From neural networks to mechanical networks...
...and vice versa!
So called physical models consitute a powerful technique to simulate the mechanics of a music instrument to reproduce its sound behavior.
Mass-spring networks in particular allow, from very simple building blocks (mostly mass and damped springs) to simulate several kinds of instruments (chords, drums, piano, etc.).
But finding the right mechanical network that will reproduce a give, music instrument is a hard problem. Considering these mechanical networks as "neural" networks opens the way to original algorithms solving this problem, inspired from "Back Propagation Through Time" and Real Time Recurrent learning".
Furthermore, the mechanical nature of these networks gives new ideas on learning architectures: coupling the learning network with the learned instrument, or letting a human play in real time with the (instrument,model) couple in a "pedagogical" way to facilitate learning.