8.25.2005

 

Modeling the brain

Swiss lab works with IBM to uncover power of brain: "The scale of the project shows just how powerful our brains are. It will take an IBM eServer Blue Gene supercomputer containing more than 8,000 processors housed in four racks to model the behavior of something less than a billionth its size: a neocortical column is a cylindrical element roughly 0.5 millimeters in diameter and 2 mm long. A rat's brain contains 10,000 of them -- and the human brain around a million, according to Henry Markram, director of the Brain and Mind Institute at EPFL.

Markram's plan is to model the cylinders at the level of individual neurons: A neocortical column contains about 10,000 of those. Over years of research, he and his colleagues have accumulated a huge amount of data about how individual neurons behave and how they interconnect and interact with one another to make up a column. That data will be used to create a detailed model of a single neuron, including its three-dimensional form, inside each processor, with the Blue Gene computer as a whole modeling a column."

This computer's 8,000 processors combine to create 22 trillion FLOPS, or roughly 2.75 billion FLOPS per processor. They are using roughly one processor per neuron.

If the human brain has a million of these 10,000 neuron columns, then the brain has 10 billion neurons. Using the techniques described in this article, then, we would need about 27,500 quadrillion FLOPS to simulate the human brain. That is much higher than previous estimates. For example, in this Wired article the estimate of the human brain's processing power is placed at 100 trillion FLOPS: "A human brain's probable processing power is around 100 teraflops, roughly 100 trillion calculations per second, according to Hans Morvec, principal research scientist at the Robotics Institute of Carnegie Mellon University. This is based on factoring the capability of the brain's 100 billion neurons, each with over 1,000 connections to other neurons, with each connection capable of performing about 200 calculations per second.". The number given in Robotic Nation is one quadrillion.

It will be interesting to see where the number actually is.

Comments:
the processing power to model a computing device and the processing power that the given device puts out are different quantities.
For instance, there are lots of different emulators of older computer architectures around on PC platforms ( like old gaming computers etc ). It takes a significant amount of processing power to "model" them, much more than each individual device was capable of. And this multiplier is increasingly larger, the bigger the architectural differences between the modelling device and the modelled are.
For example, there are some devices that can be considered "analog" computers that are really quite simple, but modelling them in digital space accurately often takes very much computing power.
Now its obvious that difference between "neuron computing" and transistor-based binary computing are enormous, so it is to be expected that modelling the neuron device takes several orders of magnitude more power than the "neuron device" actually puts out.
 
Good point kert, like Super Nintendo's Sony sound processor being emulated on a computer.

The brain as computer analogy just doesn't hold up very well. Apples and Oranges. Although their output is taste, they're two different fruits.
 
Vasser,

Designing hardware and fabricating chips are two very different things.

It is clear that a smart enough AI could probably use all the world's computation to its advantage, but the rate of increase in intelligence seems reliant on the algorithmic improvements, new processor acquisition, and (over a longer term) fabrication of new hardware.

Also, simply simulating a super-intelligence doesn't create a will or motive for that intelligence to expand itself.

So I would give it at least a few months, rather than minutes :-D
 
They're using general purpose CPU's to perform the simulation.

Once simulations are successful, then they'll start to say, "Okay, we know this works now: Now how do we optimize this circuitry?"

Like graphics cards. But they'd be neuron cards.

-- Lion
 
If you divide the time to do some mental process (e.g. recognize a fox) by the time required for one synapse transition, you will get a very low number. This should make it obvious that the article is off-base. email me at jackmullaney@yahoo.com if you are interested in discussing a more reasonable/workable approach
 
Obviously we are more than a few decades away from superior AI... but that doesn't matter. I can see the day is near when somebody makes a robot that can perform a low level service jobs. The technology to create a robot that buses tables and washes dishes must be near.
 
And no, 50 year olds wont be going back to school to learn nanotechnology 101.

Hmm...

I wonder.
 
" Obviously we are more than a few decades away from superior AI... but that doesn't matter."

Not so obvious. I know some very informed people that would be surprised if it didn't happen before 2020.

I'll leave my own point of view aside right now.
 
