MIT Creates Chip to Model Synapses

MrSeb writes with this excerpt from an Extreme Tech article: "With 400 transistors and standard CMOS manufacturing techniques, a group of MIT researchers have created the first computer chip that mimics the analog, ion-based communication in a synapse between two neurons. Scientists and engineers have tried to fashion brain-like neural networks before, but transistor-transistor logic is fundamentally digital — and the brain is completely analog. Neurons do not suddenly flip from '0' to '1' — they can occupy an almost-infinite scale of analog, in-between values. You can approximate the analog function of synapses by using fuzzy logic (and by ladling on more processors), but that approach only goes so far. MIT's chip is dedicated to modeling every biological caveat in a single synapse. 'We now have a way to capture each and every ionic process that's going on in a neuron,' says Chi-Sang Poon, an MIT researcher who worked on the project. The next step? Scaling up the number of synapses and building specific parts of the brain, such as our visual processing or motor control systems. The long-term goal would be to provide bionic components that augment or replace parts of the human physiology, perhaps in blind or crippled people — and, of course, artificial intelligence. With current state-of-the-art technology it takes hours or days to simulate a simple brain circuit. With MIT's brain chip, the simulation is faster than the biological system itself."

Comments

Re:The Interface will be a problem.

[ledow]

I think the REAL problem is that even the smallest brains have several billion neurons, with each having 10's of thousands of connections to other neurons. This chip simulates ONE such connection.

That's a PCB-routing problem that you REALLY don't want, and way outside the scale of anything that we build (it's like every computer on the planet having 10,000 direct Ethernet connections to nearby computers - no switches, hubs, routers, etc. in order to simulate something approaching a small mouse's brain - not only a cabling and routing nightmare but where the hell do you plug it all in?). Not only that, by a real brain learns by breaking and creating connections all the time.

The analog nature of the neuron isn't really the key to making "artificial brains" - the problem is simply scale. We will never be able to produce enough of these chips and tie them together well enough to produce anything conventionally interesting (and certainly nothing that we could actually analyse any better than the brain of any other species). If we did, it would be unmanageably unprogrammable and unpredictable. If it did anything interesting on its own, we'd never understand how or why it did that.

And I think the claim that they know EVERYTHING about how a neuron works (at least one part of it) is optimistic at best.

But what about efficiency?

[Pegasus]

It may be faster, but what about performance per watt? You know, the whole brain does everything on only 40-50 watts. How does this MIT product compare to brains in this area?

Re:I have my doubts

[jpapon]

You don't see how it's ethically loaded? Really?

Would the artificial brain have rights? If you wiped its artificial neurons, would it be murder? If you give it control of a physical robot arm and it hurt someone, how and to what extent could you "punish" it? The ethical questions are virtually endless when you start to play "god". I would think that would be obvious.

Re:The Interface will be a problem.

[ultranova]

That's a PCB-routing problem that you REALLY don't want, and way outside the scale of anything that we build (it's like every computer on the planet having 10,000 direct Ethernet connections to nearby computers - no switches, hubs, routers, etc. in order to simulate something approaching a small mouse's brain - not only a cabling and routing nightmare but where the hell do you plug it all in?). Not only that, by a real brain learns by breaking and creating connections all the time.

A single neuron-neuron connection has very low bandwidth, in effect transferring a single number (activation level) a few hundred times a second. Even if timing is important, you can simply accompany the level with a timestamp. A single 100 Mbs Ethernet connection is easily able to handle all those 10 000 connections.

Also, most of those 10 000 connections are to nearby neurons, presumably because long-distance communication involves the same latency and energy penalties in the brains as it does anywhere else. There are efficient methods to auto-cluster a P2P network so as to minimize total length of connections, for example Freenet does this; so, you could, in theory, run a distributed neural simulator even on standard Internet technology. In fact, I suspect that it could be possible to achieve human-level or higher artificial intelligence with existant computer power in this method right now.

So, who wants to start HAL@Home ?-)