simulated brain on supercomputer

[Admin]

For those interested in AI, from the article:
"A network of artificial nerves is growing in a Swiss supercomputer -- meant to simulate a natural brain, cell-for-cell. The researchers at work on "Blue Brain" promise new insights into the sources of human consciousness."

"This unprecedented piece of hardware consists of about 10,000 computer chips that act like real nerve cells. To simulate a natural brain, part of the cerebral cortex of young rats was painstakingly replicated in the computer, cell by cell, together with the branched tree-like structure of the synapses."

"A whole processor is currently needed to simulate the behavior of a single cell."
(I guess this shows that simulating a brain requires much more processing power than just building one that works)

http://www.spiegel.de/international/spiegel/0,1518,466789,00.html (part 1)
http://www.spiegel.de/international/spiegel/0,1518,466789-2,00.html (part 2)

My rambling thoughts:
I can imagine a human brain version of one of these being not too much greater of a step . . . minus the body to interact/learn with the world, of course . . .

In the future, I can see researchers inputting DNA, and then simulating growth of proteins and such to form an entire simulated brain. No need to input neurons 'painstakingly,' just copy/paste some DNA, add in a sprinkle of nurture, and its set to go.

Anyway, I see this project on the right step

[Nyx]

I don't think they've even completely mapped the neural pathways of the brain of an ant yet... Not to mention that rats have close to 1 million neurons, so 10000 processors simulating one neuron each most likely won't do.

In my opinion, money would be better spent trying to do a complete mapping of the brain of insects, or small animals, and then perhaps larger animals, in the hope of understanding how it all fits together, rather than trying to simulate a tiny fraction or something we know so little about.

[Somchaya]

Wow, I read the article and think it's very inspiring.. I agree that we should start mapping simple creatures first and slowly move upwards.

I do think some research into insects' brains (do they have brains?) have already been done. I remember a prof telling us about how they hijacked a cockroach's brain to remote-control it via a joystick. I'm not sure how detailed the analysis of the brain was done then (like whether or not they mapped all the neurons or just stuck electrodes in that worked) but some research into the neurons was probably done.

I think this project is really cool and if it works within a nice time scale, that would be really great!

[JesseWelling]

Nyx wrote:

Not to mention that rats have close to 1 million neurons, so 10000 processors simulating one neuron each most likely won't do.

Last time I checked 1 processor could do more calculations than 1 Neuron. The issue you have to contend with is that Processors can calculate atomicaly, but can make fewer connections.

[Nyx]

Nyx wrote:

Not to mention that rats have close to 1 million neurons, so 10000 processors simulating one neuron each most likely won't do.

Last time I checked 1 processor could do more calculations than 1 Neuron. The issue you have to contend with is that Processors can calculate atomicaly, but can make fewer connections.

"A whole processor is currently needed to simulate the behavior of a single cell."

I do think some research into insects' brains (do they have brains?) have already been done. I remember a prof telling us about how they hijacked a cockroach's brain to remote-control it via a joystick. I'm not sure how detailed the analysis of the brain was done then (like whether or not they mapped all the neurons or just stuck electrodes in that worked) but some research into the neurons was probably done.

Of course insects have brains. The cockroach thing was done by playing with the roach's antennae, so there was no direct interface with the brain.

As far as I know, the biggest problem is that researchers have no effective tools to precisely map map the circuitry inside a brain. Dissecting won't do, 3D medical scanning technologies are just not that precise. I've heard of an experiment where, to map visual pathways in monkeys, they placed a probe at various locations and tested whether or not they got a signal when the money saw a dot in different places on a screen, for example.

Ideally, in order to study this, we would need some sort of 3D magnetic resonance scanner that is precise up to a few microns, to scan specific regions and create maps of the layout. This may not be doable for a complete brain, but even if such a scanner could only study little bits at a time, scientists could make small slices, and it could still be practical.

[JesseWelling]

"A whole processor is currently needed to simulate the behavior of a single cell."

I'm going to restate my point clearly. I think it's a waste of time to do a simulation of a cell when a single
microchip is more capeable than a single cell. In part because there is more processing power. And also
in part because in the case of simulating Neurons, connections are limited.

