Source-linked AI summary
Benchmarking spiking neural networks across sensing modalities on edge devices
Xin Du, Di Yu, Changze Lv, Yuqi Zhang, Zhuo Chen, Wentao Tong, Helin Zheng, Weisong Zhang, Xiaofan Zhao, Linshan Jiang, Shijie Ji, Hui Fang, Xiaoqing Zheng, Gang Pan, Shuiguang Deng
TL;DR
Edge sensing lacks systematic evidence about when SNNs outperform ANNs under diverse modalities and deployment constraints. This paper benchmarks SNNs across modalities, model designs, and hardware settings, finding selective advantages—especially in wireless sensing—while showing that energy gains can entail latency or memory costs.
Problem
The central gap is limited systematic evidence about when SNNs provide practical edge-sensing benefits across modalities and constrained hardware.
Method
The paper constructs a unified benchmark covering five modalities, 24 datasets, seven physical devices, four neuromorphic platforms, and 4,662 evaluations.
Results
SNN benefits are selective: performance is broadly comparable to ANNs across most workloads, while wireless sensing is consistently favorable.
Takeaways & Limitations
The benchmark shows that deployment advantages are multidimensional because energy improvements do not automatically reduce latency or memory usage.
Takeaways & Limitations
The benchmark does not exhaust edge sensing, and other modalities may exhibit different signal statistics and deployment trade-offs.
Abstract
from arXiv · showhide
Edge computing systems need to support diverse sensing workloads under tight energy and memory constraints, thereby motivating deployment-aware model selection. Spiking neural networks (SNNs) are a promising alternative to conventional artificial neural networks (ANNs), yet systematic evidence for when and why they provide practical advantages remains limited. Here, we present a benchmark of SNNs across five sensing modalities and multiple edge devices, systematically evaluating spike encoding, neuron models, and network topologies under consistent training and deployment protocols. We find that SNN advantages are strongly modality-dependent: while SNNs achieve performance broadly comparable to ANNs across most workloads, wireless sensing emerges as a particularly favorable domain. Frequency-domain and feature-space analyses further explain this result by showing that spiking dynamics naturally align with the spectral-temporal structure of wireless signals. Our deployment analysis further shows that SNN advantages are not one-dimensional, with energy gains often accompanied by modality-dependent system costs. Finally, we provide an open-source framework for reproducible benchmarking and deployment profiling, offering a practical foundation for algorithm-software-hardware co-design on emerging edge and neuromorphic computing platforms.
Short title: Benchmarking SNNs in edge sensing
This study benchmarks SNNs for heterogeneous edge sensing and finds that their practical benefits depend on sensing modality, model design, and deployment hardware rather than appearing universally. The benchmark supports deployment-aware selection by jointly considering accuracy, latency, energy, and memory.
- Edge sensing systems must maintain reliable performance across diverse modalities under tight energy, latency, and memory constraints.
- SNN advantages depend on interactions among spike encoding, neuronal dynamics, network topology, input statistics, and hardware constraints.
- The benchmark covers five modalities, 24 datasets, seven physical edge devices, four neuromorphic platforms, and 4,662 evaluations under unified protocols.
- Across most modalities, SNN accuracy is broadly comparable to ANN baselines, while wireless sensing emerges as a consistently favorable domain.
- SNN deployment advantages are multidimensional because energy improvements may not reduce latency or memory usage across hardware platforms.
- Soul–NeuSim provides open-source infrastructure for reproducible multimodal benchmarking, deployment profiling, and neuromorphic simulation.
RESULTS
The framework evaluates SNN design choices and deployment behavior across heterogeneous sensing workloads and hardware settings. It compares SNNs with ANNs using unified protocols and multiple model- and system-level metrics.
- The framework organizes SNN evaluation around spike encoding, neuronal dynamics, and network topology.
- The benchmark spans five sensing modalities and compares model accuracy with deployment behavior under unified protocols.
