« Earth Day Essay | Main | Don't Pack Your Bags Just Yet »

The Early Signs of the Long Tomorrow

pbrain_sq.jpg(Or "I, for one, welcome our new cyber-mouse overlords!")

Ahoy, BoingBoing readers! I was going to update this anyway, but with the BB link, it's extra-important: this is a simulation of a cortical network with the size, link complexity and signal activity of a mouse brain, but without the structure -- so, arguably, it isn't a really a simulated mouse brain, but a functional platform upon which a mouse brain sim could run. Depending upon your perspective, this is a minor quibble or makes all the difference.

It's hard to see this as anything but a distant early warning of some pretty remarkable changes on the near horizon. IBM researchers James Frye, Rajagopal Ananthanarayanan, and Dharmendra S. Modha assembled a simulated mouse cortical hemisphere (that is, a functional half of a mouse brain) on one of the smaller BlueGene/L supercomputers. They then ran the simulation -- at ten seconds of computer processing equal to one second of brain function.

In other words: they ran a simulated mouse brain at 1/10 time.

Neurobiologically realistic, large-scale cortical and sub-cortical simulations are bound to play a key role in computational neuroscience and its applications to cognitive computing. One hemisphere of the mouse cortex has roughly 8,000,000 neurons and 8,000 synapses per neuron. Modeling at this scale imposes tremendous constraints on computation, communication, and memory capacity of any computing platform.

We have designed and implemented a massively parallel cortical simulator with (a) phenomenological spiking neuron models; (b) spike-timing dependent plasticity; and (c) axonal delays.

We deployed the simulator on a 4096-processor BlueGene/L supercomputer with 256 MB per CPU. We were able to represent 8,000,000 neurons (80% excitatory) and 6,300 synapses per neuron in the 1 TB main memory of the system. Using a synthetic pattern of neuronal interconnections, at a 1 ms resolution and an average firing rate of 1 Hz, we were able to run 1s of model time in 10s of real time!

The team published the write-up in the February 5, 2007, edition of Computer Science; a PDF is available of the one-page research report, providing a few technical details.

The human brain has some 100 billion neurons, so this mouse brain simulation is still about 1/12,500 of a simulated human brain. That may sound like a daunting challenge, until a glance at computer history makes clear that such computational capabilities will likely be possible on within 20 years, easily, if not even sooner.

But well before that point, we'll be able to run simulations of animal brains at accelerated speeds, raising a provocative test of just how important raw cognitive speed is to the emergence of artificial intelligence. Would an accelerated mouse brain simulation simply be a fast-calculating mouse, or will it have other properties and capabilities deriving from the sheer speed? Which would be smarter -- a 6,000X faster mouse brain sim, or a 1/2-speed human brain sim?

Some of that is going to depend upon how much of the simulation models actual brain structure, rather than simply the number of connections. That's likely to be crucial. The brain isn't simply a haphazard mass of neural junctions, and a functional structure simulation may well prove to be a far greater challenge than simply getting the neural connection sim working. Still, this is not an unsolvable problem, by any extent.

But this raises the question of whatt kinds of programming will be possible with these simulated brains. The IBM simulation simply showed that a functional simulation was possible; evidently, they didn't try to do anything with the cyber-mouse. It's not entirely clear what could be done with it. We're now on the brink of facing a question that had, in the past, been essentially the province of science fiction:

How does one program a simulated mind?

(Thanks to Miron for the tip!)


Interesting. However, remember that they basically "just" simulated a brain; this is a long way from simulating intelligence. Thile there is a lot of benefit possible to cognitive science here, it does not really imply much about AI or our understanding of intelligence.

Presumably there'll be some interesting insights along the way as to how important the neural structure's embodiment is. Where does that "functional structure" stop? The structure evolved to complement and work with a fully functioning mouse body, which in turn evolved to complement and work with a fully functioning ecosystem. Korzybski said we think as much with our big toe as our brains; Levi-Strauss may have added that we think as much with the natural environment as with our bodies and minds.

I guess these neural models might end up being re-purposed for unexpected, more computational and less functional-in-the-world applications before we develop their biomechanical bodily counterparts.

I don't know whether to be excited or terrified.

I'm neither excited nor terrified by this. I think it is an interesting simulation though. Hopefully there will be some insight into neural function that results from it.

>George's jaw drops

Howard, the implication here is that, like natural intelligence is an emergent result of brain structure and activity, artificial intelligence (of a mousey sort) could be an emergent result from a further development of this. Hence my emphasis on structure -- potentially, this can give us insights not just into the mechanics of cognitive networks, but into the function of cognitive networks, as well.

I understand that; my point is that brain structure + random activity does not equal intelligence, or even necessarily a meaningful probability of it emerging.

However, it may lead to a better *understanding* of how neural systems interact.

I think we're in agreement there, Howard

Huh? They did NOT simulate a mouse's brain. They simulated a bunch of random neurons. A mouse's brain has very specific connections between neurons that are tuned by a very complex process of embryonic development, neural development and pruning, interacting with sensory information from the body and the world. They simulated a mouse's brain, if the brain had been pureed in a blender. Although it's great that this computing power exists, without a real understanding of how the neurons connect, and how different connections all work, it's just a toy.

Hi Harlan

As you might have noticed, I've updated the text to make clear that this doesn't simulate the structure of a mouse brain. It is more complex and important than a pureed gray matter sim, but -- as you correctly note -- it's a cortical network devoid of the specific connection pattern of a real brain.


Creative Commons License
This weblog is licensed under a Creative Commons License.
Powered By MovableType 4.37