Medium Independence.
The neuron models are medium independent. They may be built out of
organic materials, electrical components, lenses and mirrors,
hydraulics or dominoes. The electronic implementation of a neuron [below]
is based on an operational amplifier (op-amp comparator).
![[Schematic of
electronic neuron]](neuron6.gif)
These versatile components can determine which of two voltages is larger.
They are accurate to about 0.001 volts. The three potentiometers (variable
resistors) are used to set the neural weightings in this implementation.
If one is set midway, then the associated input would have no effect on
the system because the voltage would be evenly applied between the two
op-amp inputs, resulting in a weighting of 0. Set the potentiometers
towards the top, and the op-amp is positively biased. Since op-amps have
a natural threshold of 0 volts difference, an extra bias input is required.
Any number of additional inputs and potentiometers may be inserted. It is
ironic that the lowly operational amplifier that was developed for the
long-obsolete analog computers, is making a come-back as dedicated computer
systems start to use neural networks instead of (or in addition to)
conventional microprocessors.
![[Photograph of op-amp neuron]](neuron7.gif)
To create a large neural network, one could either construct thousands of
op-amp circuits (like the one in the photograph) in parallel, or one could
merely simulate them using a program executing on a conventional serial
processor. From a theoretical stance, the solutions are equivalent since
a neuron's medium does not affect its operation. By simulating the neural
behaviour, one has created a virtual machine that is functionally
identical to a machine that would have been prohibitively complex and
expensive to build. A computer's flexibility makes the creation of one
hundred neurons as easy as the creation of one neuron. The drawback is
that the simulated machine is slower by many orders of magnitude than a
real neural network since the simulation is being done in a serial
manner by the CPU.
Previous: A Model of a Neuron.
Next: Training Nets.
Last modified: September 21, 1998
By: Neil Fraser
(neil@vv.carleton.ca)