April 29, 2007

...Learn TDD with Codemanship

Mouse Brain Simulation Breakthrough

Champagne corks may have been popping in the AI research community after news that IBM's BlueGene L supercomputer has successfully simulated "half a mouse's brain".

More specifically, they have simulated a neural network roughly as complex - though still some way off in the structure and details - as half that of the brain of one of our little furry friends. That's about 8 million neurons, each with about 6,000 connections to other neurons.

The simulation was equivalent to about 1 second of activity in a mouse's noggin, and took about 10 seconds to run: such was the complexity, they could run it no further, despite the fact that their supercomputer had 4,096 processors, each using 256Mb of memory.

BlueGene L - five seconds into the mouse brain simulation it ran off and hid behind the skirting board

Even so, in those few seconds, researchers claimed to have seen some evidence of "biologically consistent dynamical properties". The team are now working on speeding up the code and refining the overall simulation to make it more realistic.

On the back of an envelope, I've calculated that they might be able to simulate a whole mouse brain in real time with a computer 20 times as powerful as BlueGene L. At the present rate of processor development, that's probably only 6-8 years away. Even then, their simulation is only a simplified abstraction of the real thing, so maybe we should double that power, just to give them time to write the software - so that's another year or two*.

So, in about 2017, could we expect a supercomputer as smart as a mouse? Well:

* Yes - if that's all there is to it, but
* No - if there's some non-computable element to animal intelligence

Right now, after reading Roger Penrose's arguments against "strong" AI, I've got one foot and four toes in the no camp. I'm 90% certain that when they've simulated an entire mouse brain with sufficient realism, that's when we'll find out that there really is more to it than super-connected, massively parallel neural networks.

* The human brain has about 100 billion neurons, each connected by about 6,000 synapses - so a simulation of a human brain might require a supercomputer 6,000 times as powerful as that required for the mouse brain. Such a machine might be available within a further 20-30 years, perhaps by 2037?
Posted 13 years, 7 months ago on April 29, 2007