Towards Green Wearable Computing: A Physics-Aware Spiking Neural Network for Energy-Efficient IMU-based Human Activity Recognition

Abstract: Wearable IMU-based Human Activity Recognition (HAR) relies heavily on Deep Neural Networks (DNNs), which are burdened by immense computational and buffering demands. Their power-hungry floating-point operations and rigid requirement to process complete temporal windows severely cripple battery-constrained edge devices. While Spiking Neural Networks (SNNs) offer extreme event-driven energy efficiency, standard architectures struggle with complex biomechanical topologies and temporal gradient degradation. To bridge this gap, we propose the Physics-Aware Spiking Neural Network (PAS-Net), a fully multiplier-free architecture explicitly tailored for Green HAR. Spatially, an adaptive symmetric topology mixer enforces human-joint physical constraints. Temporally, an $O(1)$-memory causal neuromodulator yields context-aware dynamic threshold neurons, adapting actively to non-stationary movement rhythms. Furthermore, we leverage a temporal spike error objective to unlock a flexible early-exit mechanism for continuous IMU streams. Evaluated across seven diverse datasets, PAS-Net achieves state-of-the-art accuracy while replacing dense operations with sparse 0.1 pJ integer accumulations. Crucially, its confidence-driven early-exit capability drastically reduces dynamic energy consumption by up to 98\%. PAS-Net establishes a robust, ultra-low-power neuromorphic standard for always-on wearable sensing.

• The paper introduces the PAS-Net architecture that combines physics-aware topology mixing and causal neuromodulation to achieve high HAR accuracy while drastically reducing energy consumption.
• It employs sparse, event-driven computation with a novel Temporal Spike Error loss and early-exit mechanism, reducing inference energy by up to 98% compared to traditional DNNs.
• Experimental evaluations on seven IMU datasets demonstrate PAS-Net’s robustness in real-world, continuous wearable sensing by efficiently modeling multi-node kinematic dependencies.

Physics-Aware Spiking Neural Network (PAS-Net) for Green Wearable Computing

Introduction and Motivation

The paper "Towards Green Wearable Computing: A Physics-Aware Spiking Neural Network for Energy-Efficient IMU-based Human Activity Recognition" (2604.10458) addresses the fundamental limitations of dense DNN-based inference on battery-constrained edge devices in continuous IMU-based HAR. Existing DNNs—especially state-of-the-art CNN/LSTM hybrids, attention mechanisms, and Transformer models—require high-latency floating-point MAC operations and full temporal-window buffering, resulting in intractable power consumption and critical inference latencies for real-world, always-on wearables.

Pure SNN paradigms offer spike-driven, event-based computation, promising ultra-low energy footprints (down to 0.1 pJ/operation). However, canonical SNNs lack spatial awareness of human kinematic topology, are unable to adapt to non-stationary biomechanics with static thresholds, and struggle with gradient flow over long sequences, limiting both recognition accuracy and latency reduction.

Architecture of PAS-Net

PAS-Net introduces architectural novelties that enable state-of-the-art IMU-HAR accuracy while achieving energy and latency characteristics unattainable with standard DNN or SNN approaches.

The overall PAS-Net architecture applies the following pipeline to a raw multi-node IMU tensor:

1. Invariant Tokenization & Spiking Embedding: Temporal patching and rotation-invariant feature extraction condense and align input data, enabling a compressed, robust transition into the event-driven spike domain with LIF-based spiking embedding. Temporal-Aware Batch Normalization ensures information-preserving normalization without tempo degradation.
2. Deep Physics-Aware Spiking Core: Stacked PAS-Blocks incorporate:
   • Adaptive Symmetric Topology Mixer: This enforces bidirectional, learnable spatial routing constrained by human physical kinematics, with symmetric adjacency matrices preventing spurious or biased connection patterns and enhancing the modeling of joint biomechanics.
   • O(1) Causal Neuromodulation: Each node maintains a causal exponential moving average of spike activity, used to modulate the LIF firing threshold in real time, yielding dynamic, context-aware, energy-gated neuron activation.
   • Spiking Dilated Temporal Convolutions: Residual, dilated convolutions along the temporal axis efficiently aggregate multi-scale temporal dependencies within the event-driven regime.

     Figure 1: The PAS-Net framework, from rotation-invariant tokenization to deep adaptive topology-aware, event-driven processing.

3. Temporal-Wise Readout and Early-Exit: A fused mean+max spatial pool followed by linear classification enables stepwise output at each downsampled time index. Training employs a Temporal Spike Error (TSE) loss with membrane warmup, yielding robust early gradients and supporting streaming, confidence-based early exits.

