Search This Blog

Friday, October 19, 2007

Carver Mead on Computing

Listen to Carver Mead’s Gilder/Forbes Telecosm 2006 speech.

"I’ve said this every year for the last ten years. With all of our yotta yottas we still can’t do the computation that’s done by a common housefly. The DARPA “grand challenge” which filled up an entire van with computing stuff of the latest sort was pathetic compared with a housefly. For those of you who have tried to kill them, they’re not too easy to get rid of … You may think they’re stupid, but not nearly as stupid as our computers. And they do all that on a few milliwatts.

"If you really want more bang for a given amount of energy – more information processed, the real metric of “goodness” in this world – you divide the computation into more and more parallel things that go slower. You don’t try to make each one go a little faster; you just put lots and lots of them in parallel. But actually, we don’t really know how to do that …

"What is it about that goo in the brain of a fly that can do that? We have all of our yotta yottas and we’re still not able to do it …

"There’s nothing about the physics of goo that works with ions and membranes that can’t work with electrons and insulars. There is no reason that if we understood what this absolutely fantastic, remarkable structure is doing, that we can’t learn from it and develop a computational paradigm, which is completely different from anything that we know or have even imagined. That’s what got me into neuroscience. Because in the end, we have to make more and more parallel systems. The degree if parallelism is, in the end, going to be the efficacy of our information processing systems, and here is a working system that has the ultimate parallelism.

“So as we look at the second decade of the Telecosm, I would submit to you that we’re not really burned out. I think there is plenty to think about."

Click here to listen to the audio. Just under an hour but worth it!


Bret said...

Mead said: "...that we can’t learn from it and develop a computational paradigm, which is completely different from anything that we know or have even imagined."

Ummm. Perhaps he ought to read Hecht-Nielsen's book on Confabulation Theory?

Howard said... reason...we can't learn

a double negative

we can learn

Bret said...

My point is that I think perhaps some stuff has been developed that Mead doesn't know about or he has yet to imagine it.

Mead is innovative but tends to get stuck on things (like the silicon eye, for example) that are both ahead of their time and too late at the same time. Ahead of its time in that perhaps one day in the future his approaches will be competitive but, today, their impact lags what can be done with standard technology.

Howard said...

Yes, there are probably areas of AI related advances that Mead is unaware of and he does sometimes get stuck. I still found his description of the structure and dynamic range of biological visual systems most interesting.