- The analyses examine cross-modal design choices, wireless signal structure relative to spiking dynamics, and deployment gains across edge and neuromorphic devices.
SNN design choices exhibit strong modality dependence
SNN design choices behave differently across sensing modalities, with no universally dominant encoding, neuron model, or architecture. Wireless sensing shows the strongest and most consistent SNN gains over matched ANN baselines.
- Spike encoding: Direct encoding achieved the highest relative accuracy but is not spike-compatible; practical encodings varied substantially by modality.
- Neuron models: No neuron model dominated across modalities, with TLIF improving over LIF by up to 4.79% on BullyDetect.
- Network architectures: Transformer-style SNNs, including Spikformer and QKFormer, achieved the strongest overall performance and more consistent behavior across evaluated modalities.
- Cross-paradigm comparison: SNN and ANN accuracy was broadly comparable in vision, acoustic, motion, and neuromorphic tasks, while wireless sensing showed the largest and most consistent SNN advantage.
- Accuracy-complexity trade-off: Spikformer used 0.45M parameters versus ResNet50’s 23.55M, corresponding to an approximate 52× reduction while retaining strong wireless performance.
- Wireless sensing: On Widar, SpikingResFormer reached 81.42±0.32%, exceeding ResNet18 at 71.70±0.33% by 9.72%.
- Wireless sensing: On AOPHand, Spikformer achieved 95.12±0.42% versus the ANN best of 72.54±0.80%, a 22.58-point margin.
- Accuracy-complexity trade-off: Top-performing SNNs occupied a favorable accuracy-complexity frontier, whereas extremely lightweight recurrent ANNs showed substantial performance degradation.
Spectral and feature-space analyses explain the wireless advantage
Spectral and feature-space analyses associate wireless SNN advantages with low-pass spiking dynamics that align with discriminative wireless signal structure. Deployment profiling shows that these benefits involve trade-offs among latency, energy, memory, architecture, and hardware.
- Spectral analysis: The analysis uses synthesized multi-tone signals, Fourier transforms, and temporally pooled wireless sequences to compare signal spectra with spiking dynamics.
- Spectral analysis: LIF neurons attenuate high-frequency components as temporal low-pass filters, whereas ReLU preserves more high-frequency components.
- Wireless spectral structure: Wireless task information is primarily embedded in energy envelopes and Doppler shifts, supporting spectral alignment between wireless data and spiking dynamics.
- Feature-space analysis: SNNs can classify noisy wireless samples by filtering high-frequency noise and emphasizing cleaner low-frequency motion information.
- Deployment trade-offs: On conventional hardware, SNN latency was 3.07× ANN latency for neuromorphic event sensing and 2.84× for wireless workloads.
- Deployment trade-offs: Energy reductions reached up to 29.58× for neuromorphic event sensing but were as low as 1.27× for vision workloads.
- Deployment trade-offs: Mean memory usage increased by 65.56% for neuromorphic workloads and 41.91% for wireless workloads, reflecting retained temporal states and activations.
- Hardware dependence: Hardware heterogeneity affected latency distributions, with transformer-style Spikformer showing greater variation than convolutional backbones.
Soul-NeuSim supports reproducible benchmarking and neuromorphic co-design
Soul–NeuSim unifies algorithm benchmarking, deployment profiling, and neuromorphic simulation for reproducible edge-sensing evaluation. Its modular workflow supports modality-aware comparison and algorithm–software–hardware co-design.
- Soul provides standardized data processing, spike encoding, model construction, training configuration, and deployment-oriented evaluation.
- NeuSim translates trained SNNs into hardware-constrained representations and estimates latency, communication, resource utilization, and energy.
- Together, Soul and NeuSim span algorithm development, deployment profiling, and neuromorphic hardware simulation under a consistent evaluation protocol.
