NeuroPulse: Spiking-Transformer Hybrid Architecture for Ultra-Low-Power Continual Learning in Neuromorphic Network Processors

Conventional deep learning networks impose prohibitive energy requirements on continuously operational network intelligence applications such as anomaly detection, traffic classification, and adaptive Quality-of-Service (QoS) control. This paper proposes NeuroPulse, a spiking-transformer hybrid neural architecture that combines the temporal sparsity of spiking neural networks (SNNs) with the representational power of sparse self-attention, enabling efficient deployment on neuromorphic network processors (NNPs). We propose a Rate-Coded Cross-Attention (RCCA) module, which converts population-coded spike-trains into attention queries, allowing long-range dependency modeling within sub-milliwatt (sub-mW) power budgets. NeuroPulse also supports catastrophe-free continual learning on non-stationary network traffic distributions via a Hebbian Synaptic Consolidation (HSC) mechanism, eliminating the need for full model retraining. Experiments on NSL-KDD, UNSW-NB15, and real-world 5G RAN telemetry datasets demonstrate that NeuroPulse achieves 94.3% intrusion detection accuracy at 0.23 mW average energy consumption—a 12× power reduction over transformer-only baselines—while retaining 97.1% of accumulated knowledge after 50 sequential task updates, making it uniquely suited for always-on intelligent network nodes.