🚀 go-pugleaf

In article <4345@map.mdavcr.mda.ca> garry@mdavcr.mda.ca (Garry Holmen) writes:
>>>|> Ideally, NNs should be analog machines (... that takes us quite far
>>>|> back to the past). Unfortunately, the technology is not yet in place
>>>|> for that. I am a true believer that it is only a matter of time,
>>>|> though.
>>>
>
>NN's do not need to be analog machines. Some networks perform quite well on
>digital machines. Of course if you're implementing a Net based on arithmetic
>propagation as is used in backprop and many others an analog machine would be
>best.

The analogue/digital distinction is even more blurred than you might
imagine. You can have hardware digital (ie no software) or hardware
analogue simulated to nth degree accuracy on a digital Von Neumann machine.
Or you can have hybrid semi-digital nets (a la Hopfield), which have analogue
stages during problem solving before digital solution-set resolution, and
these can either be simulated by software (to nth degree ) and run on a VNM
or even built as hardware nets. With the advent of the Hillis connectionist
machines this picture will become even more "interesting"...

--

 / Craig Webster                      CRAIGW@pretzel.equinox.gen.nz /
/           ...now she walks through a sunken dream...             /