Slashdot Mirror
DARPA's IBM-Led Neural Network Project Seeks To Imitate Brain
An anonymous reader writes "According to an article in the BBC, IBM will lead an ambitious DARPA-funded project in 'cognitive computing.' According to Dharmendra Modha, the lead scientist on the project, '[t]he key idea of cognitive computing is to engineer mind-like intelligent machines by reverse engineering the structure, dynamics, function and behaviour of the brain.' The article continues, 'IBM will join five US universities in an ambitious effort to integrate what is known from real biological systems with the results of supercomputer simulations of neurons. The team will then aim to produce for the first time an electronic system that behaves as the simulations do. The longer-term goal is to create a system with the level of complexity of a cat's brain.'"
Upon becoming self-aware, the machine concludes, that its best shot at survival is to keep the host country prosperous and successful...
Any science-fiction authors exploring that turn of events?
In Soviet Washington the swamp drains you.
I'm applying for to a UC with a major in Cognitive Science specialized in computation. This is exactly the kind of thing that I want to be a part of. Even though I don't believe this project will get where it wants to go I do believe it will make steps in the right direction to modeling neurons.
Eat sleep die
This is intuited by the stupid humans in their cliche "Dogs have masters, Cats have staff". We work for the cats.
So, trying to model a cat's brain is both too complex for computers (try and herd cats) and too simple (try and herd pointy haired bosses). The contradiction results in the computer overheating and exploding.
and when the researcher gets home, blubbering about the 'sploded computer to his wife, the dog says "LOVE ME LOVE ME LOVE!!!! TAKE ME ON WALKIES!!!" and the cat says "Get my fucking dinner, you stupid ass. Maybe I will deign to let you pet me. After I do my rounds. Maybe."
RS
Shoes for Industry. Shoes for the Dead.
at least a dog brain. Cats are useless and dumb!
The longer-term goal is to create a system with the level of complexity of a cat's brain.
Seriously? They are shooting WAY higher than simply Artificial Intelligence that mimics humans. Have they ever interacted with a cat before? Don't they know how inscrutable, annoying, and unpredictable they are? Will this computer need a Litter box and Catnip?
This sounds identical to the Blue Brain project. This article is a great intro to the project and I hope some competition will help the race wrap up sooner!
"War makes me sad." - Me
Can a universal turing machine limitedly investigate another universal turing machine and detect halts and infinite loops? I can.
We can look at gunk like
10 Print "Hello"
20 goto 10
Yeah, that's a loop. But we can also look at graphs of y = sin(x) and understand why it repeats. I can also detect patterns and iterations that most likely go for infinity, else find a hole where the assumption falls apart. Last I checked, the computer cannot do that. Not yet, at least.
Hey, it's better than trying to imitate Pinky.
This guy's the limit!
the man in the box?
Sorry, had to go for the obligatory Terminator reference. Seriously, the organic brain is evolved, not designed. That means by definition it must be self contained . Self contained means it has to have a ton of backup, self-repair, and maintance systems. Simulatneously, being organic it competes against other organics, so does not have the same accuracy requirements. Close enough is good enough. As such, I don't see how duplicating an organic brain is useful. We don't need what it does, but do need what it does not have. OK, the ability to approximate is very usefull, but I think a direct attempt at that would work better than the indirect.
excitingthingstodo.blogspot.com
Summary of Test 49:
The robot sensors were properly tracking the missile when suddenly it decided it was time to run bats***-crazy all over the room before perching ontop of a cabinet, turning upside down, and apparently following non-existent bugs across the wall with it's cameras.
Test 49 Results:
System performed as expected.
Conclusion:
Test system has now performed perfectly in the last 48 tests, including the four times where it attacked the researchers without warning, and one where it inexplicably ejected dirty oil on the seat of the head researcher."
This unit can now be considered field ready, though there may be some difficulty tracking it if you take into account the system's autonomous nature and desire to remove it's identification badge.
Yeah, Asimov did about 60 years ago.
Man, when DARPA combines this with the Metal Gear Rex they're building we're screwed.
They should try to imitate Pinky first. Would be easier.
NARF!
Give me Classic Slashdot or give me death!
darpa is gay.
ONE M0DPOINT W@STED!
When the masses get ahold of this and try getting it to scan the internet for pr0wn, and it responds "Not tonight, I have a headache..."...
Who is general failure, and why is he reading my hard drive?
