IBM building Blue Brain (full brain neuron simulator)

http://www.seedmagazine.com/news/2008/03/out_of_the_blue.php?page=all&p=y

by andr • 18 years ago

71 55 18 years ago

This is amazing - a complete simulation of a section of cortical column. And by complete, it's 400 subsimulations per neuron. Outstanding! Thanks IBM!

mixmax• 18 years ago

This is one of the advantages of big companies: they can afford to do expensive stuff even though there is no immediate guaranteed return.

pchristensen3• 18 years ago

Don't thank IBM, thank their clients that pay too much for consulting.

albertcardona• 18 years ago

Take the results with some salt. No matter how fascinating the results, if the model is wrong then there is no advancement in our knowledge of the brain.

As far as I know, they did not go the long and hard way of actually reconstructing a cortical column from sectioned tissue (as in electron microscopical imaging of serial sections of 60 nanometers in thickness).

What they have done, even if remarkable, is "simply" a simulation of a network built from statistical data gathered from many different studies of cortical columns.

michaelneale5• 18 years ago

I watch with great interest. Is this how general AI will happen? or is it more like heavier then air flight was: we will get there, we just need to understand the fundamentals of what makes AI better.

nsrivast2• 18 years ago

I doubt it. Our modeling capabilities continue to develop much faster than our knowledge of how the brain works.

Of course, neural network modeling can still be of great use. Since we've more or less figured out how the brain does a variety of low-level processing tasks (like coordinating muscle movements, finding the visual depth and texture of observed surfaces, etc.), this knowledge can serve as a starting point for developing exciting technologies at increasingly faster speeds (robotics limbs, computer-guided navigation, etc.). Also, advanced neural networks can serve as rudimentary "existence proofs" for theories of mind and consciousness. For example, a theory that claims that our incredible computing power stems from a recurrent network of thalamo-cortical neurons is strengthened by a computer simulation that exhibits similar large-scale behavior in similar time scales under suitable parameters.

That last sentence is pretty qualified, for good reason. We shouldn't expect to write some code and discover a working brain, at least not until we figure out how it works to begin with.

greendestiny3• 18 years ago

I think it's entirely possible we'll develop general AI without understanding consciousness or even some of the other functions of the brain.

hugh• 18 years ago

I agree. I suspect that scaling up a simulation of a neuron and expecting it to act like a brain is a bit like scaling up a simulation of an iron atom and expecting it to act like a car.

xirium3• 18 years ago

Ray Kurzweil argues that we can eventually brute force the solution. However, we can do it more efficiently if we do it at a functional level. Basic acoustic function given as an example ( http://www.kurzweilai.net/articles/art0134.html?printable=1 ).

michaelneale5• 18 years ago

So if we have brutally powerful machines, that can do a human scale brain - how to we "start" it - is it going to have to learn itself, as we do, or is there tech to "read" our minds into it? (creepy sci fi style).

DanielBMarkham• 18 years ago

The day one of these simulators does something unique and interesting that can't be done with other hardware -- I think we will have turned a very important corner.

andr2• 18 years ago

Things like that make me think it's time to make Asimov's Three Laws* a legal requirement.

* http://en.wikipedia.org/wiki/Three_Laws_of_Robotics

jey2• 18 years ago

Asimov's Three Laws are a plot device, not an actual workable system for machine ethics. This is a very very important area for further research.

andr2• 18 years ago

They are a very good foundation for a workable system for machine ethics.

bugmenot2• 18 years ago

Its been condensed to "Do no evil".

moog• 18 years ago

'Markram believes that he'll be able to model a complete human brain on a single machine in ten years or less.'

Does this mean coding is in its last days? In future, will we just chat to machines instead of programming them?

SuperThread5• 18 years ago

Why am I the only person who's terrified rather than excited by every new incremental progression of AI? Do you guys actually look forward to the day when humans are made obsolete?

PS: This scientist seems to be exponentiating incorrectly. In ten years he'll have 32-ish times as much processing power; he needs 10 million times as much to get to the level of the entire brain.

tx3• 18 years ago

Every time mankind makes a major breakthrough it makes entire families of professions obsolete. This gives us opportunities to move onto more advanced forms of keeping ourselves busy.

Most of the time when people ask about "obsolete humans" in context of AI, they really mean "programmers" or "doctors" or whatever. Yes, it might. It simply means that there are professions to be invented that will be to programmers as programmers are to janitors today.

SuperThread5• 18 years ago

But once AI becomes more intelligent than humans, every conceivable role will definitionally be better filled by that AI than by actual humans.

agent86a2• 18 years ago

Apprehension about progress in artificial intelligence is entirely natural. In fact, I think apprehension and denial will lead people to continually redefine their notion of intelligent behavior so that current computers are always excluded.

