The neuron as an analog NOT digital system

[Jan Vorbrueggen]

In article <40787d$q6n at news2.deltanet.com> ma2 at delta1.deltanet.com (Marc
Anton) writes:

   Many scientists and philosophers today seem to be guilty of something I'll
   call "neural reductionism", a contemporary epistemic malady where otherwise
   infinitely complex neurons are "reduced" to the simple on and off firings
   we'd like to associate with today's computers.

If the thing is "infinitely complex", we might as well stop doing neuroscience
right now. OTOH, I know of no neuroscientist who would compare the "on and off
firings" of neurons with a computer's binary modus operandi. The similarity of
digital signals and axonal spikes is very superficial at best. 

   Though we tend to think of the neuron as essentially a digital processor
   (with its on and off firings) it's really closer to an analog processor
   with infinitely complex chemical and electrical functions above and beyond
   anything capable with digital reductionism.

And _really_ nobody, not even the people using the most simple of neural
network models, think of a neuron as a digital processor. What do you think
real-valued weights in such models are used for? Backpropagation doesn't work
if the signals and their squashing function aren't differentiable. Again, OTOH,
every such analog thing _can_ be modelled using a digital, discrete system -
it's just a question whether, when you have taken the necessary accuracy into
account, it is still cost-effective and yields results within your lifetime.
There is no such thing as being "above and beyond" digital simulation.

	Jan