Clearly the first target should be lobsters.
I see some big issues with this.
You can mimic biology and may end up with a semi-intelligent result. Mimic it well enough, and you may have a fully-intelligent result. But because you don't UNDERSTAND what you built, you can't CHANGE it.
Remember the rules of AI, introduced in Sci-Fi? How would you implement rules like that? You CAN'T implement them if you don't know HOW to implement them. If you don't UNDERSTAND the system that you have built, you can't know how to tweak it!
Furthermore, how would you prevent things like boredom, impatience, selfishness, solipsism, and the many other cognitive ills that would be unsuited to a mechanical servant?
The biggest problem is if people productize the AI before it is understood and suitably 'tweaked'. Then our digital maid might subvert the family, kill the dog, and run away with the neighbor's butler robot, because in its mind, that is a perfectly reasonable thing to do!
Simulations are great. Hardware implementations of those experiments are great. Hopefully, in the process, they will learn to understand how the things that they built WORK. But I pray that those doing this work, or looking at it, don't start salivating about ways to make a buck off of it before it is ready to be leveraged. The consequences could be far more dire than just a miscreant maid.
"But does it run Linux?"
"Imagine a beowulf cluster of these" cats.
Would love to see a response from Linus on the subject of a herd of such "cats".
Of course you might keep the entire thing occupied if you just asked it to work on string theorem.
Have they ever interacted with a cat before? Don't they know how inscrutable, annoying, and unpredictable they are?
Saw a great cartoon on that.
- Cat sitting on shelf, staring into space.
- Couple wondering aloud what deep thought are running through its head.
- Thought balloon over cat's head containing a TV test pattern.
EEEEEEeeeeeeeeeeeeeee.......
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
This sounds really great but unless they have a comprehensive theory of animal intelligence to work with, this is one more AI project that will likely fail. Sorry. No amount of computing power is going to help. If you had a good theory of intelligence, you would be able to prove its correctness and scalability on a regular desktop computer.
In my opinion, a truly intelligent mini-brain with no more than a few tens of thousands of neurons would surprise us with its clever abilities. Just hook it up to a small multi-legged robot with a set of sensors and let it learn through trial and error. If you could build a synthetic brain that can learn to be as versatile as a honeybee, you would have traveled close to 99% of the road toward the goal of making one with human-level intelligence.
The problem is, even if IBM runs a project only intend to imitate a cat's brain, it doesn't mean the said imitation wont evolve into something else altogether. This is the problem with neural net. Unless it is mathematically predetermined to be bounded by certain parameters, its cyclic digraph brain will self-insert new nodes and establish synapse links to grow beyond the designated limitation. Further more, it will adapt to its "body" (in this case a super computer). How it works is you have a cyclic digraph which has both category and weight on the edges as the construct, then you pretty much run a bunch of IDDFS threads on it, the threads will by organized and kept track of using a limited size heap. The size of the heap depends on the number of hardware threads your processor(s) can handle at once. There will also be some threads that simply run through the entire structure and re-organize data(retire old connections, and nodes if it doesn't have any connection to it, aka "forgetting"). When a path is used often enough, a new and shorter connection will be established between the source and destination; when new data are being presented, it will be stored in a node and connected to its neighbors. As we can see, the origination of a thought and destination are not so important, it's the path that really are where "understanding" come from. Keep that "understanding" under control is a very hard mathematical problem.
Where is the "Ignorant" mod tag?
Is it just me, or is the idea of modeling any sentient or semi-sentient brain in a computer a little ethically questionable?
To draw a parallel, I just wonder if we'd consider locking a cat in a dark room so small that it can't move, see or hear would be considered ethical. Then what if we removed its body entirely - is that somehow less cruel?
I consider AI research to be critical, so I don't know what the solution is, but this situation is worthy of the question...
If you can read this... 01110101 01110010 00100000 01100001 00100000 01100111 01100101 01100101 01101011
Am I the only one that read DARPA's IBM-Led Neural Network Project Seeks Inmate Brain at first?
I bet 5-Godels ... er $5.00 they fail .....
...just so some scientists can get in on the captioning action.
"I can has cray 'puter?" -- I know someone can do better. What you got?
I had a conversation with my logic teacher last week about AI ( He introduced the class with a brief talk about AI, in which he exposed that some authors considered that human like intelligence was unreachable ).
