Sparsity-Aware Event-Driven Impulse Radio Transceivers for Reliable Neuromorphic Inference

Abstract: The growing number of Internet-of-Things (IoT) based AI applications deployed at resource-constrained network edge call for ultra-reliable and low-latency data processing pipelines from distributed front-end sensors to remote inference units. Meanwhile, brain-inspired neuromorphic computing featuring spiking neural networks (SNNs) have arisen as a new paradigm for energy-efficient AI inference. However, significant energy and time expenses incurred in high-complexity transceivers that combat fading and multi-user interference hinder implementations of multi-user neuromorphic inference for edge intelligence. To address this challenge, we consider in this paper a broadband multi-user remote inference system that integrates event-based sensing and time-hopping (TH) on-off keying (OOK) based ultra-wideband (UWB) communications for reliable neuromorphic inference. Specifically, we propose a novel two-timescale repetition coding that leverages intra-frame pulse sparsity for low-latency repetition. We also develop two neuromorphic inference schemes based on: (i) digital spike encoding that recovers each pixel of the event-frame by threshold-adaptive detection via an SNN based sparsity estimator; and (ii) analog spike encoding that converts noisy correlator outputs at the receiver into analog-valued inputs for end-to-end (E2E) classification. Finally, numerical results validate the effectiveness of the proposed coding schemes, and reveal a signal-to-noise ratio (SNR)-dependent performance crossover between the two inference schemes, indicating that analog spike encoding based schemes are preferable with mild or high SNR while digital spike encoding based schemes remain robust in low SNR regime.

• The paper introduces a sparsity-aware impulse radio transceiver design that optimizes event-driven transmission for neuromorphic inference.
• It proposes a two-timescale repetition coding strategy combining inter-frame and intra-frame repetitions to enhance reliability under fading conditions.
• Empirical results on gesture classification show an SNR-dependent performance crossover between digital and analog spike encoding, guiding adaptive system design.

Sparsity-Aware Event-Driven Impulse Radio Transceivers for Reliable Neuromorphic Inference

System Architecture and Model

The work presents a comprehensive architecture for multi-user remote inference leveraging event-driven neuromorphic sensing and ultra-wideband (UWB) impulse radio (IR) communication. The system comprises $K$ spatially-distributed event-camera-equipped users and a remote inference unit implementing spiking neural networks (SNNs) for object recognition tasks.

At the sensing front-end, users deploy dynamic vision sensors (DVS) that produce highly sparse spatio-temporal event frames. Each event is mapped to a bit stream using polarity-aware binary encoding, significantly reducing transmitted data volume compared to frame-driven approaches. Transmission is performed using time-hopping on-off keying (TH-OOK), which inherently matches the sparse event structure, reducing multi-user interference and conserving energy. The receiver—after selective RAKE (SRAKE) and maximal ratio combining (MRC)—prepares received data for neuromorphic inference.

Figure 1: System block diagram encompassing DVS event generation, sparsity-matched TH-OOK transmission, and SNN-based remote inference.

Key advantages of the design arise from leveraging event sparsity via TH-OOK, ensuring low collision rates, scalable multi-user support, and compatibility with the statistical characteristics of neuromorphic outputs.

Two-Timescale Repetition Coding Scheme

The paper introduces a two-timescale repetition coding strategy exploiting event stream sparsity to enhance resilience against fading and interference without excessive latency or bandwidth cost. The approach combines:

• Inter-frame repetition: Each bit is transmitted across multiple frames to improve reliability.
• Intra-frame repetition: Within a frame, sparsity leaves numerous chips idle, which are utilized for additional pulse repetitions of '1'-bits.

This achieves $N_s = N_f N_p$ total repetitions using just $N_f$ frames, increasing coding gain and robustness while keeping transmission overhead minimal due to sparse event activation.

Figure 2: Conceptual example of the two-timescale repetition coding for two users, showing inter- and intra-frame repetition allocation.

Numerical analysis shows that increasing intra-frame repetition ($N_p$) significantly narrows the performance gap between practical deployment and ideal channel conditions, especially at moderate-to-high SNR, before collision effects emerge.

Neuromorphic Inference: Digital vs. Analog Spike Encoding

Two inference paradigms are developed and comparatively studied:

• Digital Spike Encoding (Separated Inference): Event frames are reconstructed from received statistics through adaptive thresholding using sparsity estimates provided by an SNN-based estimator, followed by conventional SNN classification.
• Analog Spike Encoding (End-to-End Inference): Channel output statistics are normalized (e.g., via a sigmoid nonlinearity) and input directly into the SNN classifier, allowing the network to learn channel-adaptive representations in an E2E fashion—bypassing hard detection.

Figure 3: Comparison of digital spike encoding (detection followed by SNN inference) and analog spike encoding (direct E2E channel-aware SNN).

A key technical contribution is the derivation of a MAP-based decision threshold for digital decoding that incorporates event sparsity and an explicit bias parameter $\lambda$, which is numerically optimized in conjunction with SNN-based prior estimation.

Empirical Evaluation and Results

The approach is evaluated on gesture classification tasks using the DVS128 Gesture dataset and ultra-wideband channel models. The experimental protocol includes:

• Pre-trained and fine-tuned SNN classifiers (DVSGestureNet) for both encoding modes.
• SNN-based sparsity estimators for adaptive thresholding in digital inference.
• Monte Carlo averaging for robust statistical performance estimation.

Numerical Results:

• The proposed intra-frame repetition coding yields clear performance gains over inter-frame-only repetition, with larger $N_p$ values further improving accuracy at mid-to-high SNR before saturating due to pulse collisions.

Figure 4: Test accuracy versus SNR demonstrating superiority of intra-frame repetition, especially at $N_p=4$ in high SNR regimes.

• TH-OOK-based schemes deliver substantial accuracy improvements over traditional PPM, validating the efficacy of collision minimization via sparsity-aware design.
• A significant SNR-dependent crossover is observed between the digital and analog inference pipelines: Digital spike encoding is robust under low SNR, while analog encoding outperforms at higher SNR due to superior end-to-end channel adaptation.

Figure 5: Comparative test accuracy across SNRs, highlighting the SNR-dependent crossover between digital and analog spike encoding approaches.

Theoretical and Practical Implications

The integration of sparsity-aware coding, ultra-wideband TH-OOK modulation, and SNN-based inference sets forth a scalable template for IoT-edge neuromorphic systems. Practically, this enables highly energy- and bandwidth-efficient multi-user deployments for real-time AI tasks in wireless and resource-constrained environments. The dual-inference architecture provides flexibility for system adaptation according to link quality and channel statistics.

Theoretically, the results underscore the importance of matching physical-layer design (modulation and coding) to statistical properties of neuromorphic outputs. The findings also chart experimental guidelines for SNR-adaptive selection between digital and analog inference modalities, advancing the regime of joint communication-inference system co-design.

Conclusion

This work systematically addresses the challenge of reliable multi-user neuromorphic inference over fading-prone wireless links by harnessing event sparsity through physical-layer and learning-layer co-optimization. The proposed sparsity-aware two-timescale repetition coding and dual inference pipelines substantially outperform conventional baselines under realistic channel stress, opening avenues for practical, scalable neuromorphic systems in wireless edge environments. The demonstrated SNR-dependent performance crossover further motivates future research on adaptive, context-aware communication-inference integration in intelligent distributed systems.