From Wikipedia on Neuron:

Neurons communicate with one another via synapses, where the axon terminal of one cell impinges upon a dendrite or soma of another (or less commonly to an axon). Neurons such as Purkinje cells in the cerebellum can have over 1000 dendritic branches, making connections with tens of thousands of other cells; other neurons, such as the magnocellular neurons of the supraoptic nucleus, have only one or two dendrites, each of which receives thousands of synapses. Synapses can be excitatory or inhibitory and will either increase or decrease activity in the target neuron. Some neurons also communicate via electrical synapses, which are direct, electrically-conductive junctions between cells.

Why not play more to the strengths (processing power) instead of simulation of a paradigm that is
below the capeabilities of the hardware? I think they should go after making a brain out of an FPGA.

However, this is a simulation to learn about how humans work, not how to make computers work better.
That's the value I see in it. I don't think there are going to be any huge strides in AI untill we under stand
human intelligence, and that's what I think is noteworthy about this. Intead of coming up with solutions
they are coming up with "the way of solving problems" but they will have a long path to travel because most
of the populace rarely thinks about "the way of solving problems"    any ways. </rant>
 

[hgordon]

There's a lot more detail on the BlueBrain website - http://bluebrain.epfl.ch/ .  Basically, they're simulating a "column" of 10,000 fully interconnected neurons (each with up to 10,000 connections) using 8192 processors, and at present, the simulation runs about 100x slower than real-time, though the goal is to optimize calculations to bring that down to 1x, as well as scaling the number of processors to model much larger collections of neurons.  With this type of architecture, I imagine they will be able to model a variety of neurological processes.  Check out some of the videos on the website - they've done some impressive work.

[Admin]

Jesse, I agree with you that simulating neurons at the molecular level is very much inefficient for processing, but thats not what they are doing . . .

They are trying to figure out why the brain works at the molecular level, not make a computer capable of calculations equal to that (or greater) of a brain

This technique is a crazy overblown technique already commonly used by biochemists - simulate molecular interactions when the traditional lab tests are more difficult or less useful to perform. (in fact, a biochemist sent me this article)

Other uses they stated were to test out current theoretical models of the brain and as a medical tool:

"But the main purpose of the artificial brain, say its creators, is to make new types of experiments possible. For example, what happens when damage is inflicted on certain types of cells whose function still isn't determined? How many cells can be switched off until the behavior of the surviving cells around them becomes erratic, or the entire circuit breaks down? Scientists know that a similar process occurs in the brains of epileptics and Alzheimer's patients. Armed with current methods, which by their very nature are based on external observation, medical science has gained only a schematic picture of the processes in these patients' brains. Using Blue Brain, researchers would simply fly to the crisis zones inside a virtual skull.

Such studies are conceivable, though, only in a model that mirrors biology. The so-called neural networks that other researchers have developed for years are very different. A neural network is also meant to behave like a brain, somehow, but how the likeness is achieved -- the circuitry underneath -- is more or less irrelevant. Someone building a cow using this principle would be content with any milk machine that moos and produces the occasional cowpat.

"That doesn't help us to understand the biology," says Markram. The researchers in Lausanne are interested in the real cow: "Our first priority is that we never use tricks to achieve the correct result," says project manager Schürmann. "If something goes wrong in the simulation, our only option to improve it is by incorporating new biological knowledge.""

[ComputerGeek]

Knowing what I know do about programming and my limited knowledge of psychology couldn't a 'brain' be simulated by a switch/case statement nested with an unreasonable number of environmental variables(I believe life to be stressful  )? If the goal for this is for cybernetics or nero-prostetics then why do we need to open the black box? Why not just take the out put and roll with it? The approach to this should be the same as a fluid dynamics model where you don't try to map every atom or molecule but instead you assign a set of properties to the mass that represent the flow/rate of diffusion. I'm checked to be notified of replies so not be argumentative but why are they seeming to take the long way? P.S. Shouldn't we have learned with Pi that you can not map an analogous concept with the digital math that we use today?

[hgordon]

... couldn't a 'brain' be simulated by a switch/case statement nested with an unreasonable number of environmental variables ...

That's certainly the approach that was taken by early AI researchers with memory models and expert systems, and it has long been out of favor, as these models work fine with inputs they've already seen, but are lost when information has to be processed out of context.

BlueBrain is an incredibly complex bit of machinery by current standards, yet it's 6-7 orders of magnitude lower in complexity than the human cerebral cortex (assuming BlueBrain could compute in real-time, else add a couple more orders of magnitude).  However, by Moore's law, we're only 20-25 years away from achieving comparable computational density, so why not try to model the real thing ?