Reservoir computing says don’t train the dynamics, just train the readout. I used the SID’s analog filter as the dynamics and one ridge regression as the readout. It’s not state of the art, but the chip earns its keep exactly where a linear model falls apart.

The fruit fly has a brain structure that acts like a hash function. I built it in Python, matched the 2017 paper’s MNIST results against a classical baseline, then got the same algorithm running on a Commodore 64.