Neuromorphic Wireless Split Computing with Resonate-and-Fire Neurons (2506.20015v1)

Abstract: Neuromorphic computing offers an energy-efficient alternative to conventional deep learning accelerators for real-time time-series processing. However, many edge applications, such as wireless sensing and audio recognition, generate streaming signals with rich spectral features that are not effectively captured by conventional leaky integrate-and-fire (LIF) spiking neurons. This paper investigates a wireless split computing architecture that employs resonate-and-fire (RF) neurons with oscillatory dynamics to process time-domain signals directly, eliminating the need for costly spectral pre-processing. By resonating at tunable frequencies, RF neurons extract time-localized spectral features while maintaining low spiking activity. This temporal sparsity translates into significant savings in both computation and transmission energy. Assuming an OFDM-based analog wireless interface for spike transmission, we present a complete system design and evaluate its performance on audio classification and modulation classification tasks. Experimental results show that the proposed RF-SNN architecture achieves comparable accuracy to conventional LIF-SNNs and ANNs, while substantially reducing spike rates and total energy consumption during inference and communication.

• The paper proposes a neuromorphic split computing architecture that leverages resonate-and-fire neurons for efficient, event-driven processing of wireless signals.
• Results demonstrate that BRF neurons achieve competitive accuracy with significantly reduced transmission and computation energy compared to LIF models.
• The system employs OFDM-based spike mapping and regularization techniques for sparsity, enhancing energy efficiency in edge computing.

Neuromorphic Wireless Split Computing with Resonate-and-Fire Neurons

Introduction

The paper "Neuromorphic Wireless Split Computing with Resonate-and-Fire Neurons" investigates a novel approach to edge computing in wireless environments, utilizing resonate-and-fire (RF) neurons. Neuromorphic computing leverages spiking neural networks (SNNs) which operate through event-driven computation. RF neurons provide oscillatory dynamics, capturing time-localized spectral features natively, thereby eliminating the need for preliminary spectral pre-processing typically required by traditional spiking models such as the leaky integrate-and-fire (LIF) neurons.

Figure 1: Neuromorphic wireless split computing architecture based on RF-SNNs: (a) The transmitter acquires time-domain signals with informative spectral components, such as audio or baseband (radio) signals.

Resonate-and-Fire Spiking Neural Network

RF neurons are introduced to support the processing of streaming time-domain signals. Such neurons are capable of resonating at preferred frequencies, acting as biological band-pass filters that accumulate input oscillations to produce spikes when input frequencies align with the neuron’s own natural resonance frequency.

This approach builds on previously introduced balanced RF (BRF) models which incorporate adaptive refractory mechanisms to adjust spiking activity while maintaining memory of previous inputs. These enhanced BRF models ensure temporal sparsity, reducing both computation and transmission energy considerably compared to LIF neurons.

Figure 2: Illustration of the dynamic of the BRF neuron ($\hat b=15$) when the input is a sinusoid with frequency $188.50~\rm{rad/s}$.

Split Computing Architecture

The proposed architecture partitions the SNN across a transmitter and a receiver, mediated by a wireless link using orthogonal frequency-division multiplexing (OFDM). The transmitter processes streaming signals into spike events, each mapped onto OFDM subcarriers, allowing real-time communication despite constraints in energy usage.

This setup reduces the number of transmitted spikes due to resonance-driven sparsity, which directly translates to reduced transmit power requirements, advantageous in wireless sensor networks where energy efficiency is critical.

Figure 3: Timeline of a neuromorphic wireless split computing system.

System Design and Energy Optimization

For efficient energy usage, the system adopts regularization techniques targeting sparsity promotion during training, minimizing the spike rate in either the encoding or decoding components of the split SNN. This reduces local compute energy and transmission energy proportional to the number of spikes handled.

The system accounts for senomic operations, showcasing that traditional energy models that ignore somatic dynamics may inadequately represent the real energy consumption in such advanced neuromorphic setups, particularly when comparing the more complex BRF neuron dynamics.

Figure 4: Digital implementation schematic of the somatic operation of an ALIF neuron: (a) ALIF neuron dynamics.

Performance Analysis

Through simulation, the paper demonstrates RF-SNN models achieving competitive accuracy to conventional models while drastically reducing total energy consumption. Specifically, it was shown that RF-based SNNs achieve better sparsity-accuracy trade-offs in both audio processing using the SHD dataset and baseband signal classification using the ITS dataset.

Results indicate BRF neurons provide superior energy efficiency with competitive accuracy due to their resonant filtering, demonstrating significant reductions in compute and transmission energy.

Figure 5: Energy consumption versus accuracy for different neuron models on the SHD dataset.

Conclusion

The introduction of RF neurons in wireless split computing architectures yields promising results, significantly reducing energy consumption while maintaining competitive accuracy for real-time signal processing tasks. The inherent spectral sensitivity of RF neurons aligns well with the requirements of streaming applications encountered in wireless communications and audio recognition. Future work can explore dynamic frequency tuning within RF dynamics to adaptively respond to diverse environmental conditions.

Overall, this research contributes substantially to the field, hinting at the potential for low-energy, high-efficiency computing, meeting the tailored demands of edge AI systems in a rapidly evolving technological landscape.