|
|
 | | From: | wtkiii | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | 24 Dec 2004 12:32:45 -0800 |
|
|
 | Of course, the parallel design of cells makes parallel hardware a good
choice, but systems of 100-1000 cells can be programmed and run on a
PC. In a peripheral program, graphics (I can't do Windows graphics)
or text output can be used to display the state of the cells. By the
way, inheritance isn't needed, composition is enough.
|
|
| | From: | BlackWater | | Subject: | Re: AI will never work in 100 years !!!! | | Date: | Fri, 24 Dec 2004 22:13:07 GMT |
|
|
| On 24 Dec 2004 12:32:45 -0800, "wtkiii" wrote:
>Of course, the parallel design of cells makes parallel hardware a good
>choice, but systems of 100-1000 cells can be programmed and run on a
>PC. In a peripheral program, graphics (I can't do Windows graphics)
>or text output can be used to display the state of the cells. By the
>way, inheritance isn't needed, composition is enough.
Neural nets, or some distillation of their function, CAN
be done using conventional microprocessors. The problem
is the increasing price of parallelization - it's not
just the raw number of 'neurons' but all of the possible
interconnections. Organic neurons may have dozens of
links to others, which may have dozens of links to
others and so on and so forth. The computing task
very rapidly mushrooms, thus limiting the practical
size of your simulations. Too few 'neurons' and you
probably won't get many worthwhile results.
Clearly we need a different approach, something closer
to nature. 3-dimensional programmable gate arrays are
probably required. Even if each simulated neuron works
rather slowly, as do real nerves, the massive degree
of interlinking possible might save the day.
The other approach is to NOT try and simulate real
nerves at all. The olde-tyme AI people tried to
substitute algorithms for 'neurons' - Minskys'
"Society of Mind" is full of this. Alas they were
MISSING something ... there was no 'glue' binding
all these relatively high-level processes together.
They could produce PARTS of thought with minimal
hardware, but the parts wouldn't go together to
create an actual 'mind' worthy of a flea, much
less a human.
Still, emulation of nature will only get us just SO
far ... and then you may as well just stick to nature
and take up genetic engineering. Nature has much to
teach - it's spent billions of years getting things
THIS good - but it's not necessarily the BEST way
to do things with electronics as we understand the
term today. I predict a composite approach - part
'nature', part algorithmic abstraction - will
eventually yeild the best results.
|
|
|
