Definition
Plain language
Using a fixed random system as a scrambler and only training a simple readout on top.
As stated in the literature
A computational paradigm where a fixed dynamical reservoir provides nonlinear features and only a linear readout layer is trained.
Why it matters: It's a reminder that much of what neural networks do can come from fixed nonlinear scrambling plus a tiny trained readout, which is both theoretically interesting and very cheap.
For example, a network of randomly connected, untrained units processes a time series and a simple linear classifier reads off the answer from their activity.
Heard on the show
“The first is a known approach — using the scattering medium as a single-token embedding engine, basically optical reservoir computing.”Episode 002 — An AI Ran a Real Optics Lab for 21 Hours and Found a Transformer-Shaped Pattern in Light