Has anyone created a neural net where the weights are expressed
as varying brightness of LEDs? That would make it interesting to
watch how the net learns and how it responds to queries.<p>When the net used 4 bit weights the LEDs would be in one of only
16 difference brightness levels.<p>Instead of LEDs, which would draw a lot of power for a 1M array,
it would be useful to use one of the huge screen TVs.<p>Even more interesting would be the ability to see how the various
"expert" subnets interact.
The original perceptron machine built by Rosenblatt used self-rotating potentiometer knobs to visualize and change weights for a perceptron that used a cadmium sulfide "pixels" camera.<p>I believe you could use LEDs and if you were really clever, you can use the fact that LEDs are also light sensitive to adjust the LED "weights".