Indexing the archive…
Your Universe of Digital Possibilities
A tangle of random wires you never train — every recurrent weight left exactly as it was thrown down. You don’t teach it; you rent its echoes. Drive it with a chaotic signal and the reservoir rings with a fading memory of everything it has heard, and from that ringing you fit one single line — a linear readout — to read the future out of it. The whole act of learning collapses to one closed-form solve.
A big pool of neurons, wired to each other by a fixed random matrix W and fed the input through random Win. Its state x is a living, fading echo of everything it has heard — a high-dimensional memory that is never trained, only listened to.
The answer is just a weighted sum of the reservoir’s state. All the learning lives in this one output layer Wout — the recurrent tangle stays random, so training a recurrent network collapses to fitting a single line.
Collect the reservoir states X against the targets Y and solve one regularised least-squares system — no backpropagation, no epochs, one matrix inverse. The ridge term β keeps the readout from chasing noise in the echo.
For the network to be a reservoir at all, it must forget where it started: scale W so its spectral radius — its largest eigenvalue — sits just below one. Below, old inputs decay and the state is a clean echo of recent history; above, the activity never settles and prediction falls apart.
It forecasts the same Lorenz system The Attractor (INST·16) draws, and hits exactly The Divergence’s Lyapunov horizon — the prediction tracks the truth for a few λ-times, then the exponential error growth of sensitive dependence pulls it off course. Like The Shadow it rebuilds a strange attractor from a single scalar signal — but here the machinery is a rented random reservoir: the recurrent weights are never touched, and the only thing learned is one linear readout, fitted in closed form.