Mar 26 2009

Brain on a Chip

Published by under Uncategorized
Comments: 24

Work continues to develop computers which more closely simulate brain activity. As I recently discussed, the brain is not a computer, and computers are not brains. But since brains operate by physical laws, there is no theoretical reason why we could not design a computer to function like a brain.

There are two types of approach to this – virtual brains and hardwired brain-like computers. The former approach uses conventional computer hardware to simulate a virtual brain. This project is far from achieving this goal, but programmers have already achieved significant milestone, such as the simulation of a cortical column of 10,000 neurons and 30 million connections.

The advantage of the virtual approach is that brain structure can be simulated directly without worrying about having to engineer functional neurons or their connections. The downside is that virtual brains require much more processing power than the brains they simulate. So virtual brains run incredibly slowly. It will likely still be 2-3 decades before raw computing power reaches the level where a desktop computer will be able to simulate a human brain in real time. (I am just talking about raw power here, not the knowledge to actually pull off the simulation.)

The advantage of the hardwired approach is that you end up with a processor that is much faster than a biological brain. The challenge, however, is in building something in silicon (or whatever) that acts like biological neurons in all the important ways.

An international team of researchers has taken a step in this direction.The project is called Fast Analog Computing with Emergent Transient States, or FACETS. Their latest product is a silicon chip that has 200,000 “neurons” which make 50 million synapses. This is still tiny compared to the human brain, which has 100 billion neurons. But they plan to connect a bunch of these wafers together into a superchip with 1 billion neurons and 10^23 synapses – still only 1% of the way to the power of the human brain.

The structure of these chips uses transistors and capacitors to recreate the same kinds of connections that neurons make. Project member Henry Markram has developed an algorithm for the synapses that he calls spike-timing dependent plasticity, which is designed to allow the device to learn and adapt. The chip is designed, like the brain, to be massively parallel. Traditional computers are serial – they can crunch numbers very quickly, but they have one stream of processing.

This chip design, it is claimed, combines the power of parallel and adaptive processing with the speed of silicon. They claim it can function 100,000 times faster than a human brain.

While the purpose of creating virtual brains is primarily to study the brain (although this line of research may result in artificial intelligence – AI), the purpose of this chip development is not to study the brain at all, but rather to develop more powerful computers and possibly computer AI.

Of course, while this is an important step, what FACETS is developing is not a silicon brain but a chip that may serve as a basic unit in such a brain. You cannot, in my opinion, just stack a bunch of these chips together and get AI.  A higher order of architecture is required. I guess, although I don’t know, that advancements in this technology – building silicon neurons – will advance more quickly than our understanding of brain architecture and functioning. The latter is progressing nicely, but the level of complexity is much greater.

So we may get to the point where we can build a massively parallel and learning silicon computer with the memory and processing power of a human brain, but not know how to put it together to create something we would call AI or, even more difficult, conscious.

I am curious as to which approach to designing the overall structure of such a silicon brain will take. The two basic approaches I am thinking of are 1 – duplicating the human brain, or 2 -designing it through trial and error from the ground up. Likely a combined approach will be taken, but we’ll see.

Also we have to consider what function we wish for the end result. Will it be placed in a robot which will serve as its interface with the outside world, or will it be on a desktop, or will it be the brains of a commercial jet or some other piece of equipment? Or – will it be implanted inside a human to serve as a supplemental brain?

I suspect that the approach that will emerge is to first design simple functions, like seeing, interacting with objects, controlling whatever physical form it is in, etc. And then progressively more complex “higher” functions will be added that control the more basic functions. This is exactly how the mammalian brain is organized. This process will be informed by our knowledge of how the brain is organized and functions, but will also have to learn through trial and error to fill in the gaps in our neurological knowledge.

And of course it is always fun to speculate about where this will lead – to Cylons, the Matrix, or to benign and obedient droids. Let’s hope it’s the latter.

24 responses so far