Josh Bell wrote:
> "Dan Minette" <[EMAIL PROTECTED]> wrote:
> >
> >How do these filters work? They cannot work algorithmically, that's been
> >proven rigorously. Penrose didn't prove it, but he has some nice proofs
of
> >the limits of algorithmic proofs in his book "the Emperor's New Mind"
>
> Are you familiar with neural networks at all?
>
Yup. From everything I've seen, they've been vastly oversold. I understand
the principal, and actually think they can work in cases like recognition of
military hardware from pictures (an application a former boss of mine used
it for) because there it is used to recognize various patterns and then put
the patterns together. But, I've more often seen it used as a buzz word by
snake oil salesmen. My understanding is that they systems are typically
very unstable, and are hard to keep working.
> The point is that they do very simple computations in a massively parallel
> way and get approximate answers.
Well, the other point is that they output values that are inputs to
neighboring cells. But, their output is still algorithmic, and they can be
mapped onto a Turning engine. The results still hold.
With regard to the chess example, Deep Blue is massively parallel, but isn't
really a neural network. It actually has a very simple weighing algorithm,
and works best by brute force. The times it appeared to play "like a human"
it was actually looking at weight differences many plies out.
>They are not guaranteed to get an exact answer, and can't follow a
"logical" progression.
They are not guaranteed to get anything close to the right answer, as far as
I've seen. What I think is amazing is that the proponents of neural
networks and strong AI consider the multiple failures meaningless.
AI, remember, was touted 20 years ago as something that was commercial.
>It's like solving a
> NP-Complete problem; if you're allowed to guess intelligently you may come
> up with a 99% answer much more quickly than you could every
algorithmically
> determine a 100% answer.
>
How are neural networks not algorithmic? Each node has a simple set of
rules, a simple algorithm. The timing of them passing information can also
be described by an algorithm. Perhaps timing difficulties exist, so the
same answer will not be obtained twice with the same input, but that can
also be described algorithmically, through the use of a pseudo-random number
generator.
Dan M.
