[sdiy] help interpreting diagrams with Mosfets

[David G. Dixon]

> training is something i keep hearing about in passing but 
> don't really understand
> 
> i heard someone say (in the SDIY archives i think) that they 
> "trained" a circuit FIRST with a calculator, then they took 
> the results and put it into a schematic.
> 
> so i think you train it first?

>From my very limited understanding, I believe that the "training" of
artificial neurons is simply the adjustment of a collection of weighting
factors on the inputs (i.e., gain factors) until the neuron (or net of
neurons) produces the desired output given certain inputs.  It's not unlike
a mathematical optimization or error minimization problem.  The weights are
often adjusted automatically using feedback.

When a neural network is used to control complex industrial processes, for
example, the network must first be calibrated with a fairly large number of
scenarios over a broad statistical range before it can be "let loose" on the
process itself.  This training can be done either with the real plant
(controlled with more traditional means) or with a simulator.  The advantage
of neural networks (as far as I can tell) is that they can be used very
effectively even if absolutely nothing fundamental is known about the
processes they are controlling.  In other words, they allow one to treat a
process like a "black box".  As long as enough inputs, outputs, and feedback
loops are connected and the network is trained with enough operating data,
it will work, provided that the behavior of the process is pseudo-linear
(meaning there are no sudden catastrophic changes in process performance
with slight tweaks in one or more variables).