Adam V1
A spiking neural network looking through the camera. It was
trained without labels on handwritten digits, and it is running here in the
browser, on your phone, with nothing sent anywhere.
What you are looking at
One hundred neurons. Each one learned, on its own, to respond to some
pattern of ink. The grid shows how many times each fired while it was
looking at the frame. The bars show how close that pattern of firing came to
the average pattern for each of the ten digits.
The bars are the interesting part. A sloppy 3 lights up part of the 8. A
1 sits near the 7. That is the network telling you what it thinks a shape
resembles, which is a different and more honest thing than a single answer.
How well it works
The control
The Control button runs the identical readout on an untrained network of
the same size and shape. If it scores close to the trained one, the readout
is doing the work rather than the network. It is here so that comparison is
visible instead of hidden.
The naming is scaffolding
Calling a pattern of firing "7" takes an outside namer. The network never
discovers that a digit is called seven, and this part of the app is
instrumentation for a person watching, not a part of the network. It comes
out when the system can produce the symbol itself.