Winfree Oscillatory Neural Network

• WONN is a phase-based neural network architecture that uses Winfree synchronization dynamics to evolve representations on the torus (S¹)ᴰ.
• It combines fixed trigonometric mappings with learnable neural networks, enabling structured oscillatory and hierarchical interactions in data processing.
• Empirical results demonstrate competitive performance on benchmarks like ImageNet and Sudoku while using significantly fewer parameters.

Winfree Oscillatory Neural Network (WONN) denotes a class of phase-based neural and coupled-oscillator systems organized around Winfree-style synchronization dynamics. In the most explicit architectural formulation, WONN is a dynamical neural architecture based on generalized Winfree dynamics that evolves representations on the torus $(S^1)^d$ through structured oscillatory interactions, combining phase-based inductive biases with flexible and hierarchical interaction mechanisms instantiated as either fixed trigonometric mappings or learnable neural networks (Dai et al., 20 May 2026). In the broader mathematical literature, the same label naturally extends to finite populations of Winfree-type oscillators coupled by a mean field, where the central issues are synchronization, periodic locking, multistability, and oscillator death (Oukil et al., 2015).

1. Conceptual position within oscillator-based neural computation

The classical Winfree model is a system of pulse-coupled phase oscillators with separable sender–receiver interactions. In one standard form,

$$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$

where $\omega_i$ is the natural frequency, $P(\theta)$ is the pulse shape emitted by each oscillator, and $Q(\theta)$ is the phase response curve (PRC) describing how oscillator $i$ reacts to the mean field (Gallego et al., 2017). A closely related formulation used in finite-$N$ synchronization theory is

$$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$

with coupling strength $\kappa$ and spectrum width $\gamma$ of the intrinsic frequencies (Oukil et al., 2015).

WONN inherits this Winfree principle of separable interaction. The 2026 neural architecture begins from

$$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$0

where $$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$1 is a sensitivity function and $$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$2 is an influence function (Dai et al., 20 May 2026). This differs from the Kuramoto model,

$$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$3

because Kuramoto depends only on phase differences, whereas Winfree-style dynamics use separable interactions in which the receiver sensitivity and sender influence are distinct objects (Dai et al., 20 May 2026).

This distinction is important in oscillator-based neural computation. The data repeatedly frames computation as emerging from phase locking, synchronization, and structured collective oscillatory states rather than from conventional weighted sums alone. A plausible implication is that WONN should be understood not merely as an oscillatory variant of a recurrent network, but as a synchronization-centered architecture whose representational geometry and dynamical inductive bias are both inherited from Winfree-type phase dynamics.

2. State space, variables, and update equations

In the neural-architectural formulation, each unit is represented by a phase

$$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$4

and the full hidden state is

$$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$5

The model also maintains a frequency state

$$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$6

so WONN has a dual-state design: a fast oscillatory phase state $$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$7 and a slower frequency or carrier state $$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$8 (Dai et al., 20 May 2026).

The input is embedded into the initial frequency state,

$$\dot\theta_i=\omega_i+ Q(\theta_i)\,\frac{\varepsilon}{N}\sum_{j=1}^N P(\theta_j), \qquad i=1,\dots,N,$$9

while the initial phase is randomly sampled,

$\omega_i$0

Within a layer, the core recurrence approximating Winfree dynamics is

$\omega_i$1

with shared parameters across recurrent steps (Dai et al., 20 May 2026).

The forward pass is organized as stacked synchronization layers. For each layer $\omega_i$2 and recurrent step $\omega_i$3, the model computes sensitivity and patch-wise influence, then updates phases according to

$\omega_i$4

$\omega_i$5

After $\omega_i$6 steps, the network performs cross-layer updates,

$\omega_i$7

and a final output head consumes $\omega_i$8 (Dai et al., 20 May 2026).

Although the state space is toroidal, practical computation is performed through the trigonometric embedding

$\omega_i$9

The phase update is described as

$P(\theta)$0

which preserves circular geometry while using standard neural operations (Dai et al., 20 May 2026).

WONN also introduces grouped Winfree dynamics. Oscillators may be partitioned into spatial groups or patches $P(\theta)$1, within which states are aggregated into a shared influence signal. The group size $P(\theta)$2 controls interaction scale: $P(\theta)$3 gives point-wise interactions, whereas larger $P(\theta)$4 gives more shared contextual influence. Coupling can be local via convolution or global via attention; the reported experiments mostly use global attentive coupling as the default WONN variant (Dai et al., 20 May 2026).

3. Synchronization theory and the Winfree mechanism

A central analytical notion in Winfree systems is bounded phase dispersion. In the finite-$P(\theta)$5 mean-field model, synchronization is defined by the requirement that

$P(\theta)$6

remain bounded uniformly in time $P(\theta)$7 (Oukil et al., 2015). This does not mean identical phases or convergence to a fixed phase difference; it means that the pairwise phase spread remains uniformly controlled forever.