DISCUSSION
SNNs provide selective rather than universal benefits for edge sensing, with wireless sensing emerging as the clearest favorable domain. Their practical value reflects interactions among modality, model design, and deployment hardware, with energy gains sometimes accompanied by latency and memory costs.
- Wireless sensing was the clearest favorable domain for SNNs, while most other modalities achieved accuracy broadly comparable to ANN baselines.Wireless workloads were more sensitive to spike encoding and neuronal dynamics, and several spiking architectures substantially outperformed ANN counterparts.
- LIF-type neurons act as temporal dynamical filters that attenuate high-frequency components while preserving slower variations in wireless signals.Spectral analysis, feature-space visualization, and reconstruction analysis connect this behavior to wireless signal structure.
- Spike encoding, neuronal dynamics, and network topology jointly determine which information is preserved and how temporal, spatial, and cross-channel dependencies are represented.The results indicate that these design components should be evaluated together rather than independently.
- SNN advantages cannot be reduced to energy efficiency because multi-timestep processing often increases latency and membrane-state memory on conventional edge devices.Energy reductions were substantial in temporally sparse workloads, especially neuromorphic event-sensing tasks.
- Hardware heterogeneity shaped SNN trade-offs, with GPUs partly reducing latency penalties and CPUs exposing stronger sequential execution overhead.Convolutional spiking models had more stable latency distributions than transformer-style spiking models across devices.
- Soul–NeuSim provides a reproducible pipeline for comparing SNNs across algorithmic and system-level dimensions.The framework integrates multimodal processing, encoding, neuron construction, benchmarking, edge profiling, and neuromorphic simulation.
- The benchmark does not exhaust edge-sensing diversity, and its mechanistic analysis is intentionally focused on wireless sensing.Biomedical, tactile, industrial, and multimodal workloads may exhibit different signal statistics and deployment trade-offs.
MATERIALS AND METHODS Experimental design
The benchmark compares SNNs and ANN baselines across five sensing modalities using representative datasets and controlled preprocessing, temporal, and evaluation protocols. It covers conventional and neuromorphic sensing while preserving modality-specific input structures where possible.
- SNNs were compared with ANN baselines across heterogeneous modalities, model designs, and deployment substrates under controlled training and evaluation protocols.The experimental design evaluated spike encoding, neuronal dynamics, and network topology.
- The benchmark covered vision, acoustic, motion, wireless, and neuromorphic event sensing, using four representative datasets per modality.Four additional wireless datasets expanded evaluation across WiFi channel-state-information and millimeter-wave sensing workloads.
- The study represented distinct signal types, including frame-based spatial perception, time-frequency acoustics, inertial time series, wireless channel or radar measurements, and sparse event streams.These modality definitions support comparisons across heterogeneous sensing workloads.
- Dataset-specific training and test splits and preprocessing conventions were followed, including augmentation for images and mel-frequency spectrograms for acoustic signals.Motion samples were constructed as multivariate inertial inputs, with further preprocessing details supplied separately.
- Non-neuromorphic inputs were converted into temporally structured inputs before SNN processing, using four default timesteps.Neuromorphic workloads used 10 timesteps to preserve their native event-driven temporal structure.
Spike-encoding strategies
The study evaluates multiple spike encodings, neuron models, and surrogate-gradient settings to isolate how temporal representation and neuronal dynamics affect SNN behavior. Direct coding serves as an analog upper-bound reference rather than a spike-compatible format.
- The encoding set included rate, time-to-first-spike, burst, phase, and temporal-switch coding for converting continuous sensor inputs into spike trains.All methods generate binary event streams over a temporal window.
- Direct coding replicated normalized inputs across timesteps as constant analog drive and was used as an upper-bound software reference.Because it relies on real-valued synaptic input, it was not treated as spike-compatible for spike-only neuromorphic hardware.
- Encoding configurations were selected by dataset and input format, then held constant across models within each benchmark setting.This design ensured comparability among models.
