Network-Optimised Spiking Neural Network for Event-Driven Networking

TL;DR

The paper tackles the challenge of achieving low-latency, energy-efficient decisions in time-critical networks under sparse, bursty telemetry. It introduces Network-Optimised Spiking (NOS), a compact two-state neuron with bounded excitability, differentiable resets, and graph-local, delayed inputs that map directly to queueing semantics, enabling gradient-based optimization. Through equilibrium and stability analysis, including a Perron-mode spectral threshold, and by incorporating stochastic arrivals, the work demonstrates how NOS maintains headroom and mitigates cascades across chain, star, and scale-free topologies while offering practical calibration and neuromorphic deployment guidance. Empirically, NOS yields competitive or superior early-warning F1 scores and detection latency compared with ML baselines under a residual-based protocol, and the framework provides a principled link between node-level physics and network topology, facilitating topology-aware control in edge/neural hardware deployments.

Abstract

Time-critical networking requires low-latency decisions from sparse and bursty telemetry, where fixed-step neural inference waste computation. We introduce Network-Optimised Spiking (NOS), a two-state neuron whose variables correspond to normalised queue occupancy and a recovery resource. NOS combines a saturating excitability nonlinearity for finite buffers, service and damping leaks, graph-local inputs with per-link gates and delays, and differentiable resets compatible with surrogate gradients and neuromorphic deployment. We establish existence and uniqueness of subthreshold equilibria, derive Jacobian-based local stability tests, and obtain a scalar network stability threshold that separates topology from node physics through a Perron-mode spectral condition. A stochastic arrival model aligned with telemetry smoothing links NOS responses to classical queueing behaviour while explaining increased variability near stability margins. Across chain, star, and scale-free graphs, NOS improves early-warning F1 and detection latency over MLP, RNN, GRU, and temporal-GNN baselines under a common residual-based protocol, while providing practical calibration and stability rules suited to resource-constrained networking deployments. Code and Demos: https://mbilal84.github.io/nos-snn-networking/

Figures (16)

• Figure 1: NOS unit with graph-local inputs, weighted per-link gates and delays, exogenous shot-noise, bounded excitability, recovery dynamics, stochastic threshold, and differentiable reset. Three non-overlapping lanes: left for $u\!\to\!v$, centre for $v\!\to\!v_{\mathrm{th}}$, right for reset $\to v$.
• Figure 2: Operational margin $\Delta_{\mathrm{op}}(\chi, I_{\max})$ as a function of subthreshold damping $\chi$ and maximum steady input $I_{\max}=\max_i I_i^*$. The red contour marks $\Delta_{\mathrm{op}}=0$; points above this curve (larger $I_{\max}$ for the same $\chi$) lie beyond the small-signal existence bound. Parameters $\{\alpha,\kappa,\lambda,\beta,a,b,\mu,\gamma,v_{\mathrm{rest}}\}$ are held fixed as in the text. methods\ref{['app:global_stability']} gives the analytical definition of $\Delta_{\mathrm{op}}$ and one-dimensional sweeps versus $I_{\max}$, together with sensitivity to $\rho(W)$ and $g$.
• Figure 3: Network-level dynamics in NOS. (a) Delay effects: introducing link delays $\tau_{ij}>0$ shifts the stability boundary, causing oscillations to emerge at lower coupling $k$. (b) Absolute topology effects: chain networks destabilise last (resilient), stars first (fragile), with scale-free in between. (b$"$) Relative to Perron-mode prediction: chain networks deviate strongly above the $1/\rho(W)$ line, while star and scale-free graphs track the prediction closely.
• Figure 4: Test-range alignments under train-calibrated residual thresholds. A marker at the truth dot is correct, to the right is late, to the left is early. Extra markers indicate false positives, and missing markers near truth indicate false negatives. NOS remains close to truth across all three topologies while avoiding small oscillations.
• Figure 5: Summary metrics under the label-free residual protocol. Higher bars are better for F1, precision, and recall. Lower curves are better for MAE, RMSE, and latency. NOS achieves the best F1, the lowest forecast error, and the earliest starts across all three topologies.

Theorems & Definitions (6)

• Lemma 1: Monotone stabilisation by saturation
• proof : Proof sketch
• Corollary 1.1: Network threshold increases with saturation
• Corollary 1.2: Spectral collapse of the coupling threshold
• Lemma 2: Existence and uniqueness
• Corollary 2.1: Algebraic condition and networking view