   Figure 2: Stepwise cumulative accuracy illustrating PAS-Net's sub-second early-exit mechanism for select activities; complex transitions require longer evidence accumulation.

Experimental Evaluation

PAS-Net is benchmarked against seven public IMU datasets (PAMAP2, Daily-Sports, TNDA-HAR, HuGaDB, HAR70+, Parkinson FOG, USC-HAD) under rigorous subject-independent splits. All comparative models are implemented and optimized under uniform protocols with PyTorch/SpikingJelly. Hardware energy profiles are computed leveraging 45nm CMOS logic cost models.

Accuracy

PAS-Net establishes new SNN SOTA in all evaluations, with absolute gains of up to 12% over the strongest previous SNNs on complex multi-node datasets. More strikingly, PAS-Net often exceeds mainstream heavy DNNs (e.g., DeepConvLSTM, ResNet-SE), refuting the assumption that event-driven computation is necessarily at odds with top accuracy (PAMAP2: +7% over ResNet-SE; USC-HAD: +9% over DeepConvLSTM). On challenging episodic datasets, PAS-Net remains the top SNN, with transformer-based DNNs occasionally offering marginal improvements.

Figure 3: Confusion matrix for DeepConvLSTM on PAMAP2; severe off-diagonal confusion for geometrically similar actions.

Figure 4: Confusion matrix for PAS-Net on PAMAP2; inter-class confusion significantly reduced due to explicit topology routing.

Energy Efficiency

PAS-Net executes nearly all inference using sparse, event-driven integer accumulates, reducing total energy consumption per inference by up to 98% compared to floating-point DNNs, with dynamic adaptation to input complexity. For PAMAP2, PAS-Net requires only $2.57~\mu$J (vs. 94.62~$\mu$J for DeepConvLSTM), and on TNDA-HAR, the reduction is $>99\%$ compared to the most demanding attention mechanisms.

Responsiveness and Early-Exit

PAS-Net's temporal supervision allows sub-second response with minimal accuracy degradation for most activities. On datasets with highly discriminative kinematics, the model exits after a single downsampled patch, yielding $>98\%$ energy savings. Only ambiguous or transition-heavy inputs require full-horizon evaluation.

Ablations

The removal of adaptive topology routing or causal neuromodulation severely degrades performance, especially on multi-node, biomechanically complex tasks. The TSE loss is essential for both peak accuracy and for enabling robust, low-latency early exit.

Figure 5: Ablation study on topology routing, causal neuromodulation, and TSE loss. Removal of any component substantially degrades performance across diverse datasets.

Theoretical and Practical Implications

PAS-Net fundamentally demonstrates that multiplier-free, event-driven SNNs, when augmented with explicit physical topology and causal modulation, can match or exceed DNNs in accuracy for complex, noisy, real-world sequences—without incurring the hardware infeasibility or latency of dense models. The method's architectural approach generalizes: any system that requires fast, low-power, always-on inference of streaming, structured sensor signals could benefit from explicit model-informed topology, adaptive firing thresholds, and confidence-based streaming supervision.

Practical deployment is limited only by the maturity of neuromorphic hardware. PAS-Net is architecture-ready for chips like Loihi and TrueNorth, and the paper's energy accounting is grounded in real CMOS operation counts. Robustness to sensor displacement and generalization to adversarial or highly non-stationary domains will depend on future advances in online domain adaptation and continual learning within event-driven frameworks.

Visualizing Biomechanical Awareness

Layerwise analysis shows PAS-Net learns physically meaningful anatomical dependencies (Figure 6), adaptively routing information across nodes in accordance with established kinetic chains, from localized self-connections to global joint coordination.

Figure 6: Layerwise visualization of learned symmetric topology masks for sparse human full-body topology (hand, chest, ankle sensors in PAMAP2).

Conclusion

PAS-Net achieves best-in-class accuracy on standard HAR benchmarks while reducing energy-delay product by more than an order of magnitude. Its explicit modeling of human kinematics, event-driven dynamic gating, and streaming early-exit strategy establish it as an archetype for genuinely deployable, robust, ultra-low-power wearable AI. The design principles underpinning PAS-Net are likely to influence not only IMU-based HAR but also other real-time, structured sensor interpretation problems in embedded contexts.

For the complete experimental details, comparative model implementations, datasets, and code, see "Towards Green Wearable Computing: A Physics-Aware Spiking Neural Network for Energy-Efficient IMU-based Human Activity Recognition" (2604.10458).