EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures
Neuromorphic chips slash inference power while staying within 1.2% of CNN accuracy
EdgeSpike is a co-designed spiking neural network framework that brings brain-inspired computing to resource-constrained edge devices. The researchers integrated four key innovations: a hybrid surrogate-gradient/direct-encoding training pipeline, a hardware-aware neural architecture search (NAS) bounded by energy and memory budgets, an event-driven runtime supporting Intel Loihi 2, SpiNNaker 2, and ARM Cortex-M microcontrollers, and a local plasticity rule for on-device adaptation without backpropagation. Across five representative sensing tasks—keyword spotting, vibration-based machine fault detection, sEMG gesture recognition, 77 GHz radar human-activity classification, and structural-health acoustic-emission monitoring—EdgeSpike achieved 91.4% mean accuracy, only 1.2 percentage points behind strong INT8 CNN baselines (92.6%). The real breakthrough is energy efficiency: on neuromorphic hardware, inference energy dropped 18–47× (mean 31×); on Cortex-M, it dropped 4.6–7.9× (mean 6.1×). End-to-end latency never exceeded 9.4 ms. A seven-month, 64-node wireless field deployment confirmed a 6.3× extension in projected battery lifetime (from 312 to 1,978 days at 2 Wh per node) and showed that on-device adaptation limited accuracy drift to just 0.7 pp under seasonal conditions versus 2.1 pp without. The NAS explored 8,400 candidates, producing a 12-point Pareto front. The entire framework will be released as open source, including reproducible training pipelines, portable runtimes, and benchmark suites.
- EdgeSpike achieves 91.4% mean accuracy (within 1.2 pp of CNN) while reducing energy per inference by 18–47× on neuromorphic chips and 4.6–7.9× on ARM Cortex-M.
- A 64-node field deployment over 7 months extended battery life from 312 to 1,978 days (6.3× improvement) with only 0.7 pp accuracy drift via on-device adaptation.
- Hardware-aware NAS evaluated 8,400 candidates and produced a 12-point Pareto front; framework is open source with reproducible pipelines and portable runtimes.
Why It Matters
EdgeSpike makes autonomous sensing viable for years-long deployments with negligible accuracy loss, enabling truly low-power IoT AI.