Well, let's say we got a human-level intelligence in 2050- just to pick a date that is easy to calculate from.

Within minutes, it'll be able to reason... ...about as much as a human would be able to reason.

Right? I mean, because, by definition, we're talking about a human level intelligence.

It will have some advantage, because it'll be "native" to the Internet, and be able to copy it.

But it won't overtake the world within minutes, even with the ability to copy itself. I mean, consider if you could copy yourself.

Would you be a super-power capable of taking over the world, within minutes, even with your copies? I doubt it.

It'll take at least 2 years for Moore's laws doubling to increase your intellectual power.

And we don't know if intellect increases linearly, exponentially, or logrithmicly with increased processing available.

So, I doubt that the first human-level intelligence will take over the world within minutes of appearance.

(And: I also wonder: Wouldn't the first human-level intelligence be saying "goo goo ga ga" if it's a neural simulation?)

-- Lion
 
I spoke to a guy in Manhattan Beach (CA) who has business contacts in cutting edge computer science and engineering. He said they already have a working model of a nanoscale computer that is about the size of one die (singular of the word, dice) and has supercomputer capability. These are the things that are happening now, that will end up trickling into the consumer market based on regular market forces in the coming years.
 
I am not an expert in artificial intelligence, but would advise you to refer to an excellent book from Jeff Hawkins (founder of Palm and Handspring), "Intelligence". Hawkins spent years on studying neurology and has a very interesting approach on how to reproduce human intelligence.

The type of research mentioned above follows numerous initiatives to discover how signals are processed in the human brain. Hawkins pushed towards top bottom approaches that would help reconciliating these efforts and help reproducing inetelligences cose to human once.
 
Thanks for the info, I added you to my favorites.
If you are interested in business home online opportunity work
, I have a business home online opportunity work
site.
Feel free to drop by and tell me what you think.
 
Extremely late correction:

The brain does have about a million cortical columns, making your estimate of neurons roughly accurate... for the cortex. Most of the neurons in the brain are not in the cortex, however, so the estimate will have to be increased several fold. That, and the blue brain project is only just now working toward molecular simulation, which they don't even believe is possible on current computing platforms, which makes even the single column much, much, much more complex.

The computational power of the brain is very, very much higher than most estimates.
 
-----warhammer gold warhammer money warhammer accounts tibia money tibia gold tibia item runescape accounts buy runescape accounts runescape money runescape gold runescape gp runescape power leveling runescape powerleveling cheap rs2 powerleveling runescape equipment buy rs equipment runescape runes cheap rs2 runes runescape logs cheap rs2 logs runescape items buy runescape items runescape quest point rs2 quest point cheap runescape questpoint runescape gold runescape items runescape power leveling runescape money runescape gold buy runescape gold buy runescape money runescape items runescape accounts runescape gp runescape accounts runescape money runescape power leveling runescape powerleveling tibia gold dofus kamas buy dofus kamas wow power leveling wow powerleveling runescape questpoint rs2 questpoint Warcraft PowerLeveling Warcraft Power Leveling World of Warcraft PowerLeveling World of Warcraft Power Leveling Hellgate money Hellgate gold buy runescape logs buy rs2 items cheap runescape items Hellgate London gold Guild Wars Gold buy Guild Wars Gold runescape items rs2 accounts cheap rs2 equipments lotro gold buy lotro gold buy runescape money buy runescape gold buy runescape runes lotro gold buy lotro gold runescape money runescape gold cheap rs2 powerleveling eve isk eve online isk buy runescape power leveling rs2 power leveling tibia gold tibia item runescape accounts Fiesta Silver Fiesta Gold Scions of Fate Gold Hellgate Palladium Hellgate London Palladium SOF Gold Age Of Conan Gold AOC Gold ArchLord gold tibia money tibia gold runescape accounts runescape gold cheap rs2 powerleveling buy ArchLord gold DDO Plat Dungeons and Dragons Online Plat
 
Post a Comment

<< Home
Archives © Copyright 2005 by Marshall Brain
Atom RSS

This page is powered by Blogger. Isn't yours?