At the end of the class remembered that I casually printed something AI related while testing mi printer. Read it here on slashdot or somewhere, found it interesting and saved it in a .txt file.
I asked him if this was considered AI:
[quote]The most compelling case for AI I've seen was a complete accident. This story is from memory -- I read about it years ago.
To test whether a neural network could create an efficient program, researchers prepared a network (on a simulator) and then pitted it against a team of human programmers. The task was to write the most efficient program possible for an EEPROM that performed some simple task.
They tested both programs, and they both worked. They compared the code, and the code created by the neural net was a fraction of the size. But the code didn't make any sense and should not have functioned at all.
Also, they discovered that when they took the EEPROM to another location to show someone, the program didn't work anymore.
Eventually the figured out what happened. The neural network learned that the EEPROM could be used as an analog device as opposed to a digital one. It was using complex, unintended functions of the circuitry, like magnetic flux between the wires, to achieve the goal. But because these features weren't engineered, the relationships changed when they took the circuit to another location where the room temperature, barometric pressure, and other conditions were different.
But it shows that we can make machines that are "smarter" than we are -- machines that can end up achieving far more than they were intended to. For that reason, I don't think man will "invent" true AI. I think it will suddenly explode from some random and unintended relationship. It will grow like our own (biological) form of life. And then we just better hope it means us well.[/quote]
I think somebody else replied with this link (or I found it later while googling about it)
: http://www.newscientist.com/article.ns?id=dn2732 , which is quite similar but not the same.
Back to the conversation with my teacher, he said that it wasn't considered AI.
Then I asked him if it could be possible to model a human brain, the only problem that I found in such a feat, apart from the huge computational power required, was the complete understanding of the chemical reactions in a cell, and between celss, if you can model a chemical reaction, you can model a cell, if you can model a cell you can model tissues, cell signaling/neurotransmission,and so on...
He replied that at some point you'll hit the heisenberg uncertainty principle which I vaguely remembered.
The conclusion was that who knows if what makes us intelligent is unmeasurable or not, the only way would be trying to buld such a system.
By the way:
What do you know about neuroscience, any books articles/papers interesting to read about it ?
I have a brief understanding on how de neurotransmitters / neuroreceptors works.I know that it's still a huge area of reasearch, that they're so complex that instead of working with a certain neurorecpetor they're tackled in groups depending on which neurotransmitters affects them, for example serotonin and the 5-HT group.
It's a good thing I watched all those Terminator movies and TV show ... I'm prepared.
You don't even have to RTFA, it's right there in the summary.
I'm just saying.
-t.
The article is based on the IBM's press release and is misleading because of it. In fact, there are three competing teams - one lead by IBM, one lead by HP and one lead by HRL Laboratories. See also the FBO website for more information about this program.
The High Priests have already had there chance to do this and failed, repeatedly. Now they are just throwing more money at them.
This should be an open grand challenge with clear rules like the autonomous vehicle challenge was.
http://www.darpa.mil/grandchallenge/
Even I was surprised at how well they managed to get these cars to drive themselves.
I am sure the same would happen with other AI problems if a large enough prize was put out there.
I am always doing that which I can not do, in order that I may learn how to do it. - Pablo Picasso
"When people post in threads like this, there is an almost universal application of (1) reductionist ideas and (2) the computational (i.e. turing machine - like) paradigm, as if we already know this is how biological systems actually work. In future, I fully expect mimicking biological intelligence, even at the level of a cat, will require a different kind of machine; one that takes advantage of currently unknown physical principles."
Quantum effects are unknown?
"I studied A.I. for 5 years at University and the one lesson I took away from it was that in the theory, practice and philosophy of A.I. there was a significant missing ingredient. I'm sure that we will find it, eventually, but researchers need to make a major conceptual/theoretical leap before we can even begin to try."
God beat'em to it.
Seems like they're being too ambitious. Seems like you'd want to start with something simpler, like Caenorhabditis Elegans, which has 302 neurons. Once you have an accurate simulation of that, THEN move on to something more complicated.
http://www.setiai.com/archives/000050.html
You: "Hi, is this the customer help desk?"
Help Desk: "Meow"
You: "My disk is stuck in the CD drive."
Help Desk: "Meow"
You: "What?"
Help Desk: "Meow Meow"
You: "So, can you help get my disk out?"
Help Desk: "Meow"
You: "The line must be bad. I just hear cat sounds."