- Soul-NeuSim decoupled neuron models from network topology through a modular neuron factory supporting 13 spiking neuron models and 13 surrogate-gradient formulations.The models varied in integration, leakage, thresholding, reset behavior, and adaptive dynamics.
- Surrogate-gradient backpropagation was used for training, with the arctangent surrogate function as the default benchmark setting.Holding topology and training settings fixed while changing neuron models isolates neuronal-dynamics effects from architecture confounders.
Network architectures and ANN baselines
The benchmark spans diverse SNN backbones under unified training and evaluation protocols, while using a shared LeNet baseline to reduce architecture-level bias in ANN–SNN comparisons.
- The benchmark includes 12 SNN backbones and hybrid models spanning convolutional, residual, transformer-based, and hybrid attention-based designs.
- LeNet provides a common lightweight ANN–SNN baseline, enabling comparisons of activations and neuronal dynamics under the same architecture.
- Model capacity was controlled by matching parameter counts or aligning computational complexity when exact architectural matching was unavailable.
- Models were trained with Adam, cosine annealing, an initial learning rate of 1×10^-3, batch size 16, and 150 epochs unless otherwise specified.
- All tasks used top-1 accuracy, while deployment evaluation measured inference latency, energy consumption, and peak memory footprint.
- Inference latency was measured after warm-up as wall-clock time per sample, with P90 latency additionally characterizing runtime variability.
Physical edge-device deployment
The deployment study evaluates standardized ANN and SNN execution across physical CPU-, GPU-, and NPU-based edge devices, alongside wireless signal analyses using frequency-domain representations.
- Seven physical edge devices cover CPU-, GPU-, and NPU-based execution environments, including Raspberry Pi 4B, Jetson, Redmi K80, Pixel 6, and Huawei Mate 40 platforms.
- Deployment comparisons reused the same trained model, input representation, and preprocessing pipeline whenever supported by each backend.
- Latency, energy, and peak memory were recorded under identical batch-size and numerical-precision settings within each device.
- Wireless analysis reduced non-temporal dimensions through mean pooling, then transformed the resulting temporal proxy signals into the frequency domain using discrete Fourier transforms.
- Millimeter-wave analyses included feature-time energy maps, feature-Doppler maps, and micro-Doppler spectrograms derived with Fourier and short-time Fourier transforms.
- Task-relevant frequency structure was quantified with a Fisher-criterion discriminative spectrum over frequency bins.
Frequency-matching score
The frequency-matching score measures how closely LIF neuron dynamics retain task-discriminative wireless frequency content, complemented by feature-space and reconstruction analyses.
- The frequency-matching score measures alignment between wireless discriminative spectra and the frequency response induced by LIF neuron dynamics.
- The LIF membrane decay factor controls frequency attenuation and the neuron's effective passband because LIF dynamics act as a first-order low-pass filter.
- For each decay factor, the method computes a DC-normalized power response and weights retained discriminative spectral content to produce the score.
- Higher scores indicate greater overlap between frequencies emphasized by spiking dynamics and frequencies carrying task-relevant wireless information.
- t-SNE visualizations and silhouette scores assess whether ANN and SNN latent representations form separable class clusters under the same wireless task.
- Reconstruction comparisons qualitatively examine whether representations retain noise-dominated components or preserve cleaner task-relevant structure.
Compilation, partitioning, and mapping
NeuSim compiles SNNs into hardware-constrained neuromorphic simulations by extracting connectivity, partitioning workloads, and mapping logical cores onto a physical mesh.
- NeuSim extracts high-level neural-network layers into neuron–synapse connectivity represented as a directed acyclic graph.
- Partitioning groups neurons and synapses into logical cores subject to per-core neuron and synapse capacity constraints.
- Sequential partitioning is the default, while user-defined strategies include locality-preserving space-filling-curve ordering followed by binary segmentation.
- Mapping assigns logical cores to physical cores in the two-dimensional mesh after capacity-valid partitioning.