For the model

$P(\theta)$8

with $P(\theta)$9, the 2015 synchronization theorem identifies an open region $Q(\theta)$0 in the $Q(\theta)$1-plane such that $Q(\theta)$2 is independent of $Q(\theta)$3, independent of the specific frequency distribution, and its closure contains

$Q(\theta)$4

For every $Q(\theta)$5, any sufficiently clustered initial condition remains synchronized forever, and for some initial condition the system exhibits periodic locking of the form

$Q(\theta)$6

where all $Q(\theta)$7 are periodic with common period $Q(\theta)$8 (Oukil et al., 2015).

The key hypothesis introduced there is

$Q(\theta)$9

After changing variables to the mean phase $i$0, the dispersion inequality becomes

$i$1

This is why H3 matters: if the average of $i$2 over one period is positive, then one can build a periodic dispersion curve that traps the dynamics and preserves bounded phase spread (Oukil et al., 2015).

In the continuum limit, the Ott–Antonsen ansatz yields an exact low-dimensional representation for the classical Winfree model. For Lorentzian frequency distribution,

$i$3

the Kuramoto order parameter $i$4 obeys the exact closed ODE

$i$5

with $i$6 a function of $i$7 determined by the pulse shape (Gallego et al., 2017). The reduced system exhibits two structurally distinct synchronization scenarios: one organized by an ordinary Bogdanov–Takens point, and another organized by a mutated Bogdanov–Takens point denoted BT'. The transition between them depends on pulse shape and PRC offset $i$8, and the paper infers a rule of thumb: pulses that decay faster to zero near $i$9 are more prone to the BT' scenario (Gallego et al., 2017).

This body of theory places several recurrent WONN themes on rigorous footing. Synchronization is not only a function of coupling magnitude; it depends on the interplay among heterogeneity, pulse shape, PRC offset, and the separable geometry of sensitivity and influence. A plausible implication is that neural architectures adopting Winfree-style interaction functions inherit a dynamical design space in which pulse waveform and response asymmetry can qualitatively alter the accessible collective regimes.

4. Oscillator death, inertia, and generalized Winfree dynamics

Winfree systems do not only support rotating synchronized states; they also support oscillator death. For the standard sinusoidal first-order model

$N$0

equivalently

$N$1

a coupling strength exceeding twice the maximal magnitude of the intrinsic frequencies,

$N$2

guarantees convergence for Lebesgue almost every initial data (Ryoo, 3 Jan 2026). The same paper gives the lower bound

$N$3

for the limiting order parameter and shows that the total number of distinct equilibria modulo $N$4 is at most $N$5 (Ryoo, 3 Jan 2026).

The inertial Winfree model introduces second-order phase dynamics,

$N$6

or

$N$7

For this model, the 2026 analysis proves a pathwise oscillator-death theorem with explicit smallness thresholds

$N$8

yielding

$N$9

for every oscillator (Moreno-Earle et al., 3 May 2026). The same work proves a qualitative zero-inertia synchronization result: if inertia and spreads in natural frequencies and initial velocities are sufficiently small relative to coupling, then oscillator death occurs and

$$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$0

for any prescribed $$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$1 (Moreno-Earle et al., 3 May 2026).

By contrast, the 2022 inertial study establishes that small coupling can still produce phase locking. With phase diameter

$$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$2

its deterministic theorem proves

$$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$3

under explicit small-dispersion assumptions, and in the stochastic multiplicative-noise case it gives the lower-probability estimate

$$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$4

for the pathwise emergence of bounded phase diameter (Kang et al., 2022).

Winfree dynamics have also been generalized beyond scalar phases. On the special orthogonal group, the matrix-valued model

$$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$5

reduces to the classical Winfree model when $$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$6 (Ha et al., 27 Apr 2026). In the non-identical case, the paper proves a positively invariant trapping region, a leader–follower mechanism, and $$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$7-exponential stability of solutions; in the identical-oscillator regime, it proves complete state synchronization and oscillator death exponentially fast with explicit decay rates (Ha et al., 27 Apr 2026).

Taken together, these results show that WONN-type dynamics are compatible with at least three distinct coherent regimes: bounded rotating phase-locking, quiescent oscillator death, and higher-dimensional relaxation on manifolds. This suggests that the computational role of synchronization in WONN cannot be reduced to a single notion of “more coherence.”

5. Learning, circuit realization, and implementation regimes

The 2026 WONN architecture parameterizes the interaction maps in two ways. A fixed trigonometric parameterization uses

$$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$8

while a learnable parameterization implements $$\dot{x}_{i}=\omega_i-\kappa \frac{1}{N}\sum_{j=1}^{N}P(x_j)R(x_i), \qquad i=1,\dots,N,$$9 and $\kappa$0 as MLPs (Dai et al., 20 May 2026). Under the continuous-time trigonometric system

$\kappa$1

if the coupling matrix is symmetric and the natural frequencies vanish, the interaction energy

$\kappa$2

is a Lyapunov function, with

$\kappa$3

when $\kappa$4 (Dai et al., 20 May 2026). In the reported Maze-hard experiments, this energy is used as a diagnostic signal: multiple trajectories are sampled and the one with lowest final interaction energy is selected (Dai et al., 20 May 2026).