Not so long ago many people said a computer could not beat a grandmaster at chess without being intelligent. Enter Big Blue. Others have stated computers will never compose music that is emotionally meaningful to humans without being intelligent. Enter Experiments in Musical Intelligence and other widely-acclaimed composition programs.

Until the Turing test is passed people will be able to plausibly deny any advances in artificial intelligence. No matter how advanced such "brain in a box" models becomes, they won't pass the Turing test without being embedded in a rich environment with which they can interact.

anewaccountname2• 18 years ago

Can you link to a single good composition program? I have heard Wolfram's and I thought it was crap, and the only decent one I have heard was one which could only emulate old composers.

newt03112• 18 years ago

Human bodies may be made obsolete but I am keeping my fingers crossed. If we have computers which can simulate human brains (and presumably do it more efficiently/with more capabilities), we should have methods of transferring consciousness back and forth eventually at which point, I can implant myself into the computer and have all the advantages. For greater detail on this, see Space Odyssey 2001.

dusklight• 18 years ago

Anyone watching the new TV show "Terminator: Sarah Connor Chronicles"?

It seems to me, whether or not to be scared about the progress of AI, depends on a very simple question.

Should we treat AI like our slaves, or should we treat AI like our children?

bugmenot2• 18 years ago

Skynet really exists: http://en.wikipedia.org/wiki/Skynet_%28satellites%29

anupamkapoor• 18 years ago

> Should we treat AI like our slaves, or should we treat AI like our children?

more likely scenario would probably be, how _we_ would be treated by _them_.

s3graham• 18 years ago

Awesome (in the archaic sense) that the size estimate is only 200x what Google has currently. Seems trivially within reach in, what, only 8 doublings of performance? They guess at 10 years for "one machine", so what, 15-20 years before it's pda size/priced? An extra full-real-thinking-feeling-creating-art-type-brain in my laptop?! Holy shit.

I was actually just wondering why this project wasn't happening yet while reading "The Singularity is Near". If you haven't read it, I recommend it. It's hella dry and boring in parts, but overall worth reading if only to make you think about what a startup might look like in only 5 or 10 years.

xirium3• 18 years ago

The central argument from the book The Singularity Is Near by Ray Kurzweil is on his website ( http://www.kurzweilai.net/articles/art0134.html?printable=1 ). It was almost completely ignored when it was posted to this forum a few days ago ( http://news.ycombinator.com/item?id=126183 ).

marvin2• 18 years ago

I think the reason most people are ignoring Kurzweil is that he sounds like a crackpot. I happen to be a guy who owns an autographed copy of Kurzweil's latest book, but I acknowledge that the idea of the technological singularity sounds too good to be true. There have been thousands of cults proclaiming that the end is near; it would be a coincidence of historical proportions if this one happened to be right.

Kurzweil presents many convincing arguments, but the central tenet is that the doubling of processing power keeps up for another two decades. Due to the nature of nature, what we observe as exponential growth in the end invariably turns out to be logistic growth. If this is so, then at any time the doubling of processing power could slow, and this would mean that we would only, at the most, get one more doubling. Such an event would bring the dream of a near-term universe full of computers to a screeching halt.

I believe that we should be able to create machines that think, but it is important that we aren't blinded by ideology. A massive "singularity movement" that ends up disappointing people, will in the best case delay the advent of such technology due to funding issues. I think the history of AI has demonstrated this quite clearly.

If Kurzweil's estimates are correct, though, another 8-10 years of "exponential growth" in computer hardware will put full-scale simulation of a human brain into supercomputer territory. And frankly, I doubt that evolution has managed to find the most computationally efficient way to achieve intelligence. There's got to be a way to do this.

truck• 18 years ago

I wonder how they plan to do parameter estimation? Millions of nonlinear dynamic models running, each with at least 3 free parameters in the simplest of neuron models, with an almost infinite number of possible topologies of the networks. That is quite the search space and neuroscience can provide very little prior knowledge.

bayareaguy• 18 years ago

The human brain requires about 25 watts of electricity to operate. Simulating the brain on a supercomputer with existing microchips would generate an annual electrical bill of about $3 billion. If computing speeds continue to develop at their current exponential pace, and energy efficiency improves, Markram believes that he'll be able to model a complete human brain on a single machine in ten years or less.

I wonder how many megawatts Blue Brain will need to do something simple simulating only 1e5 cells.

According to wikipedia,

- the average power consumption of a human cell is around 1e-12 watts.

- 50 years ago ENIAC needed 150 kW to do what you could probably do today in a few µW.

Based on those numbers, I think Markram is way too optimistic about simulating the brain on a signle machine in 10 years. If things continue as they have, it's more likely to take closer to 50 years.

I'll pay more attention to Kurzweil's singularity when the difference in power consumption between a tiny fraction of a simulated brain and a real one drops below ten orders of magnitude.

IBM building Blue Brain (full brain neuron simulator)

Related Stories