Unverified Re-invented 2026

Signed Phase-Attractor Layer

Implementation & benchmark of arXiv:2607.00286 — Self-Organized Learning in Oscillatory Neural Networks with Memristive Signed Couplings

Usefulness6/10

Difficulty5/10

Novelty6/10

Dynamics Accuracy Stability Sample efficiency Rnn Retrieval Graph nn Dynamical systems Geometry Optimization Algebra

Source paper: Self-Organized Learning in Oscillatory Neural Networks with Memristive Signed Couplings arXiv:2607.00286 ⓘ · analyzed Aug 29, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Build a recurrent associative-memory layer whose hidden variables are phases rather than unconstrained activations, and whose symmetric couplings may be positive or negative. Positive edges attract two units to the same phase, while negative edges attract them to phase difference $\pi$, allowing memories containing both correlation and anti-correlation constraints to remain stable after the external input is removed.

Formulas

$$\dot{\phi}_{i}=\omega_{i}-\sum_{j\neq i}K_{ij}(x_{ij})\sin(\phi_{i}-\phi_{j})$$

$$F(\phi\mid x)=-{1\over 2}\sum_{i<j}K_{ij}(x_{ij})\cos(\phi_{i}-\phi_{j})$$

$$R(x)=R_{\mathrm{off}}(1-x)+R_{\mathrm{on}}x,\qquad G(x)=R(x)^{-1}$$

$$K_{ij}=\frac{\kappa}{N}A_{ij},\qquad A_{ij}=A_{ji}=\sum_{\mu=1}^{P}\cos\!\left(\theta_i^\mu-\theta_j^\mu\right),\qquad \phi_i^{t+1}=\phi_i^t-\Delta t\sum_{j\ne i}K_{ij}\sin(\phi_i^t-\phi_j^t)$$

Mathematical statement

For $N$ oscillators, $\phi_i(t)\in\mathbb{R}$ is the phase of unit $i$, $\omega_i$ is its intrinsic frequency, and $K_{ij}=K_{ji}$ is a signed symmetric coupling. The paper uses $\dot{\phi}_i=\omega_i-\sum_{j\ne i}K_{ij}(x_{ij})\sin(\phi_i-\phi_j)$, where $x_{ij}\in[0,1]$ is a memristor state and $K_{ij}(x_{ij})$ is the effective edge weight. In the homogeneous rotating frame $\omega_i=0$, the dynamics are gradient descent on $F(\phi\mid x)=-\frac12\sum_{i<j}K_{ij}(x_{ij})\cos(\phi_i-\phi_j)$. Therefore $K_{ij}>0$ favors equal phase and $K_{ij}<0$ favors phase difference $\pi$. For a software layer, use normalized signed couplings $K_{ij}=\kappa A_{ij}/N$, with $A=A^T$, and interpret the final phase vector modulo $2\pi$. A phase-memory coupling can be constructed as $A_{ij}=\sum_{\mu=1}^{P}\cos(\theta_i^\mu-\theta_j^\mu)$, where $\theta^\mu\in[0,2\pi)^N$ is stored pattern $\mu$. The hardware resistance model is $R(x)=R_{\mathrm{off}}(1-x)+R_{\mathrm{on}}x$ and $G(x)=R(x)^{-1}$, with $R_{\mathrm{on}},R_{\mathrm{off}}>0$; signed effective weights require differential excitatory/inhibitory branches rather than a single nonnegative conductance.

Implementation notes

(1) Exact integration point: implement this as a recurrent refinement layer between an encoder and a classifier or retrieval head. The encoder maps an input to an initial phase vector $\phi^0\in[0,2\pi)^N$, for example $\phi_i^0=2\pi\,\sigma(z_i)$. The recurrent layer performs 10–50 relaxation steps. Store a symmetric coupling matrix $K$ as a fixed Hebbian buffer for the first experiment, or make it trainable with the exact symmetry parameterization $K=(W+W^T)/2$. (2) Pseudocode: inputs are $\phi^0$, symmetric $K$, step size $\Delta t$, and number of steps $T$. For each $t=0,\ldots,T-1$, compute $D_{ij}=\phi_i^t-\phi_j^t$, $g_i=\sum_{j\ne i}K_{ij}\sin(D_{ij})$, and update $\phi^{t+1}=\operatorname{wrap}_{2\pi}(\phi^t-\Delta t g)$. Decode using $[\cos\phi,\sin\phi]$, or retrieve pattern $\mu$ by maximizing $\sum_i\cos(\phi_i-\theta_i^\mu)$. The paper's continuous-time equation supplies the vector field and its $F$ supplies an energy diagnostic. (3) Compute $K$ from the phase-Hebbian formula, symmetrize it exactly, and normalize by $N$. Estimate basin size, energy decrease, and retrieval accuracy empirically. Check $F(\phi^{t+1})\le F(\phi^t)$; reduce $\Delta t$ if energy rises. For learned couplings, use $K_{ij}=\kappa\tanh(W_{ij})/N$ to bound interaction strength. (4) First cheap experiment: use binary MNIST or CIFAR-10 class prototypes converted to phases, corrupt 20–50% of phases with additive noise or flips by $\pi$, and compare against a standard Hopfield network with the same parameter count and recurrent-step budget. Include memories with deliberately anti-correlated phase relations, where unsigned couplings cannot encode the target constraints. Measure retrieval accuracy versus corruption level, energy monotonicity, relaxation steps to convergence, and robustness after removing the input clamp. Success means higher anti-correlated-pattern retrieval at equal recurrent FLOPs, a larger basin of attraction, or faster convergence; failure is no improvement over the unsigned control.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.