A hardware-oriented design line appears in the 2023 ONN paper, which explicitly uses the same synchronization-based computational paradigm associated with Winfree oscillatory neural networks. There, the target hardware is a network of resistively coupled 7-inverter ring oscillators described by

$\kappa$5

with physical coupling resistances

$\kappa$6

Positive or in-phase coupling connects node 3 of one oscillator to node 3 of another, while negative or anti-phase coupling connects node 3 of one oscillator to node 6 of another (Rudner et al., 2023).

Learning in that circuit model is performed by Backpropagation Through Time applied to the ODE simulator. The workflow is explicitly: simulate the coupled oscillator network over time, compute a loss from the final or late-time oscillator state, backpropagate through the time-unfolded dynamics, update the learnable circuit parameters, then hard-wire the learned values into hardware. The implementation uses PyTorch and torchdiffeq, including differentiable ODE solvers and the adjoint method to reduce memory cost (Rudner et al., 2023).

These two implementation regimes are distinct. The 2026 WONN is a torus-valued neural architecture with grouped hierarchical interactions, whereas the 2023 design is a circuit-level ONN based on resistively coupled ring oscillators. The papers nevertheless converge on the same computational motif: phase-locked collective states are the computational substrate.

6. Benchmarks, applications, and adjacent phase-based models

On image recognition and reasoning tasks, the reported WONN results are unusually broad for a synchronization-based architecture. The paper evaluates CIFAR-10, CIFAR-100, ImageNet-100, ImageNet-1K, Maze-hard, and Sudoku (Dai et al., 20 May 2026). On CIFAR-10, WONN reaches about 95.26% with 11.84M parameters; on CIFAR-100, about 76.20% with 11.86M parameters. On ImageNet-100, it reaches up to 82.88% accuracy, and on ImageNet-1K, up to 76.78% with about 12.28M parameters (Dai et al., 20 May 2026). The same paper states that WONN is, to its knowledge, the first synchronization-based oscillatory architecture to scale competitively to ImageNet-1K (Dai et al., 20 May 2026).

On Maze-hard, WONN uses point-wise interactions $\kappa$7, with $\kappa$8 and $\kappa$9. The reported numbers are 76.2% accuracy with 0.396M parameters, and 80.1% with energy voting. The comparison table gives HRM at 74.5% with 27M parameters and AKOrN at 36.2% with 1M parameters, so WONN achieves 80.1% accuracy using only 1% of the parameters of prior state-of-the-art models (Dai et al., 20 May 2026). On Sudoku, WONN achieves 100% accuracy with 1.58M parameters (Dai et al., 20 May 2026).

The 2023 ONN design paper reports complementary results in a hardware-realizable oscillator setting. For binary associative memory on the $\gamma$0 MNIST subset, the reported table gives: Hebbian: 1176 parameters, MSE $\gamma$1; Proposed fully connected ML: 2352 parameters, MSE $\gamma$2; Proposed nearest-neighbor ML: 312 parameters, MSE $\gamma$3 (Rudner et al., 2023). For multi-class classification on MNIST, the same paper reports about 70–75% accuracy for an FFNN-like oscillatory structure, about 65–70% for a winner-take-all ONN ensemble, and 96.7% for an ONN preprocessing layer plus a small trained neural output layer; for binary classification, both the fully connected and nearest-neighbor ONN classifiers reach about 98% accuracy (Rudner et al., 2023).

A common source of terminological confusion is the conflation of Winfree-type and Kuramoto-based oscillator networks. The 2025 Sudoku solver is explicitly based on the Kuramoto model,

$\gamma$4

not on a classical Winfree pulse-coupled model (Porfir et al., 4 Aug 2025). Its digit-to-phase encoding uses

$\gamma$5

and for standard $\gamma$6 Sudoku it uses 81 oscillators and an $\gamma$7 weight matrix instead of 729 neurons and a $\gamma$8 matrix (Porfir et al., 4 Aug 2025). The reported benchmark trend is that both ONN and Hopfield baselines achieve 100% correct solutions at 5 unknown cells, the ONN reaches 95% at 10 unknowns while the Hopfield network drops to 71.43%, the ONN remains above 90% at 15 unknowns while the Hopfield network falls to 42.86%, and the ONN still achieves over 33% at 25 unknowns where the Hopfield model reaches 0% (Porfir et al., 4 Aug 2025).

This adjacent literature clarifies the scope of WONN. Strictly speaking, the named architecture in (Dai et al., 20 May 2026) is Winfree-based, torus-valued, and hierarchically structured. More broadly, however, WONN sits inside a larger family of phase-based computing systems in which synchronization, phase locking, and dynamical relaxation implement memory, classification, and constraint satisfaction. The exact dynamical law—Winfree, inertial Winfree, matrix-valued Winfree, or Kuramoto—determines the formal theory, but the computational primitive remains organized around collective phase evolution rather than static feed-forward activation.