Quantum reservoir computing

Most of a recurrent network's power comes from its rich internal dynamics, not from tuning every weight. Reservoir computing leans into that: leave a large, fixed, randomly-connected dynamical system untouched, let the input stir it into a high-dimensional trace of its own recent history, and train only a single linear layer to read the answer out. The reservoir supplies memory and nonlinearity for free; the learning is a one-shot least-squares fit. Drive one below and find where it remembers best.

A fixed reservoir, a trained readout

The reservoir state $\mathbf{x}$ updates by a leaky, saturating recurrence — a fixed random matrix $W$ scaled to a chosen spectral radius $\rho$, plus the input. Only the readout weights are fitted, by ridge regression, to a target. Here the target is a delayed copy of the input, $y_t=u_{t-k}$: recovering it with a linear readout is possible only because the reservoir still carries an echo of $u_{t-k}$ in its present state. How long that echo lasts is set by $\rho$ and the leak — the reservoir's fading memory.

$$\mathbf{x}_t = (1-a)\,\mathbf{x}_{t-1} + a\,\tanh\!\big(\rho\,W\mathbf{x}_{t-1} + W_{\rm in}u_t\big),\qquad \hat y_t = W_{\rm out}^\top[\mathbf{x}_t;1]$$

Live — the classical echo-state baseline on a memory task (exact library kernel)

target y = u(t−k) readout prediction error vs ρ

spectral radius ρ 0.90 leak rate a 0.60 recall delay k 1

readout MSE: — reservoir size: — nodes trained weights: readout only

The left panel overlays the delayed target and the reservoir's readout of it; when they sit on top of one another, a single linear layer has recovered a memory the raw input alone could not supply. The right panel sweeps the spectral radius and re-fits the readout at each: the error falls into a trough near the edge of chaos ($\rho$ a little below one), where the reservoir's dynamics are lively enough to hold a long memory but not so wild that they wash it out. Ask for a longer delay and the whole trough lifts — memory is finite.

Deeper: why this is the baseline a quantum reservoir must beat

This is the deterministic classical echo-state network the library carries as its yardstick: a quantum reservoir replaces the $\tanh$ state vector with the measured features of a driven quantum system, but is scored on exactly this kind of task against exactly this readout, so any claimed advantage is a like-for-like comparison rather than a favourable framing. The panel runs the library's reservoir-and-ridge kernel — the leaky recurrence and the least-squares readout — on embedded fixed weights; its predictions and errors match the library reference to machine precision. The library's quantum-reservoir surface is explicit that it certifies only deterministic synthetic tasks with exact-statevector features, and makes no claim of hardware deployment or quantum advantage.

Where it fits

Reservoir computing is a natural home for a coupled-oscillator quantum device: the very dynamics that are hard to program can instead be left alone and used as the reservoir, with only a cheap classical readout to train. It rhymes with the rest of the portal — a fixed rich dynamical system doing the work, a light touch of learning on top — and it is the setting in which a quantum substrate is compared, honestly and on the same task, against a strong classical baseline.

Piece	Role	Trained?
Reservoir $W$	fixed random recurrence, supplies memory & nonlinearity	no
Spectral radius $\rho$	sets the memory length; best near the edge of chaos	chosen
Leak rate $a$	how fast the state follows the input	chosen
Readout $W_{\rm out}$	linear map to the target	yes — one ridge fit

Coupled-oscillator networks Applications Quantum neuromorphic Back to overview

Evidence boundary: the panel runs the library's classical echo-state reservoir and ridge readout exactly, on embedded fixed weights, live in your browser — predictions and error matched to the library reference to machine precision. It demonstrates the reservoir-computing principle and is the classical baseline for the quantum reservoir; it is not itself a quantum run, and the library flags quantum advantage for reservoir computing as an open question on synthetic tasks only.