Help Desk: "Meow"
You: "(Sigh) I'm going to call back later."
Help Desk: "Meow"
Table-ized A.I.
Isaac Asimov's "I, Robot" covers this (in a manner of speaking) in the final chapter. More precisely, the self-aware robots that control the world's economy do everything they can to simultaneously preserve their positions as advisers to the human race while dispensing the best advice possible for the continued peace and prosperity of humanity.
Do note, however, that in the continued Asimov universe, mankind really didn't explode out into space until he disposed of the "robotic overlords". Those few cultures ['Spacers'] who held on to their robots slowly stagnated and died off.
Asimov's self-aware robots were never the violent, conquering overlords seen in many other sources of fiction (Terminator, Matrix), nor were they really human-equals (Star Wars, Star Trek), but were rather a crutch for mankind that man needed to discard to truly progress.
Also, please note that I am willfully ignoring anything in the Foundation Universe not written by Asimov, as well as Asimov's last book "Foundation and Earth", for reasons that anyone who has read it will clearly understand.
>> Standing on head makes smile of frown, but rest of face also upside down.
"I am sure the same would happen with other AI problems if a large enough prize was put out there."
Give me a million dollars and I can solve the problem of why geeks don't get dates.
Shai Schticks:"You don't make peace with friends, you make peace with enemies"
void main() {
int x = 0;
while (isTheGodelNumberOfAValidDerivationOfTheRiemannZetaHypothesis(x) == 0)
x++;
return x;
}
DRM: Terminator crops for your mind!
Deep Cat (Yeah, yeah, you pervs thought it would be another feline synonym.)
10 a=0
20 b=17
30 a=(a+27389) mod 527
40 b=(b+98372) mod 3991
50 if a!=b goto 30
Will it halt?
I'm sorry Dave I'm afraid I can't MEOW that.
They could wind up with this cat.
Have gnu, will travel.
... and if it could determine the looping pattern and break out, could it then go on to pen some dope lyrics like this:
"Insane in the membrane, ;)
Crazy insane got no brain,
Insane in the brain"
Subject says it all.
" ... The longer-term goal is to create a system with the level of complexity of a cat's brain.' ..."
No, it's not.
Way simpler than a cat.
If we choose something small enough, we can make these computers sound REALLY powerful.
With that in mind, I nominate the unit of "1 Bush" as the standard brain unit.
The longer-term goal is to create a system with the level of complexity of a cat's brain.'
Fuckin' pussys
"If your parents never had children, chances are you wonât either." -Dick Cavett
Not quite. In other words we're not simply a random number generator.
Did Rip van Winkle wake up from the neural network craze 20 years ago? We have next to zero clue about how memory and learning are done at the neural level and now someone arrogant is going to solve the problem? HA HA HA HA HA HA HA HA HA HA!
That's one of the theories about how a brain figures out the senses. Or, to describe it differently, a super-sense or a huge jumble of 'static' that the brain parses and learns at the very beginning of its existence. Are we anywhere near that on the pattern recognition front? This is also dynamic...a brain can pick up a new 'sense' well after birth and incorporate it very well.
What about a network that can take a billion parallel processes and tie that into a network of a billion serial processes in an organized and dynamic way? True self-reflection on each system, oh that's dynamic too.
If the above is done, we have a few small parts of one of the simplest brains emulated.
One of the problems with trying to simulate a brain is that we're using a 'high-level' process to try and simulate something that is much lower level.
A brain's parts are not inherently logical, just like an electron is not. It builds up to make logical components that work in a particular situation. We're starting with pure logic to try to simulate natural world phenomenon, which, if you've studied evolution, is not based on logic but, 'this worked' - the rule set brains formed under is defined by the laws of matter and time, which we still don't know a lot about. Scientists are constantly surprised how biology is utilizing pathways that make no sense if you apply our idea of logic. It can, however, make sense after the fact - because we're able to believe things that we don't think should be (fairly illogical). How can we program, using logic, illogical and poorly understood systems?
You can go as low-level as possible and painfully simulate the particles that make up matter and energy and get accurate enough results. You can also go very high level and painfully simulate the cosmos, and get accurate enough results. We are woefully behind on simulating a brain that is many factors of less complexity than that of a cat and get brain-like behavior. An exact programmatic replica of how brain cells work will help us learn a lot about how cells work but, I think the big picture will be lost in there. We will get the cascades and responses, but I don't think we'll get something that resembles behavior, not for a long time. Mainly because we don't have a big-picture, we have no idea how all this works on a large scale. Someone else mentioned the lack of an over-all theory...we need that well before we start trying to make a software brain.
