7 comments

[ 2.4 ms ] story [ 30.7 ms ] thread
> But the human brain, which most certainly must be a kind of computer

Must?

It's one thing to observe that the brain appears to do computation, and to have as a working hypothesis that everything the brain does is a form of computation.

It's something else to say that it must be true, when, for the moment, the brain also appears to do things that we have no idea how to make computers do, that we have significant difficulty getting computers to even pretend to do.

On which sources are you basing your statements? What else could it be, if not a kind of associative/predictive/computation+memory system?

And as to the IBM effort, I think they are doing the right thing. Whatever brings us close to having the hardware that can do the job (needs to be at least on the same scale as the human brain, that features around a petasynaps of connectivity/memory)...

> What else could it be

It doesn't matter what else it could be. Either there's convincing evidence/argument that we can get computers to do everything the brain apparently does, or there isn't.

Right now, we're in the later state.

I dunno. But after I've seen a neuro network dreaming about hand-written digits, I couldn't think about the brain as of something other than a large neuro-network any more. It is still beautiful and magical to me, but it is magic without magic.

The video that I've mentioned is still available here: http://www.cs.toronto.edu/~hinton/digits.html

If you read the paper * , the IBM programming language's model really isn't far from that. Even the article indicates that you are basically building a modular neuro-network.

* Find the link somewhere at http://lambda-the-ultimate.org/node/4797

How they put "most certainly" and "must" in the same sentence is very cool. Probably the author is "a little pregnant"?
Seems somewhat like hype. FPGAs and ASICs are the only fast enough solutions for direct parallel nets of inputs and outputs at the current moment. There is no way a processor is going to perform as well, especially if running instructions from a memory chip.