Stuart-Landau Network Explorer
Interactive companion to the paper

Task structure and dynamical regimes shape information processing in physical oscillator networks

Every network on this page is a trained Stuart-Landau recurrent network, running live in your browser. Change the oscillator parameters and watch the computation succeed or fail.

input force Re z Im z

Some physical systems compute by moving. A laser, a circuit or a population of neurons takes a signal in, lets its own dynamics churn on it, and reads an answer out. You cannot program such a system line by line. You can only tune its physics: how fast it oscillates, how strongly it is damped, how hard the input pushes. This page lets you do exactly that, to networks that have already learned three real tasks.

The three networks below were trained on real benchmarks: reading handwritten digits one pixel at a time, judging film reviews, forecasting a chaotic signal. Their learned connections never change while you play. The sliders move only the physics underneath, and each task breaks in its own direction. Which regime a task needs turns out to follow from the temporal structure of its input, and that is what the paper reports. Everything here runs in your browser; nothing is sent anywhere.

Start here

  1. The dials in the middle show where the oscillators ended up after the whole sequence. Each circle is one oscillator. Press Play to watch them get there.
  2. Drag λ, the first slider, to the left. Damping rises, the dots collapse to their centres, and the digit readout on the right falls apart.
  3. Press Reset to trained values to bring it back.
  4. Switch to IMDb sentiment and try the same thing. This task fails in the opposite direction, which is the whole point.

Task

input sequence u(t)
What the network is asked to do, and the signal it receives one value at a time. Pick a different example to feed it.

Signal fed to the network

Network of — Stuart-Landau oscillators

each dial is one node's complex plane
The machine itself. Every dial is one oscillator: the dot is its current state, the blue trail its recent path, the amber stub the push the input is giving it right now, and the lines the learned connections. Every pair of oscillators is drawn, in both directions, with weight and opacity set by the strength of that connection, so a dense mesh is a densely coupled network. The arc hugging each dial is that node's connection to itself.
t = 0 / 784

Readout

linear map of [Re z(T), Im z(T)]
The answer. A single fixed weighted sum of the oscillators' final positions, nothing more. If the answer is wrong, the state the dynamics arrived at did not carry the information.

Oscillator parameters

shared by all nodes · weights stay as trained
These six numbers are the physics. On a real device they would be a pump current or a bias voltage. Move them and the trained network is forced to compute in a regime it was not trained for.
Bifurcation parameter. λ < 0: trajectories relax to the origin with rate |λ|. λ > 0: self-sustained oscillation (if γreal < 0).
Rotation rate in the complex plane, radians per step. Set from the input's period: 2π/28 for sMNIST, 2π/175 for IMDb, 0.15 for Mackey-Glass.
Cubic amplitude term. Negative values cap the amplitude at r = √(−λ/γreal); positive values make growth self-reinforcing.
Shear: the rotation speed depends on amplitude, φ̇ = ω + γimag r². Nonzero values twist trajectories in phase space.
Strength of the tanh-bounded force α·tanh(I) pushing the real part. These networks are trained in the weakly driven regime, α = 0.04, where the nodes' own dynamics dominate.
Integration step of z ← z + h·ż, one per input sample. Trained at h = 1. Smaller h slows the whole dynamics relative to the input.
—
Try

What the dynamics are doing

Everything below is computed from the network you are looking at, with the parameters currently set, and with its trained connections held fixed. Drag the point on the regime map to move the whole network through its bifurcation landscape. Keep in mind throughout that you are testing one already-calibrated machine in conditions it was not calibrated for, which is a different question from asking how well a network could do if it were trained in those conditions.

Regime map (λ, γreal)

drag the point
Analytical stability of a single unforced node (Appendix A of the paper). The origin is a stable spiral for λ < 0 and unstable for λ > 0; a limit cycle of radius √(−λ/γreal) exists only when λ and γreal have opposite signs, stable in the supercritical quadrant and unstable (a basin boundary) in the subcritical one. The amber dashed line is where the forward-Euler map used by the network actually loses stability at the current ω and h, λc = (√(1 − h²ω²) − 1)/h, slightly left of the analytic λ = 0. Tinted background: this network's score on the shipped test samples at each grid point, computed live, with its trained weights left exactly as they are.

Read this map the right way

The scan concentrates the good scores (the darkest cells) in a small region around the settings this network was trained at. That is expected, and it is not the same thing as a map of which regimes can solve the task. The two ask different questions.

Here, one network, moved out of its world. Its connections were learned at one setting and then frozen, so changing the parameters is like letting a device drift after calibration: the readout is tuned to a machine that no longer exists. This measures tolerance, how far the physics may wander before the calibration is worthless.

A regime sweep trains a fresh network at every point. Each setting gets its own weights, learned from scratch there. That measures capability, the best a network could do with that physics.

A broad optimum in such a sweep and a narrow one here are therefore consistent: a regime can be excellent to train in and still unforgiving to drift away from once you have committed to it.

One node, unforced

phase portrait at current parameters

Amplitudes |zi(t)| during the sequence

all nodes · current sample

Effective dimension of the state

PCA of the 2N-dimensional trajectory

Memory of a single input

linear response kernel

Readout in the state

final states of the shipped samples

Confusion matrix

live, at the current parameters

Accuracy against network size

every N shipped

Performance against ω

frequency sweep

The networks on this page

The equations, integrator, weight layout and preprocessing are those of the public code (stuart-landau-oscillator-network on GitHub). The browser implementation was checked against the PyTorch model to within single-precision round-off.

Background

How the model works, and what the networks show

The Stuart-Landau oscillator and the normal form of a Hopf bifurcation, how the network is wired and integrated, what each of the three tasks demands of it, and the findings task by task. About a ten-minute read, on its own page so this one stays an instrument.

Read the background
—
—
—