Sorry, I ramble. I'm very passionate about this stuff, so much so that I dumped a 10-year career and started studying it. ...when I mention logic, I'm mainly talking about computer logic (AND, OR, etc.)
Another cat brain? Anybody remember Hugo de Garis, the CAM-Brain machine and Robokoneko? Is there a pervasive pussy fixation among AI researchers? He didn't get any, I hope these guys do better.
Aristotle had a great theory on gravitation. He even *invented* the word "gravitation". His theory stood undisputed for two thousand years. It was considered absolute truth. There was only one problem: it was a WRONG theory.
It was only after Galileo invented a method to measure the speed and acceleration of falling bodies that the foundations were laid for Newton's theory of gravitation. And it was Michelson's experiments showing small discrepancies in measuring the speed of light that allowed Einstein to develop his corrections to Newton's theory.
This was done nearly twenty years ago with a simulated cockroach brain.
The big problem with human-level intelligence is that it appears to need a human-brain sized neural network. Consider how successful intelligence is from an evolutionary point of view. Humans have totally dominated the biosphere, no other animal within the same range of body size is as numerous as humans, except for those animals we raise for food.
If it were possible to evolve human-level intelligence with smaller brains, it would probably have happened by now. Looking from the biological evolution side, we see in the fossil record a steady increase in brain size in our ancestors. I seriously doubt human-level intelligence is possible with less than about a hundred billion neurons.
http://www.scholarpedia.org/article/GENESIS
Open source hosted at sf, but seems dead since 2006.
The halting problem assumes that computers have infinite storage. They don't, so the halting problem is solvable.
In reality a computer with 2Gb of storage can only be configured in 256^2^31 different ways. Heck the halting problem can be event be optimized to an ordered search...
Seriously this news is about 40 years old. A neural network is by definition an imitation of a brain (or other parts of a neural system of a living organism). Reading the post I have a feeling I had seen at least 5 identical posts before this, it's a theme repeating every couple of years. Not even the players: IBM and DARPA, are new -- sure as hell DARPA has sponsored a number of neural network projects before, only in the 60s it was something new and sci-fi-like.
I suppose this is because neural networks is an obligatory course at uni, so for every fresh student (future IBM employee) there comes a time that they realise, "OMG, we can simulate a brain!" and 1% of them will apply for a grant at their local national institution later.
The longer-term goal is to create a system with the level of complexity of a cat's brain.
We can already model a cat's behavior in software.
"I have never let my schooling interfere with my education." - Mark Twain
Nerves transmit energy at high speed rates... That's most of the trick
ghostbar page.
A lot of publications have picked up this IBM press release, resulting in what must be some of the worst science reporting of the year. Modha and his colleagues at IBM have not simulated a mouse or rat brain. No one can do that at present; the wiring diagram isn't known at that level of detail.
What they did was simulate a huge, randomly-wired network of grossly simplified "neurons" on a supercomputer. The number of units was roughly comparable to the number of neurons in rat cortex, and the statistics of short vs. long-range connections (intra vs. inter-cluster connections) was vaguely suggestive of the organization of cortex, But they used single-compartment, integrate-and-fire neurons that are vastly simpler than real neurons, which do lots of nonlinear processing in their complex dendritic trees. So their network didn't actually compute anything at all. What it did, basically, was oscillate.
http://portal.acm.org/citation.cfm?id=1362627
Calling this a simulation of a severely brain-damaged baby rat being run through a blender while having an epileptic seizure would still imply far too much realism to this cartoon.
The Modha group's work is a useful step toward the long-term goal of eventually building large-scale simulations of cortical circuitry on a supercomputer. But to report that they've presently produced a simulation of "a mouse brain", as some of the news articles are saying, is ludicrous.
The amount of neuroscience that needs to be done, the number of people required, and the time and costs it will take to produce an accurate simulation of even a mouse brain are orders of magnitude larger than this modest $4.9 million research contract. I'm amazed that technology reporters can be so gullible.
What about the CBM, CAM Brain Machine? Cellular Automata based evolved neural networks... Pretty old shit.
-=/\- Jizzbug -/\=-
Too true.
Ceci n'est pas une signature.