intelligence.

• [Anonymous]

  There is an interesting article about this at the New Atlantis, way at the bottom of the front page.

  http://www.thenewatlantis.com/publications/why-minds-are-not-like-computers